NASA ARSET_ Estimation of PM2.5 from AOD – Methodologies and Available Datasets

0:02 Speaker 1: Welcome to Part two of our aur set training on

0:05 Speaker 1: estimating surface PM two point five using satellite data and

0:08 Speaker 1: other information sources. In this part, we'll be learning about

0:12 Speaker 1: the estimation of PM two point five from AOD and

0:16 Speaker 1: discussing the methodologies and available data sets using these methodologies

0:21 Speaker 1: to create PM two point five estimates. The trainers for

0:26 Speaker 1: this part will be doctor Aaron von Donkolar, a research

0:30 Speaker 1: associate at the Washington University in Saint Louis, doctor Paulan Gupta,

0:34 Speaker 1: a research scientist in Godard Space Flight Center, and they'll

0:38 Speaker 1: be joined by doctor Jin Juncio, a assistant research scientist

0:43 Speaker 1: at Morgan State University. Our objectives for this training will

0:49 Speaker 1: be by the end of part two that participants will

0:52 Speaker 1: be able to explain the geophysical, hybrid and machine learning

0:56 Speaker 1: methodologies used to infer surface PM two point five from satellite,

1:00 Speaker 1: AOD information and other data sources. Second, we hope you'll

1:04 Speaker 1: be able to differentiate between these available data products for

1:07 Speaker 1: surface PM two point five from NASA and the Washington

1:10 Speaker 1: University in Saint Louis based on their methodologies and the

1:13 Speaker 1: strengths and weaknesses of the two data products. And finally,

1:18 Speaker 1: you'll be able to use NASA tools and the SATPM

1:20 Speaker 1: website to access these surface PM two point five data

1:24 Speaker 1: products for a region of interest and a time period

1:26 Speaker 1: of interest to you. Recall from Part one last week

1:32 Speaker 1: that PM two point five is a measure of the

1:34 Speaker 1: in situ mass concentrations of particles smaller than two point

1:38 Speaker 1: five microns and aerodynamic diameter, and that aerosol optical depth

1:42 Speaker 1: or AOD is an optical measurement of the atmospheric column

1:46 Speaker 1: aerosol loading, that is, the total presence of aerosols in

1:50 Speaker 1: the atmosphere from the surface up to the top of

1:52 Speaker 1: the atmosphere. AOD can be well related to PM two

1:56 Speaker 1: point five under certain conditions, but that relationship will do

2:00 Speaker 1: grade or break down if, for example, the aerosols are

2:03 Speaker 1: prevalent in the atmosphere well above the surface level, if

2:07 Speaker 1: coarser particles that has particles larger than two point five

2:11 Speaker 1: microns are dominant among the aerosols, if the aerosols are

2:15 Speaker 1: absorbing water, and if the humidity is variable and the

2:19 Speaker 1: particles absorb water under high humidity conditions. If the atmospheric

2:24 Speaker 1: or the surface conditions prevent reliable retrieval of AOD, for example,

2:29 Speaker 1: over highly reflective surfaces or under dense cloud cover, if

2:34 Speaker 1: the spatial resolution of the AOD does not capture the

2:37 Speaker 1: local PM two point five variability as measured by ground

2:40 Speaker 1: based PM two point five monitors, or if the timing

2:44 Speaker 1: of the PM two point five measurements does not well

2:47 Speaker 1: coincide with the satellite overpasses. As a reminder, if you

2:52 Speaker 1: have any questions during this part, please put them into

2:55 Speaker 1: the questions box and we'll address them all at the

2:58 Speaker 1: end of the webinar. You can put the questions in

3:01 Speaker 1: as you go. We'll collect those questions and those questions,

3:04 Speaker 1: along with the answers, will all be posted in a

3:07 Speaker 1: document which will be available on the training website about

3:10 Speaker 1: a week following today's training. So with that, I'll hand

3:15 Speaker 1: things over to doctor Aaron von Dongler to talk about

3:18 Speaker 1: geophysical and hybrid PM two point five estimation methods.

3:24 Speaker 2: Thanks Carl So. Last week we touched on using a

3:27 Speaker 2: simple linear regression between ground based observations of PN point

3:31 Speaker 2: five and satellite retrievals of aerostoptical depth as it means

3:35 Speaker 2: to use AOD to locally estimate P and point five concentrations.

3:39 Speaker 2: We saw that in certain cases this simple approach works

3:42 Speaker 2: really quite well, but in others there can be some

3:45 Speaker 2: severe limitations. For this part of the workshop, I want

3:48 Speaker 2: to focus on two other methods of using AOD to

3:51 Speaker 2: estimate P and T point five that specifically seek to

3:54 Speaker 2: either avoid the use of ground based observation altogether or

3:57 Speaker 2: at least ensure the information they provide me applied over

4:00 Speaker 2: a larger area. I wanted to start by showing the

4:05 Speaker 2: density of publicly available ground based P and twenty five

4:08 Speaker 2: observations around the world. Here of updated figure from Martin

4:12 Speaker 2: at Ull twenty nineteen showed the average population weighted distance

4:15 Speaker 2: to ground based monitors for countries around the world. You'll

4:19 Speaker 2: note that this average proximity can vary quite significantly, with

4:22 Speaker 2: populations in large regions in North America, Europe, and China

4:26 Speaker 2: is well a few others, typically residing within less than

4:29 Speaker 2: fifty kilometers in the nearest P and twenty five monitor.

4:32 Speaker 2: Populations within South Asia as a whole typically reside within

4:35 Speaker 2: seventy kilometers, Central Latin America within one hundred and thirty klomebers,

4:40 Speaker 2: and North Africa at least typically within a few hundred clombers.

4:43 Speaker 2: Even over a monitor dense region such as North America,

4:47 Speaker 2: the typical distance can vary significantly, from less than fifteen

4:50 Speaker 2: kilometers over western states to twenty five to fifty over

4:53 Speaker 2: more central ones. The net impact of all this with

4:57 Speaker 2: respect to estimates of P and twenty five from satellite

5:00 Speaker 2: is that any method that seeks to be globally applicable

5:03 Speaker 2: will need to either minimize its direct use of ground

5:05 Speaker 2: based observations or find an effective way to account for

5:09 Speaker 2: how the relationship used between AOD and P five changes

5:12 Speaker 2: with distance from the monitor locations used to constrain that relationship.

5:18 Speaker 2: One way to deal with this challenge is to remove

5:20 Speaker 2: the need for P and twoentty five bonders altogether. To

5:23 Speaker 2: do this, we need to revisit a relationship that was

5:25 Speaker 2: introduced to you last week that calculates AOD given P

5:28 Speaker 2: and twenty five concentration or set of assumptions, namely that

5:32 Speaker 2: skies are cloud free, the aerosols are well mixed, with

5:34 Speaker 2: none above the boundary layer, and last so that the

5:37 Speaker 2: aerosols present are optically homogeneous. That is similar, given these conditions,

5:42 Speaker 2: you can see that AOD is proportional to P and

5:44 Speaker 2: point five concentration itself as well as the boundary layer

5:48 Speaker 2: height and the extinction coefficient, which relates to how much

5:51 Speaker 2: the aerosol in question attenuates that it scatters or absorbs

5:54 Speaker 2: any passing light particle. Effective radius and density are inversely

5:59 Speaker 2: related with the potential for an additional effect from relative

6:02 Speaker 2: humidity if the aerosols hydrophilic. In order to broaden the

6:06 Speaker 2: applicability of this equation, we can relax a few of

6:09 Speaker 2: the initial assumptions and rearrange to solve for p and

6:12 Speaker 2: twoin five given the retrieval of AOD. Here, rather than

6:16 Speaker 2: assuming the aerosols are well mixed throughout the boundary layer,

6:19 Speaker 2: we're going to assume that we know what the fraction

6:22 Speaker 2: of AOD that occurs near the surface up to a

6:24 Speaker 2: heighth delta z and is associated with PAN twoin five.

6:28 Speaker 2: We also remove the assumption that there is no aerosol

6:30 Speaker 2: above the boundary layer and that all the aerosols are similar,

6:33 Speaker 2: instead just focusing on the qualities of those near surface aerosol.

6:38 Speaker 2: With the exception of this fractional term, the formula itself

6:41 Speaker 2: is quite similar, except that now the variables related to

6:43 Speaker 2: aerosol properties are no longer representative of the entire boundary layer,

6:48 Speaker 2: but again just to those aerosols that are near the

6:50 Speaker 2: surface within that delta z height. This equation shown again

6:55 Speaker 2: here forms the basis of the geophysical method to estimate

6:58 Speaker 2: PN twoiny five from ALI. It is theory based, physically driven,

7:03 Speaker 2: and completely independent of ground based P and twoin five observations.

7:08 Speaker 2: The challenge, however, is that it requires some assumptions or

7:11 Speaker 2: knowledge about the aerosols that are producing the retrieved AOD.

7:15 Speaker 2: This includes information on aerosol type which impacts a density,

7:19 Speaker 2: effective radius and extinction coefficient. It also needs an understanding

7:24 Speaker 2: as to how those aerosols respond to changes in relt

7:27 Speaker 2: of humidity. And lastly, it requires some knowledge with the

7:31 Speaker 2: vertical profile, that is, how much of the local airsol

7:33 Speaker 2: are located near the surface or again within that delta

7:36 Speaker 2: zt height. For simplicity, this large massive terms relating AODP

7:41 Speaker 2: and twenty five is typically reduced to a single scalar

7:44 Speaker 2: term called ATA. The challenge of the geophyscal method is

7:48 Speaker 2: how to best represent ata. This is where chemical transfer

7:53 Speaker 2: models play an essential role in most applications of the

7:56 Speaker 2: geophysical method. As a reminder, a CTM is a computer

8:00 Speaker 2: model that simulates atmosphere composition using the physical and chemical

8:04 Speaker 2: equations that govern the state of the atmosphere. It does

8:08 Speaker 2: this by inputting assimilated meteorology and natural and anthrogenic emissions

8:12 Speaker 2: into the model and allowing those equations to represent the

8:15 Speaker 2: atmosphere's chemical and dynamic response, as well as the transport

8:19 Speaker 2: and deposition of these individual chemical constituents or tracers. A

8:23 Speaker 2: CTM provides a complete representation of the atmosphere and in

8:27 Speaker 2: the context of geophyscal p point five, provides all the

8:30 Speaker 2: information we need to calculate ATA. Here we have an

8:35 Speaker 2: example of a CTM predicted ATA, in this case for

8:39 Speaker 2: the month of July twenty twenty three. The different colors

8:42 Speaker 2: are associated with the contributions of different aerosol types to ATA,

8:46 Speaker 2: with yellow fermeneral dust, blue for sea salt, green for

8:49 Speaker 2: organic aerosol, black for carbonaceous aerosol, and red for inorganics

8:53 Speaker 2: being sulfate, ammonium, and nitrate. The ATA values plotted here

8:57 Speaker 2: represents the relationship to instantaneous rather than twenty four hour

9:01 Speaker 2: Pene point five, and you can very clearly see a

9:03 Speaker 2: pulsing and intensity that circles the globe with a diurnal rhythm.

9:08 Speaker 2: This corresponds to the impact of solar heating on bound

9:11 Speaker 2: do lair heights, essentially diluting concentrations at the surface as

9:14 Speaker 2: the bound dolair increases with rising temperatures associated daylight hours,

9:18 Speaker 2: which in turn alters the vertical structure to decrease the

9:21 Speaker 2: value of ATA. Also visible are the regional differences in

9:25 Speaker 2: the particular aerosol sources and types driving the AOD to

9:28 Speaker 2: Pine point five relationship, such as dust emissions over the

9:31 Speaker 2: Sahara organic carbon VIEO biomass burning were parts of South

9:35 Speaker 2: America and Central Africa, or anthropogenic inergantic aerosol over China,

9:39 Speaker 2: Europe and North America. You can even see the impact

9:42 Speaker 2: of the long range transport of dust and biomass burning plumes,

9:45 Speaker 2: where ATA essentially drops to zero, implying that the aerosols

9:49 Speaker 2: almost exclusively aloft and therefore the relationship between AOD and

9:53 Speaker 2: P twointy five is effectively removed. As is pretty clear

9:57 Speaker 2: from this example, using a CTM to represent ATA has

10:00 Speaker 2: some definis advantages. You can see how the spatial and

10:04 Speaker 2: vertical variability impacting ATA is captured via the model's response

10:08 Speaker 2: to emissions. In meteorology, temporal variability is also captured which

10:12 Speaker 2: can have advantages when compensating for sellid overpast times and sampling.

10:16 Speaker 2: And also quite clear is the global applicability of this

10:19 Speaker 2: approach when compared to limiting yourself to justice locations that

10:23 Speaker 2: have dense ground based observations. Counter to these advantages, of course,

10:27 Speaker 2: CTM inherently has limitations to its resolution, whether by code,

10:32 Speaker 2: structure or resource limitations, and most global models are typically

10:35 Speaker 2: about it a resolution on the order of one hundred

10:37 Speaker 2: klometers or so. Any model is also, of course, only

10:42 Speaker 2: as good as the inputs that go into it, and

10:44 Speaker 2: regional uncertainty in emission, inventories or meteorology will impact the

10:48 Speaker 2: quality of the simulation. It should also be noted that

10:52 Speaker 2: while these simulations can be quite effective at representing general features,

10:56 Speaker 2: individual events are much more of a challenge and often

10:59 Speaker 2: limit it to the most effective applications of a simulated

11:02 Speaker 2: eight to two longer term means. And lastly, even at

11:06 Speaker 2: one hundred klumeter resolution, these simulations are quite computationally intensive

11:10 Speaker 2: to run, which can limit an individual's ability to run

11:13 Speaker 2: these simulations themselves and require a dependency on whatever simulation

11:18 Speaker 2: output is already available via other sources. All that said,

11:23 Speaker 2: a CTM based simulation of ATA is often an effective

11:27 Speaker 2: means to quantify how AOD relates to PM two point five,

11:31 Speaker 2: especially as the distance to the nearest ground based monitor increases.

11:36 Speaker 2: Here you can see an example over Eastern Asia, primarily China,

11:39 Speaker 2: for annual mean values during twenty twenty one. The left

11:42 Speaker 2: hand column shows AOD, whereas the right hand column shows

11:45 Speaker 2: the corresponding geophysical PM two point five. Black dots show

11:49 Speaker 2: the locations of the ground based monitors used for the

11:52 Speaker 2: comparisons in the lower row. In this example, you can

11:56 Speaker 2: see quite clearly that AOD is well related directly to

11:59 Speaker 2: P and TOY five concentrations over this region with an

12:02 Speaker 2: R squared of a round point four. Converting this AOD

12:06 Speaker 2: to PM two point five using a simulated ATA not

12:09 Speaker 2: only allows a direct interpretation of this AOD as PM

12:12 Speaker 2: twoint five, but additionally improves the R squared to approximately

12:16 Speaker 2: point five. This example nicely shows some of the strengths

12:21 Speaker 2: as well as from the challenges of this geophysical approach.

12:24 Speaker 2: You can see how major spatial features in PM twoint

12:27 Speaker 2: five are captured and even improved upon compared to AOD alone,

12:30 Speaker 2: and unlike AOD, the air quality implications can be more

12:35 Speaker 2: directly interpreted. On the side of challenges, there is potential

12:39 Speaker 2: for bias and scatter, which is difficult to evaluate without

12:42 Speaker 2: some level of ground based observations. As a result, purely

12:46 Speaker 2: geophysical estimates are typically best used for large spatial and

12:50 Speaker 2: temporal scales. Of course, ground based observations do not strictly

12:56 Speaker 2: have to be used for only evaluation, and can also

12:59 Speaker 2: be to understand any bias is present in the geophysical values.

13:04 Speaker 2: This idea forms the basis of what is known as

13:07 Speaker 2: hybrid PM tw point five or geophysical hybrid PM two

13:11 Speaker 2: point five. Hybrid PM two point five essentially builds on

13:16 Speaker 2: geophysical PM two point five using ground based observations. To

13:21 Speaker 2: give an example, here, i've shown the difference between ground

13:24 Speaker 2: based observations of PM two point five and coincidently sampled

13:28 Speaker 2: geophysical estimates. As you look at this figure, you'll begin

13:32 Speaker 2: to notice regional structures and features within these differences, such

13:36 Speaker 2: as a large scale underestimate of a few microgrants per

13:39 Speaker 2: cubic meter on the part of the geophysical values in

13:42 Speaker 2: the southeastern US, and a large scale overestimate of the

13:45 Speaker 2: same magnitude over the northeastern Great Lakes region. We don't

13:49 Speaker 2: see these large scale features over western regions, but rather

13:53 Speaker 2: we see a high degree of variability between seemingly neighboring

13:57 Speaker 2: monitor locations. Hybrid methods therefore, look at these features using

14:03 Speaker 2: statistical models and training to predict biases in geophysical PM

14:07 Speaker 2: two point five. In other words, these hybrid methods predict

14:11 Speaker 2: limitations associated with AOD retrievals and ATA simulations and their

14:16 Speaker 2: geophysical implications. So how do we do that well. Multiple

14:22 Speaker 2: statistical techniques can be used to make these associations, but

14:26 Speaker 2: perhaps the most intuitive approach uses multiple linear aggression. From

14:31 Speaker 2: last week's session and from other work, most of you

14:34 Speaker 2: will be quite familiar with single linear aggression, such as

14:37 Speaker 2: the example shown in the upper right. Here, a line

14:40 Speaker 2: of best fit is created between a single independent variable

14:43 Speaker 2: X one and some other desired dependant variable, why with

14:47 Speaker 2: a slope of beta one. One way to think of

14:50 Speaker 2: this result is that beta one shows the response of

14:52 Speaker 2: Y to changes in X one. Multiple linear aggression is

14:57 Speaker 2: an extension of single linear aggression, where as the name implies,

15:01 Speaker 2: multiple dependent variables are used. Rather than fitting a line

15:06 Speaker 2: of best fit, you're fitting a plane. But the concept

15:08 Speaker 2: is very similar and the result produces a series of

15:11 Speaker 2: predictor coefficients denoted by betas in this case, that can

15:15 Speaker 2: be interpreted as the linear response of your dependent variable

15:18 Speaker 2: to each independent predictor variable. For some systems, these beta

15:25 Speaker 2: predictor coefficients are a universal truth, that is, they don't

15:29 Speaker 2: change in space or time, but for our particular application

15:33 Speaker 2: estimating the bias in geophysical PM twoenty five, this is

15:36 Speaker 2: unlikely to be the case. As a result, an extension

15:40 Speaker 2: of multiple linear aggression called geographically weighted regression or GWR,

15:44 Speaker 2: is recommended. In effect, GWR allows the predictor coefficients from

15:49 Speaker 2: MLR to vary in space. In the equation shown here,

15:53 Speaker 2: this is denoted by the addition of the I and

15:55 Speaker 2: J subscripts, which are simply identifying different locations on a

15:59 Speaker 2: grid such as the one shown on the right here.

16:03 Speaker 2: In practice, the spatially unique coefficients are calculated by running

16:08 Speaker 2: a series of MLR regressions, one for each location on

16:11 Speaker 2: the grid, where the data points used for each regression

16:14 Speaker 2: are modified based on their local significance. This can either

16:18 Speaker 2: be through some sort of waiting scheme associated with the

16:20 Speaker 2: data points during regression, or their removal entirely if they

16:24 Speaker 2: are not relevant to the location at hand. It's worth

16:28 Speaker 2: noting also that while grids can be regularly spaced, such

16:32 Speaker 2: as I've shown here, they can also be irregular in nature.

16:36 Speaker 2: An irregular grid can be advantageous to capture changes in

16:40 Speaker 2: these relationships in regions where we either expect rapid relational

16:44 Speaker 2: changes in beta or have additional training capacity due to

16:48 Speaker 2: more ground based monitors, or alternatively reducing the computational costs

16:53 Speaker 2: associated with running on a fine grid over regions where

16:57 Speaker 2: neither of these conditions are expected. Okay, so what variables

17:02 Speaker 2: are going to be most effective at predicting the geophysical bias? Well,

17:07 Speaker 2: the ants would really tie back into one of two sources,

17:10 Speaker 2: either uncertainties in the retrieval of AOD itself or uncertainties

17:14 Speaker 2: in the relationship that's being used to relate AOD to

17:17 Speaker 2: PM two point five. Between this week and last, we've

17:21 Speaker 2: covered a lot of the sources of this potential uncertainty,

17:24 Speaker 2: which typically falls into four broad categories. Firstly, composition, Different

17:31 Speaker 2: aerosol types can be associated with different artical properties that

17:34 Speaker 2: may challenge AD retrievals, but probably more impactful in this

17:38 Speaker 2: context is the associations they can have with different emission sources,

17:42 Speaker 2: vertical profiles, and growth factors, all of which impact the

17:46 Speaker 2: AOD to PM two point five relationship. In this way,

17:50 Speaker 2: the presence of a particular aerosol component, such as mineral dust,

17:54 Speaker 2: for example, could be insightful due to uncertainties in the

17:58 Speaker 2: simulated representation of its emission transport, which in turn have

18:02 Speaker 2: the potential to bias ATA. Similarly, meteorological inputs also feedback

18:10 Speaker 2: on ATA within a chemical transport model, with local winds

18:14 Speaker 2: and temperature impacting several natural emission sources for example, or

18:19 Speaker 2: atmospheric conditions impacting the chemical production of aerosol. Land type

18:26 Speaker 2: information can also be related to emission type uncertainties, but

18:30 Speaker 2: primarily I tend to think of this sort of predictor

18:33 Speaker 2: as being associated with uncertainties in the AOD retrieval itself.

18:37 Speaker 2: The dominant assumptions in satellite aidor retrievals typically relate to

18:42 Speaker 2: the amount of light that is reflected off the Earth's surface,

18:44 Speaker 2: which naturally is strongly connected to the type of surface

18:48 Speaker 2: that is doing the reflecting. Urban and desert biases, for example,

18:52 Speaker 2: have both been documented and improved upon in AD retrievals

18:55 Speaker 2: over the years, but surface type remains a solid predictor

18:59 Speaker 2: of some of the challenge is facing an AOD retrieval,

19:01 Speaker 2: and any bias in AOD will directly carry forward into

19:05 Speaker 2: a bias in geophysical PM two point five. Lastly, elevation

19:11 Speaker 2: or more specifically, changes in elevation, can be an important

19:15 Speaker 2: predictor in geophysical bias. In this case, what is being

19:19 Speaker 2: represented is not so much of an error on the

19:22 Speaker 2: part of a model to simulate ATA, but rather changes

19:25 Speaker 2: in the AOD to PM point five relationship that are

19:28 Speaker 2: occurring at a resolution below that of our simulation. You'll

19:32 Speaker 2: recall that simulations are often run at fifty or one

19:34 Speaker 2: hundred klomebers, a mountain or valley can often occur at

19:38 Speaker 2: scales well below this, with obvious implications for how AOD

19:41 Speaker 2: relates to PM two point five. In essence, a predictor

19:46 Speaker 2: such as elevation changes is not so much accounting for

19:48 Speaker 2: errors in a model which could be correctly represent representing

19:53 Speaker 2: the average conditions at the simulated resolution, but still need

19:57 Speaker 2: adjustments to account for the impactive features that occur below

20:02 Speaker 2: the model's resolution. Bringing this all together, we end up

20:06 Speaker 2: with a GWR based equation such as you see here,

20:10 Speaker 2: where the difference between the ground based observations of P

20:13 Speaker 2: and point five and their coinstantly sampled geophysical estimates are

20:17 Speaker 2: locally regressed against a series of predicted variables associated with

20:21 Speaker 2: simulated composition neear logical parameters, elevation parameters, and land type parameters,

20:27 Speaker 2: and below you can see an example of the results

20:29 Speaker 2: that can be obtained. On the left we have the

20:32 Speaker 2: original biases that were observed against the ground based observations,

20:35 Speaker 2: and on the right we see the total GWR predicted

20:38 Speaker 2: bias in the geophysical PN point five. You can see

20:41 Speaker 2: other This approach captures the major features we discussed earlier,

20:44 Speaker 2: such as the overestimate in the Great Lakes region, but

20:47 Speaker 2: beyond that, it predicts that this feature extends further north

20:52 Speaker 2: and west into parts of the Canadian Prairies, a region

20:54 Speaker 2: with far less ground based constraints. One of the real

20:59 Speaker 2: strengths this GWR technique for predicting biases compared to some

21:03 Speaker 2: recent and arguably more advanced machine learning techniques, is a

21:07 Speaker 2: relative ease of understanding the impact of individual variables and

21:11 Speaker 2: how they shape the overall bias prediction. For example, I've

21:16 Speaker 2: plotted here both the predictor coefficients and that impact of

21:19 Speaker 2: the nitrate and percent urban variables on the overall predicted bias.

21:24 Speaker 2: As you can see in this case, nitrate large explains

21:28 Speaker 2: give shape to that Great Lakes Region over estimate, as

21:31 Speaker 2: well as its extension to under monitored regions. This opens

21:35 Speaker 2: up a question about what it is about nitrate that

21:38 Speaker 2: produces this bias, in this case potentially related to uncertainties

21:42 Speaker 2: in either the chemical reactions of all the local emissions.

21:45 Speaker 2: But knowing these connections can in turn direct developments and

21:49 Speaker 2: improvements within the CTM, which can then feed back into

21:53 Speaker 2: the original geophysical p in twenty five with potentially global benefits.

21:58 Speaker 2: You can also see the advantage of allowing predictor coefficients

22:01 Speaker 2: to vary in space. Again shown the left hand panels.

22:05 Speaker 2: Here is a very clear East West divide in the

22:08 Speaker 2: strength of association between urban content and predicted geophysical bias,

22:12 Speaker 2: with the coefficient approximately tripling over western North America compared

22:16 Speaker 2: to the east. Given the fine spatial scale at which

22:21 Speaker 2: present urban impacts hybrid PM twoint five, this may suggest

22:25 Speaker 2: something regionally unique in the way urban surfaces impact the

22:28 Speaker 2: AOD retrieval, or it could be a resolution impact on

22:32 Speaker 2: ATA where the urban environment impacts its relationship on a

22:35 Speaker 2: scale below that of a simulation more strongly in the west.

22:41 Speaker 2: In either event. Putting the impact of all these predicted

22:44 Speaker 2: variables together, we can demonstrate how hybrid PM two point

22:48 Speaker 2: five offers some significant gains compared to geophysical PM two

22:52 Speaker 2: point five alone. Here I've shown an overall global map

22:56 Speaker 2: of PM two point five provided by the hybrid methodology

22:59 Speaker 2: as as well as the improved agreement it offers, with

23:02 Speaker 2: an R squared increasing to zero point nine from approximate

23:05 Speaker 2: point seven, and that root means score difference decreasing by

23:08 Speaker 2: about a factor of two. One of the big strengths

23:12 Speaker 2: of this approach is that it builds on the fundamental

23:15 Speaker 2: strength of the geophysical method while simultaneously taking advantage of

23:19 Speaker 2: the constraint provided by ground based observations. There are, of course, limitations,

23:24 Speaker 2: and first and foremost, unlike purely geophysical estimates, we now

23:28 Speaker 2: require ground based observations, and the challenge becomes how to

23:32 Speaker 2: incorporate those measurements in a way that is most representative

23:36 Speaker 2: of your particular region of interest. Related to this, setting

23:40 Speaker 2: up a model such as GWR or alternatively, machine learning

23:43 Speaker 2: based algorithms like a convolutional know network or Ingredient boosting

23:48 Speaker 2: can be quite complicated, including but not limited to the

23:51 Speaker 2: selection of appropriate predictor variables. Lastly, evaluation of these models,

23:57 Speaker 2: like any statistical models, can be quite challenging, and to

24:00 Speaker 2: that point, appropriate cross validation is key. It's relatively easy

24:05 Speaker 2: to create a model which is very high agreement against

24:08 Speaker 2: the data points and locations that we're used to train it.

24:11 Speaker 2: It can be much more challenging to determine how well

24:14 Speaker 2: your model performs at other locations and times. With this

24:19 Speaker 2: in mind, I'm going to spend a few minutes touching

24:21 Speaker 2: on a few different methods of cross validation. Cross validation

24:25 Speaker 2: is a technique that's used to understand the uncertainty of

24:28 Speaker 2: statistical models. It involves the creation of a subset of

24:32 Speaker 2: data points that can be used for independent evaluation of

24:35 Speaker 2: that model. The idea being that they provide insight into

24:39 Speaker 2: how well a model will perform a part or away

24:42 Speaker 2: from the data that has been used to train it.

24:45 Speaker 2: Cross validation can be time consuming and realistically often has

24:49 Speaker 2: to balance its complexity, accuracy and representation. Above all, the

24:55 Speaker 2: cross validation data set must maintain what I would call

24:58 Speaker 2: true independence. The nature of which may depend on the

25:02 Speaker 2: questions being asked. For applications making use of the resulting model,

25:07 Speaker 2: The most basic form of cross validation is what is

25:10 Speaker 2: referred to as random cafold cross validation. Here, we randomly

25:14 Speaker 2: withhold the set fraction often ten percent of the data

25:17 Speaker 2: set from model training and use the remaining ninety percent

25:20 Speaker 2: for training purposes. We do this multiple times until we've

25:24 Speaker 2: withheld a sufficient amount of the original data set from

25:27 Speaker 2: at least one of the cross validation models, and use

25:30 Speaker 2: the cross validation models predictions of the withheld data points

25:33 Speaker 2: to evaluate the model quality. An advantage of this approach

25:38 Speaker 2: is that it's fairly simple to understand and implement. Is

25:41 Speaker 2: also perfectly suitable to many situations where the points in

25:44 Speaker 2: the data set being modeled are themselves fully independent of

25:48 Speaker 2: one another. For our particular application, those last point is

25:52 Speaker 2: often not the case, with ground based monitors often being

25:55 Speaker 2: clustered around urban centers and the concentration of PM two

25:59 Speaker 2: point five present at one station being very much connected

26:02 Speaker 2: to the concentration that we might expect at another. Buffered

26:07 Speaker 2: cross validation methods have been developed specifically to address this

26:11 Speaker 2: last point. The concept is that when a particular data

26:15 Speaker 2: point where station is withheld for cross validation. Not only

26:18 Speaker 2: is that point withheld for model training, but also any

26:21 Speaker 2: sites located within a given spatial distance. As with kfold

26:26 Speaker 2: cross validation, multiple models are trained and used to evaluate uncertainty,

26:31 Speaker 2: with the distinction being that the sites excluded from each

26:34 Speaker 2: cross validation model are not only the cross validation evaluation sites,

26:39 Speaker 2: but also those within the spatial buffer. The initial implementation

26:45 Speaker 2: of this approach was called buffered leave one out or

26:47 Speaker 2: blue cross validation. It essentially withheld one side at a

26:52 Speaker 2: time pluss buffer points running through each data point in

26:55 Speaker 2: the overall training data set. This is effective, but as

26:59 Speaker 2: you can imagine, is also computationally very expensive, as you

27:03 Speaker 2: need to train a unique cross validation model for each

27:06 Speaker 2: point in your data set. Buffer leave cluster out, and

27:11 Speaker 2: buffer leave isolated sites and cluster out methodologies were developed

27:15 Speaker 2: to combine the improved independence of the blue methodology with

27:19 Speaker 2: improved computational efficiency. Specifically, these methods withhold groups of nearly

27:26 Speaker 2: located sites plus or overlapping buffers out of each cross

27:29 Speaker 2: validation fold, in essence running multiple Blue like models simultaneously,

27:35 Speaker 2: thereby providing similar information. Is Blue both far less training

27:39 Speaker 2: of the individual models. The last type of cross validation

27:45 Speaker 2: that I wanted to highlight is temporal holdback, which is

27:48 Speaker 2: used to address a slightly different question. In this case,

27:52 Speaker 2: the focus of the model being developed is not to

27:55 Speaker 2: represent missing values at different locations, but rather at the

27:58 Speaker 2: same locations as the training data set, just for different

28:01 Speaker 2: time periods. As before, the goal here is to separate

28:06 Speaker 2: a set of data from the training of the model

28:08 Speaker 2: that is independent in the specific way that the model

28:12 Speaker 2: is to be used. In the case of a temporarily

28:14 Speaker 2: focused model, this requires the removal of specific blocks of

28:18 Speaker 2: time from the training process, which can then be used

28:21 Speaker 2: for subsequent validation. This approach does not inform both spatial

28:26 Speaker 2: uncertainties and should not be used in that way, but

28:29 Speaker 2: is quite insightful for specific temporal applications and can be

28:33 Speaker 2: combined with a spacely buffer analysis to provide a complete

28:37 Speaker 2: representation of model uncertainties. Okay, So, in summary, geophysical methods

28:43 Speaker 2: to estimate PM two point five avoid the use of

28:47 Speaker 2: ground based monitors by using the theory based physically driven

28:50 Speaker 2: relationship between total column aerosoloptical depth and PM two point five.

28:56 Speaker 2: CTMs are invaluable for this approach as they are able

28:59 Speaker 2: to represent all the necessary parameters and provide global coverage.

29:04 Speaker 2: Hybrid methods build on the geophysical methods using a statistical

29:08 Speaker 2: framework to predict the bias in geophysical PM two point five,

29:12 Speaker 2: and Lastly, effective cross validation of these or any models

29:17 Speaker 2: must consider how to ensure strong independence of the validation

29:21 Speaker 2: points that are being used. And now we're going to

29:25 Speaker 2: switch gears a little bit and focus specifically on the

29:28 Speaker 2: sat PM two point five data set produced at Washington

29:31 Speaker 2: University in Saint Louis by the Atmospheric Compositional Analysis Group

29:35 Speaker 2: or AGAC. The sat PM two point five data product

29:41 Speaker 2: is the culmination of fifteen to twenty years of research

29:44 Speaker 2: using aerosol optical depth to estimate near surface PM two

29:47 Speaker 2: point five concentrations. The goal of this long running project

29:52 Speaker 2: is to provide a consistent global, long term PM two

29:55 Speaker 2: point five data set, and its current iteration is observationally

29:59 Speaker 2: construc using both satellite AOD and ground based observations within

30:03 Speaker 2: the hybrid geophysical framework that we've been discussing is monthly

30:08 Speaker 2: with regular updates provided annually and at a fairly high

30:11 Speaker 2: resolution on a zero point zero one degree grid, which

30:14 Speaker 2: is about one kilometer. A key feature of this data

30:18 Speaker 2: set from its inception is that it is publicly available

30:21 Speaker 2: currently with its own dedicated website at SATPM dot org.

30:26 Speaker 2: On the right you can see its basic development structure,

30:29 Speaker 2: again following the hybrid geophysical framework, where AOD are converted

30:34 Speaker 2: to geophysical PM twointy five using a simulated representation of ATA,

30:38 Speaker 2: and these values are refined using a hybrid approach. At present,

30:42 Speaker 2: we do maintain a geographically weighted regression type algorithm such

30:45 Speaker 2: as we've been discussing, but primarily we encourage users to

30:48 Speaker 2: make use of a more current algorithm that makes use

30:51 Speaker 2: of a more modern machine learning method convolutional neural networks.

30:59 Speaker 2: I'm going to take a few minutes and address some

31:01 Speaker 2: of the unique challenges in setting up a full fledged

31:04 Speaker 2: hybrid geophysical PM two point five algorithm and how we've

31:07 Speaker 2: addressed these within the sat PM data set. Geophysical PM

31:12 Speaker 2: twoint five all starts with AOD, So the first questioning

31:16 Speaker 2: phase is which AOD retrieval is best suited for PM

31:19 Speaker 2: two point five estimation. The answer, of course can varya

31:23 Speaker 2: both space and time depending on the specific conditions. So

31:26 Speaker 2: to give you a better idea, of how AOD retrievals

31:29 Speaker 2: can vary. I've plotted here four separate AOD retrievals, all

31:33 Speaker 2: monthly means of available data for April twenty fourteen, taken

31:37 Speaker 2: from the same satellite platform. Two instruments represented MODUS and MISER,

31:42 Speaker 2: with the dark Target, Deep blue and MAIC retrievals being

31:45 Speaker 2: applied to the Modus instrument and the MISER instrument having

31:49 Speaker 2: its own specific retrieval algorithm. What will strike you is

31:53 Speaker 2: how in some ways these retrievals are also similar, but

31:56 Speaker 2: in other ways they look quite different. In mind, these

32:00 Speaker 2: are all good, high quality retrievals. Some differences are by design.

32:04 Speaker 2: Dark Target, for example, was never intended to retrieve AOD

32:07 Speaker 2: over deserts, and so data over such regions is simply

32:10 Speaker 2: missing from this retrieval. Deep Blues development initially focused on

32:15 Speaker 2: this very type of surface and well captures the features

32:18 Speaker 2: we might expect over such regions. MEEK specializes in complex terrain,

32:23 Speaker 2: using multiple overpasses to build up a representation of surface reflectants,

32:28 Speaker 2: something which MISER can do in a single overpass, but

32:31 Speaker 2: with far less spatial coverage each time, which results in

32:34 Speaker 2: almost stuttered look for MISER. Despite its high quality of

32:37 Speaker 2: retrieval owing to its reduced sampling of the monthly means

32:41 Speaker 2: shown in all cases, you can see data missing from

32:45 Speaker 2: the far North, where snow cover inhibits the use of

32:49 Speaker 2: any of these methods. So what can we do? We

32:53 Speaker 2: have multiple satellite instruments with multiple AOI retrievals, all which

32:58 Speaker 2: contain useful information, but not necessarily equally under all conditions.

33:03 Speaker 2: In our case, we turned to airnet. Aaronnet was briefly

33:07 Speaker 2: introduced last week, and for our purposes, all we really

33:10 Speaker 2: need to know is that aaronet is a ground based

33:13 Speaker 2: network of sun photometers that provides accurate long term measurements

33:17 Speaker 2: of AOD at different locations around the world. This allows

33:21 Speaker 2: the evaluation of satellite based AOD, but only at locations

33:24 Speaker 2: that have an AARNet site. So the question becomes, how

33:28 Speaker 2: can we extend this site specific information globally to understand

33:33 Speaker 2: the quality of a satellite retrieval anywhere in the world.

33:38 Speaker 2: What we've done is group AARNet locations together and perform

33:42 Speaker 2: separate month specific comparisons for each satellite instrument and or retrieval.

33:47 Speaker 2: These comparisons are grouped by land type DDI or normalized difference,

33:51 Speaker 2: vegetation index, and weighted by distance. We combine the findings

33:55 Speaker 2: of these comparisons based on local conditions around the world

33:59 Speaker 2: to produce a contents consistent definition of uncertainty for each

34:03 Speaker 2: AOD data source. You can see an example of this

34:07 Speaker 2: on the right, where I've shown the normalized root means

34:09 Speaker 2: square difference for Dark Target. Based on this approach. For

34:13 Speaker 2: those of you familiar with this retrieval, it shows a

34:16 Speaker 2: number of features we might expect, such as higher uncertainties

34:19 Speaker 2: over the Western US, where the surface tends to be brighter,

34:22 Speaker 2: presenting a challenge to some of dark Target's inherent assumptions,

34:26 Speaker 2: compared to the Eastern US, where richly vegetated surfaces align

34:29 Speaker 2: very well with those same assumptions. Once we've completed these

34:34 Speaker 2: comparisons for different AOD data sources, we need a way

34:38 Speaker 2: to bring this information together so that we locally rely

34:41 Speaker 2: on each data set relative to its accuracy at any

34:45 Speaker 2: given location. To do this, we employ the relationship shown

34:48 Speaker 2: in the lower right to represent each retrieval's overall uncertainty

34:52 Speaker 2: compared to the total uncertainty of all the other AOD

34:56 Speaker 2: data sources combined. Here are location's final AOD is determined

35:01 Speaker 2: as the weighted average of all available AOD data sources,

35:04 Speaker 2: with each one's weight based on a combination of its

35:07 Speaker 2: local normalized routing square difference. Bias is represented by the

35:11 Speaker 2: slope of the line of best fit and its availability.

35:15 Speaker 2: This equation can be evaluated anywhere in the world, which

35:18 Speaker 2: gives us the overall weighting factors for each retrieval shown here,

35:21 Speaker 2: which provide a measure of the relative local amount of

35:24 Speaker 2: uncertainty for each AOD retrieval. What's encouraging about these results

35:29 Speaker 2: is that these spatial features make intuitive sense based on

35:32 Speaker 2: an understanding of each AOD data source. You can see,

35:36 Speaker 2: for example, dark target contributing heavily over more vegetative regions,

35:41 Speaker 2: deep blue over many of the so called brighter surfaces,

35:44 Speaker 2: and may contributing most over some of the more complicated conditions.

35:49 Speaker 2: We've also brought in CTM simulated AOD as an additional

35:53 Speaker 2: AOD data source, which contributes predominantly over those regions where

35:58 Speaker 2: satellite based retrievals cannot find function, such as snow covered

36:01 Speaker 2: regions in the far North. Here, simulated AOD uncertainty follows

36:06 Speaker 2: a similar concept of comparisons against airnet, but rather it

36:10 Speaker 2: uses categories based on speciation and elevation, which are more

36:13 Speaker 2: globally connected to CTM uncertainty than land surface types in NDBI,

36:19 Speaker 2: and bringing this all together we can produce an overall

36:22 Speaker 2: combined AOD such as shown here, which takes advantage of

36:25 Speaker 2: each data set where it is most accurate. An additional

36:30 Speaker 2: challenge to bringing together multiple AOD data sets is that

36:33 Speaker 2: often there are different spatial scales or time periods. Here

36:37 Speaker 2: I have listed the satellites of retrievals as well as

36:39 Speaker 2: the model we currently use in our sat PM data set.

36:43 Speaker 2: You'll note that some of the earlier retrievals available from

36:45 Speaker 2: seaweeds using deep Blue are available from nineteen ninety eight

36:49 Speaker 2: to twenty ten at a resolution of fourteen kilometers, whereas

36:52 Speaker 2: the most recent editions include a veer's instrument starting in

36:55 Speaker 2: twenty eighteen at resolutions ranging from one to six kilometers. Gs.

37:00 Speaker 2: Chemical transfer model, for its part, is run at a

37:03 Speaker 2: resolution about fifty kilometers throughout this whole time period. One

37:10 Speaker 2: option to combine all these sources would be to simply

37:12 Speaker 2: interplate each source down to the finest one klometer resolution

37:16 Speaker 2: and combined, but this runs the risk of washing out

37:19 Speaker 2: fine scale features when relatively coarse AOD sources are weighted

37:23 Speaker 2: against truly fine ones. To address this possibility, we combine

37:27 Speaker 2: these data sets in a tiered fashion, first averaging all

37:31 Speaker 2: data sets onto the coursest fifty klounder grid. We then

37:36 Speaker 2: apply the ten kilometer relative variation in AOD within that

37:40 Speaker 2: fifty klomter grid according to the satellite based sources. And lastly,

37:46 Speaker 2: we apply the one clometer relative variation within the ten

37:49 Speaker 2: clumdra grid according to MAIC algorithm, which is the only

37:52 Speaker 2: satellite retrieval providing information at such a fine spatial scale. Finally,

38:00 Speaker 2: we linearly interprelate the simulated AOD to PM twenty five

38:03 Speaker 2: relationship down to the one cloonder grid which assumes smooth

38:07 Speaker 2: variation and ATA, and apply it to the one cloonder AOD.

38:13 Speaker 2: As I've stressed throughout this presentation, the geophysical method relies

38:17 Speaker 2: heavily on the inclusion of a chemical transport model. The

38:20 Speaker 2: Atmospheric Compositional Analysis Group is fortunate to be a proud

38:24 Speaker 2: part of the GSM community. GSM is an advanced chemical

38:27 Speaker 2: transport model with a specific mission to advance the understanding

38:31 Speaker 2: of human and natural influences on the environment through a

38:34 Speaker 2: comprehensive state of the science, readily accessible Global Model of

38:37 Speaker 2: Atmospheric Composition. It is continually developed and maintained by an

38:42 Speaker 2: international host of scientists, many of whom are shown in

38:45 Speaker 2: this photograph taken two years ago during the biannual GSM meeting,

38:48 Speaker 2: and from the map below, which shows locations of various

38:51 Speaker 2: GSM contributors for our particular purposes, representing the aerosol driven

38:57 Speaker 2: geophysical relationship between AOD and m PO point five. It

39:01 Speaker 2: is particularly relevant that GSM is open source and community driven.

39:05 Speaker 2: This allows us to include developments specific to our needs,

39:09 Speaker 2: such as recent developments and mineral dust emissions, or particular

39:12 Speaker 2: diagnostics that are well suited to ADA itself. GSM also

39:16 Speaker 2: boasts a detailed aerosol oxiden model representing the inorganic sulfate

39:20 Speaker 2: nitrate ammonium system. Carbonaceous aerosols and natural sources such as

39:24 Speaker 2: mineral dust and sea salts. Emissions are up to date

39:27 Speaker 2: and include meteorologically driven sources such as mineral dust and

39:31 Speaker 2: biogenic VOCs, as well as high resolution anthroogenic inventories such

39:35 Speaker 2: as SAIDs. Lastly, biomass spurning is informed using satellite observations

39:40 Speaker 2: using inventories such as GFAs and g FED. The whole

39:44 Speaker 2: system is driven using assimilated meteorology with variable resolution, but

39:48 Speaker 2: in our case we typically run at about fifty kilometers

39:53 Speaker 2: hybrid PM two point five, as you're aware, depends not

39:57 Speaker 2: only on satellite retrievals and CTM simulations, but requires high

40:01 Speaker 2: quality ground based observations. At AGAC, we actually maintain a

40:05 Speaker 2: database covering approximately fifteen thousand monitor locations around the world.

40:10 Speaker 2: The coverage of this database continues to grow, as you

40:13 Speaker 2: can see from the coverage map shown here, with different

40:15 Speaker 2: colors representing ground based monitors that were brought in during

40:18 Speaker 2: successive SAPPM releases. This database predominantly relies on government based

40:24 Speaker 2: networks and is limited to high quality monitors as opposed

40:28 Speaker 2: to low cost sensors which would complicate the usage in

40:30 Speaker 2: its context. Most of the data sources are publicly available,

40:34 Speaker 2: but there are exceptions which are obtained through direct contact.

40:38 Speaker 2: I've listed the various countries and sources of the monitors

40:41 Speaker 2: we use. This includes data from open AQ, which is

40:44 Speaker 2: an online nonprofit specifically devoted to the aggregation and harmonization

40:49 Speaker 2: of open access air quality data, but additionally, as I mentioned,

40:53 Speaker 2: direct access to many government environmental agencies. Akin to the

40:59 Speaker 2: challenge we face combining multiple AOD sources over a long

41:02 Speaker 2: time period, ground based observations are not temporarily complete. That is,

41:06 Speaker 2: there a very few ground based monitors that have been

41:09 Speaker 2: in continuous operation at the same location for the past

41:11 Speaker 2: twenty five years. As a result, changes to the distribution

41:16 Speaker 2: of monitors over time has the potential to weight the

41:19 Speaker 2: adjustments inferred from these monitors towards regions with the longest

41:22 Speaker 2: histories of monitors or worse introduced temporal changes in these

41:26 Speaker 2: adjustments that are based on changes to the distribution of

41:29 Speaker 2: available monitors over time rather than a meaningful representation of

41:33 Speaker 2: how the impact of bias itself has changed. To help

41:37 Speaker 2: limit this potential impact, we produce a temporarily complete PM

41:41 Speaker 2: two point five record at all ground based monitors prior

41:45 Speaker 2: to incorporating them into the hybrid framework. To complete this task,

41:49 Speaker 2: missing ground based values are inferred using relationships developed during

41:53 Speaker 2: overlapping periods with other available data sources. These other sources

41:58 Speaker 2: include not only geophysical and simulated PIN twoin five, but

42:02 Speaker 2: also other ground based observations from nearby or related stations.

42:07 Speaker 2: Multiple techniques are used to fill in this missing data.

42:10 Speaker 2: All are compared using a temporal holdback, and then combined

42:14 Speaker 2: based on the relative uncertainties. You can see the overall

42:18 Speaker 2: agreement between the inferred and observed P and twoint five

42:21 Speaker 2: and the lower scatterplots shown here, with a left hand

42:23 Speaker 2: panel showing the agreement when other ground based monitors are

42:26 Speaker 2: included within the inference calculations, compared to the right hand panel,

42:30 Speaker 2: which showed the agreement when based strictly on geophysical and

42:33 Speaker 2: simulated P and Twoiny five. A couple sample time series

42:38 Speaker 2: are shown above, with the red line showing the directly

42:41 Speaker 2: observed concentrations, the blue line showing the inferred values that

42:44 Speaker 2: are benefited from other local monitors, and the cyan color

42:47 Speaker 2: showing inferred values where no other local monitors were available.

42:52 Speaker 2: This representation of missing observation obviously adds some of its

42:56 Speaker 2: own certainties, but is overall very much a net benefit

42:59 Speaker 2: for during long term consistency and applying the insight gain

43:04 Speaker 2: from more recent heavily monitored time periods to earlier years.

43:09 Speaker 2: For the actual hybrid calculations, our most recent algorithms make

43:13 Speaker 2: use of a state of the science convolutional neural network

43:16 Speaker 2: or CNN. These types of machine learning algorithms are specifically

43:20 Speaker 2: designed to identify and learn from images, and in this

43:23 Speaker 2: way are highly effective for spatial tasks such as the

43:26 Speaker 2: relationship of geophysical PM twointy five with surrounding predictor variables.

43:31 Speaker 2: We train models separately for each month, which allows us

43:34 Speaker 2: to capture the temporal variability that we would expect relationships

43:38 Speaker 2: between the predictors and the geophysical bias. We additionally maintain

43:42 Speaker 2: an earlier GWR based algorithm, but this is a little

43:45 Speaker 2: bit more of a niche market. It's quite useful for

43:47 Speaker 2: a few particular applications, but in general we recommend the

43:51 Speaker 2: CNN based version. Six. Predictor variables are in line with

43:56 Speaker 2: the GWR discussion we had previously, including satellite based products,

44:01 Speaker 2: CTM simulations and related factors, emissions, meteorology, monitor location details,

44:07 Speaker 2: and geological type of information such as land type and elevation.

44:13 Speaker 2: We spend a lot of time developing cross validation strategies

44:17 Speaker 2: to help us understand the impacts of and ensure an

44:20 Speaker 2: effective model away from ground based monitor locations. These monitors

44:25 Speaker 2: provide an invaluable source of information and constraint, but they

44:28 Speaker 2: also run the risk of fooling us into misrepresenting or

44:31 Speaker 2: misunderstanding the qualities of our data set, especially as we

44:35 Speaker 2: get farther away from these training locations. On the left,

44:39 Speaker 2: here i've show on the effect on our squared of

44:42 Speaker 2: changes to buffer radius during our buffered cross validation analysis.

44:47 Speaker 2: In effect, this can be interpreted is how the agreement

44:49 Speaker 2: of hybrid PM two point five changes with distance from

44:53 Speaker 2: the nearest ground based monitor. For a series of differency

44:56 Speaker 2: and N models that were created during the development of

44:58 Speaker 2: our version six SAT PMD data set, the agreement of

45:03 Speaker 2: a purely geophysical PM twoenty five estimate shown in red,

45:07 Speaker 2: is independent of monitor distance, as shown the flat straight line. Interestingly,

45:13 Speaker 2: all the hybrid models we developed performed similarly at the

45:16 Speaker 2: monitor locations themselves if a buffer were included. However, models

45:21 Speaker 2: that excluded information from a CTM that did not include

45:25 Speaker 2: a physically and chemically meaningful process based understanding the atmosphere

45:30 Speaker 2: fell below the quality of a purely geophysical approach within

45:33 Speaker 2: about fifty kilometers. More advanced models that did include predictors

45:38 Speaker 2: based on CTM output fared better, but again within about

45:42 Speaker 2: one hundred two one hundred and fifty klometers underperformed the

45:45 Speaker 2: results provided by the geophysical PM twoenty five. Our CNN

45:50 Speaker 2: model has been optimized in a way that recognizes the

45:53 Speaker 2: relative strength of the purely geophysical estimates with increasing distance

45:58 Speaker 2: to the training sites. This allows improved agreement compared to

46:02 Speaker 2: the geophysical lessons alone, even at great distance from modern locations.

46:08 Speaker 2: Another way to look at this result is through what's

46:10 Speaker 2: called a shaply additive explanation or SHAP analysis. I'm not

46:14 Speaker 2: going to go through the details of this approach here,

46:16 Speaker 2: but only to say that a SHAP analysis is a

46:19 Speaker 2: method to rank the relative importance of individual predictor variables

46:23 Speaker 2: within a machine learning algorithm. I've shown such a result

46:27 Speaker 2: on the right, with the variables again ranked by their

46:30 Speaker 2: impact on the results of our hybrid CNN. As you

46:34 Speaker 2: can see, the process based predictors and the geophysical PM

46:37 Speaker 2: two pint five itself rank as the most influential predictors,

46:41 Speaker 2: again highlighting the limitations that would be present in a

46:44 Speaker 2: more purely statistical approach that didn't build upon a physically

46:48 Speaker 2: sound framework. And lastly, the results here these figures are

46:54 Speaker 2: briefly cycling through the global monthly hybrid geophysical PM two

46:58 Speaker 2: point five data set that were how to produce and

47:01 Speaker 2: please to make publicly available. The evaluations you see are

47:04 Speaker 2: against the highest possible standards, being both temporarily and spatially buffered.

47:09 Speaker 2: The sat PM two point five data set is particularly

47:12 Speaker 2: well suited for users that want to include the impact

47:14 Speaker 2: of features at a high space resolution and want to

47:17 Speaker 2: ensure consistency both over a long time period but also

47:21 Speaker 2: a distance from ground based constraints. So, in summary, AGAX

47:27 Speaker 2: sat PM two point five offers a global PM two

47:29 Speaker 2: point five data set that uses the geophysical hybrid methodology

47:33 Speaker 2: with multiple satellite based AOD products, an advanced chemical transfer model,

47:38 Speaker 2: and thousands of ground based observations, all within the state

47:41 Speaker 2: of the science convolutional neural network. We focus on consistency

47:46 Speaker 2: and quality, and we're proud to make this whole data

47:48 Speaker 2: set freely and publicly available. To the last section of

47:53 Speaker 2: my part of this talk, I'd like to give a

47:55 Speaker 2: brief walkthrough of the SATPM website and how to access

47:59 Speaker 2: the SATPIM data sets. For that, I'm going to switch

48:02 Speaker 2: over to a demo of the website itself. All right,

48:07 Speaker 2: So the easiest way to access the SATPM two point

48:11 Speaker 2: five data set is via its dedicated website, which can

48:15 Speaker 2: be found at SATPM dot org. The main homepage shown

48:19 Speaker 2: here gives a brief description of the data set with

48:22 Speaker 2: links to many of the associated information sources we've already

48:25 Speaker 2: covered today. One element I particularly like to draw your

48:29 Speaker 2: attention to is this form at the bottom of the page,

48:31 Speaker 2: which allows you to sign up for our mailing list.

48:34 Speaker 2: We primarily use this list to let our users know

48:37 Speaker 2: when an update has been released. Given the active nature

48:40 Speaker 2: of this data set, this happens at least annually is

48:42 Speaker 2: we extend sat PM two point five's time series forward

48:46 Speaker 2: in time, which we also coincide with expanded and updated

48:50 Speaker 2: ground based monitors and any other algorithm developments we've been

48:53 Speaker 2: working on. We really encourage our users to sign up

48:56 Speaker 2: so that they can stay informed about these updates and releases.

49:02 Speaker 2: The menus are fairly self explanatory, with access to our

49:05 Speaker 2: most recent data sets via the data access tab. Here,

49:08 Speaker 2: you'll find our latest global CNN based product V six

49:12 Speaker 2: GLO three, as well as the corresponding GWR based version

49:16 Speaker 2: V five GL six, and while not our focused today,

49:20 Speaker 2: there's also a North American regional CNN based product V

49:23 Speaker 2: six and AO one, which offers a compositional data set

49:26 Speaker 2: over North America following a similar methodology. Selecting any of

49:30 Speaker 2: these such as V six YLO three will then direct

49:33 Speaker 2: you to the product specific page. Here you'll find more

49:39 Speaker 2: detailed information related to this specific version, with a particular

49:44 Speaker 2: focus on any updates compared to the previous version. You

49:47 Speaker 2: also find reference to each data sets associated publication, which

49:51 Speaker 2: will provide you a detailed understanding of the data sets development.

49:55 Speaker 2: A little further down, you'll find details about the format

49:58 Speaker 2: and usage of the data itself, along with with access

50:00 Speaker 2: to the data set. For ease of access, we currently

50:03 Speaker 2: post this data set at both zero point zero one

50:06 Speaker 2: and zero point one degree resolution or about one in

50:09 Speaker 2: ten kilometer, and available from two separate repositories. This courser

50:14 Speaker 2: zero point one degree resolution is much easier for file

50:18 Speaker 2: size to work with and especially for the global files,

50:20 Speaker 2: but the zero point zero one degree files are therefo

50:23 Speaker 2: any projects that require it. For access, the first access

50:27 Speaker 2: point is via box folders using the links provided. To

50:31 Speaker 2: avoid downloading the whole data set, we've parsed the data

50:34 Speaker 2: by regions, say for South America, AF for Africa, GL

50:38 Speaker 2: for Global and A for North America, AS for Asia

50:41 Speaker 2: and EU for Europe. Within these you'll find subjectories for

50:45 Speaker 2: either monthly and or annual mean values, and then the

50:49 Speaker 2: net CDF formatted files themselves.

50:52 Speaker 3: Access is also.

50:53 Speaker 2: Available via the AWS Registry of Open Data via a

50:57 Speaker 2: dedicated S three bucket. Browsing the data via this repository

51:01 Speaker 2: is similar to the box folders, with subdirectories for region

51:04 Speaker 2: and timescale. Data can also be downloaded via the AWS

51:08 Speaker 2: command line interface, with specific instructions given here. Lastly, we

51:19 Speaker 2: recognize that often the questions being asked of our data

51:22 Speaker 2: set don't require access to the full high resolution data set.

51:26 Speaker 2: In fact, it can even be cumbersome to do so.

51:28 Speaker 2: To simplify such applications, we provide a select number of

51:32 Speaker 2: process data sets via the links here at the bottom

51:34 Speaker 2: of each version's page. These links will direct you to

51:38 Speaker 2: process national and regional summaries in a simple text based

51:42 Speaker 2: CSV format. These files can be downloaded and viewed within

51:48 Speaker 2: any number of programs, such as Excel, which I'm bringing

51:51 Speaker 2: onto the screen here. They contain national or subnational population

51:57 Speaker 2: and geographically weighted mean p and two point five concentrations

52:01 Speaker 2: by year covering the entire sat PM time series, as

52:04 Speaker 2: well as a breakdown of the percent of the population

52:06 Speaker 2: above certain thresholds. So if for example, you're interested in

52:12 Speaker 2: how exposures have changed over China for the past fifteen years,

52:16 Speaker 2: you could do a very quick searchdown to where China

52:19 Speaker 2: is being stored here the Chinese data that is, of course,

52:22 Speaker 2: select the last fifteen years or so, as well as

52:26 Speaker 2: the corresponding p twenty five concentrations, and end up by

52:29 Speaker 2: just simply plying that with an insert scatter plot you

52:32 Speaker 2: can see how concentrations have changed or China or any

52:35 Speaker 2: country you're interested in for the past little while. Of course,

52:37 Speaker 2: the exact structure of doing this would depend on your

52:40 Speaker 2: exact software of choice, but it's meant to be a

52:43 Speaker 2: very simple access point for answering these very questions quite quickly. Well,

52:58 Speaker 2: not our main focus, we do offer few tools and

53:00 Speaker 2: examples of accessing and reformatting the data, mainly provided by

53:04 Speaker 2: some of our users, which can help get you started

53:06 Speaker 2: on more advanced applications. The archive in Publications tab provide

53:13 Speaker 2: access to historic versions and to our group's publications related

53:17 Speaker 2: to sat PM two point five, and finally, the support

53:20 Speaker 2: team tab gives you a point of contact if you

53:22 Speaker 2: have any questions need to reach out, as well as

53:25 Speaker 2: specific contact for the members of a group responsible for

53:28 Speaker 2: supporting and overseeing this product. Lastly, if this SATPM product

53:34 Speaker 2: is of interest to you, I'll put in one final

53:36 Speaker 2: plug for joining our mailing list. We don't use extensively

53:39 Speaker 2: and so you won't be fluttered with emails, but it

53:41 Speaker 2: really does allow us to most effectively inform our users

53:44 Speaker 2: when new versions are released. And with that, I'll turn

53:53 Speaker 2: over to Pawan and Junhayung who will be discussing a

53:56 Speaker 2: direct machine learning approach for surface PM point five estimation.

54:15 Speaker 3: Thanks Eren for covering that. Now I'm going to talk

54:19 Speaker 3: about some of the direct machine learning approach to get

54:21 Speaker 3: the PM two point five using satellite data and model outputs.

54:28 Speaker 3: So before I start, I like to introduce three main

54:30 Speaker 3: categories of approaches used to estimate surface PM two point

54:34 Speaker 3: five from satellite data and how direct machine learning fit

54:38 Speaker 3: into that landscape. The first two approach are geophysical and

54:43 Speaker 3: hybrid approach, which we have covered in previous section by Eron.

54:48 Speaker 3: The third is direct mL approach. In this there is

54:52 Speaker 3: no city in the loop. Instead, machine learning models learn

54:56 Speaker 3: the mapping directly from the inputs such as satellite davids

55:01 Speaker 3: and reanalysis variable straight to surface PM two point five.

55:06 Speaker 3: Everything is learned end to end from the data. There

55:09 Speaker 3: are many examples in the community, and we'll explore one

55:13 Speaker 3: of them in here that is marred to CNN.

55:21 Speaker 2: Now.

55:21 Speaker 3: This slide shows how a machine learning based PM point

55:26 Speaker 3: four five product can support applications across three time horizon,

55:32 Speaker 3: the past, the present, and the future. On the left

55:36 Speaker 3: is the historical case answering question like what was PM

55:40 Speaker 3: two point five in twenty ten? Here we use complete

55:44 Speaker 3: retrospective archives, long term satellite data, reanalysis, and ground monitors.

55:51 Speaker 3: These products are ideal for exposure studies, trend analysis, and

55:55 Speaker 3: building decade long records. In the middle one is near

56:01 Speaker 3: real time what is PM to pint five right now?

56:04 Speaker 3: This relies streaming satellite data such as geostationary AEROSO, optical

56:11 Speaker 3: depth and short term metrilogical forecast. These rapid updates supports

56:17 Speaker 3: operational monitoring, wildfire response, dust storms, and public health alerts.

56:23 Speaker 3: And on the right is forecasting what will be PM

56:26 Speaker 3: to point five tomorrow. These models rely only on forecast

56:32 Speaker 3: metrilogy since feature satellite observations are ut available, and emission

56:37 Speaker 3: forecast forecasts can help with planning public alerts and proactive

56:44 Speaker 3: exposer mitigation strategies. Together, these three time horizons shows how

56:51 Speaker 3: machine learning can help us understand past air quality, monitor

56:56 Speaker 3: current conditions, and anticipate future. The typical workflow for machine

57:06 Speaker 3: learning based PM two point five estimation. We start with

57:10 Speaker 3: satellite aerosoloptical depth and reanalysis AEROSO product as the primary input.

57:18 Speaker 3: Then we bring in axillary variables such as humidity, boundary air,

57:24 Speaker 3: high temperature, windland cover, mostly from re analysis or forecasting

57:29 Speaker 3: models like AR five and MIREMERATU. These fields helps the

57:35 Speaker 3: model understand meteorological context. Finally, we use ground based PM

57:40 Speaker 3: to point five observation from networks like air now and

57:44 Speaker 3: open Aqui platform as target variables. All of these inputs

57:49 Speaker 3: feeds into machine learning models such as random forest ex

57:53 Speaker 3: you boost deep learning to produce surface PM to pint

57:57 Speaker 3: five estimates. Here I walked through the full pipeline we

58:06 Speaker 3: used to estimate surface PM to point five from satellite observation.

58:11 Speaker 3: We begin by acquiring data satellite aud and ground based

58:15 Speaker 3: PM to point five measurement. Next, we co located these

58:18 Speaker 3: data sets in time and space to ensure they represent

58:23 Speaker 3: the same atmospheric conditions. Then we build statistical or machine

58:28 Speaker 3: learning models that learn the relationship between aerosoloptical depth and

58:33 Speaker 3: surface PM too point five. Once train, the model estimate

58:37 Speaker 3: PM too point five across the satellite grade, providing continuous

58:41 Speaker 3: special coverage. As an optional final step, we convert those

58:48 Speaker 3: PM to point five values into air quality index categories

58:51 Speaker 3: to support public health communication. This is a reference like

58:59 Speaker 3: showing some example of machine learning use cases. Overall, this

59:04 Speaker 3: progression shows how machine learning is increasingly integrated into both

59:10 Speaker 3: observation driven and model driven PM to point five system

59:14 Speaker 3: improving accuracies across time scales. Here I'm highlighting the key

59:23 Speaker 3: strength and limitation of machine learning based PM to point

59:27 Speaker 3: five estimates. Machine learning approaches are powerful because they can

59:32 Speaker 3: capture nonlinear relationship between aerostoptical depth and PM to point five,

59:39 Speaker 3: incorporate diverse predictors or inputs, and perform well even when

59:45 Speaker 3: emission inventories are spares. They also reduce some systematic biases

59:51 Speaker 3: found in physics based models and allow fast inference one strand. However,

59:59 Speaker 3: there performance depends heavily on the density and quality of

1:00:04 Speaker 3: ground monitors. They may struggle with extreme events if those

1:00:08 Speaker 3: conditions are underrepresented in training. Database Generalization across region and

1:00:17 Speaker 3: time period is sometimes difficult in machine learning approaches. Okay, now,

1:00:23 Speaker 3: let's move on to the examples of machine learning. Deride

1:00:27 Speaker 3: PM to point five data sets called bias corrected MARA

1:00:32 Speaker 3: two PM two point five at our ly scale produced

1:00:35 Speaker 3: as a NASA's hay Cost project. Before we get to

1:00:42 Speaker 3: bias corrected MARAP to PM two point five, take a

1:00:47 Speaker 3: look at comparison of several major global PM too point

1:00:52 Speaker 3: five product. These data sets offer in different temporal coverage

1:00:58 Speaker 3: special resolution in the method used to generate PM two

1:01:02 Speaker 3: point five. For example, deepcamp and HAP combine the satellite

1:01:12 Speaker 3: ground and deep learning approach to provide high resolution hourly

1:01:18 Speaker 3: or daily estimates. Berkeley Earth relies primarily on the ground

1:01:24 Speaker 3: reasurement with interpolation, while the Washington University data sets which

1:01:30 Speaker 3: we just learned about, integrate satellite model and ground input

1:01:35 Speaker 3: at coursal temporal scale. Then, finally, MARA to CNN hay

1:01:41 Speaker 3: Cast PM two point five product, we will learn more

1:01:43 Speaker 3: on it incoming slide. The slide provides some more specific

1:01:52 Speaker 3: details on the MARA to CNN PM two point five

1:01:57 Speaker 3: data sets, although it says only available until twenty twenty four,

1:02:01 Speaker 3: but periodically continue to produce the data and currently as

1:02:06 Speaker 3: we will see, available until twenty twenty five. Temporal resolution

1:02:10 Speaker 3: is one hour with special resolution is about half a degree,

1:02:15 Speaker 3: which is original special resolution of MARA to data product.

1:02:20 Speaker 3: Each daily file contained global data in net city of

1:02:24 Speaker 3: format and the recently publication provides more details on the

1:02:30 Speaker 3: methodology and validation efforts. Each PM two point five estimation

1:02:38 Speaker 3: over each Mara to grid comes with a qaflag. This

1:02:43 Speaker 3: qaflag provides a simple grid level measure of confidence in

1:02:48 Speaker 3: the estimated PM two point five value. It determines by

1:02:52 Speaker 3: factors such as proximity to the surface, monitor line covered,

1:02:57 Speaker 3: type and latitude. Area with more monitoring or a stable

1:03:02 Speaker 3: surface condition generally receives a higher scores. The Quality Assurance

1:03:09 Speaker 3: flag range from one to four, where one represent low

1:03:13 Speaker 3: quality and four represent highest. For most quantitative analysis, we

1:03:22 Speaker 3: recommend using data with a flag value of three or four.

1:03:28 Speaker 3: The map on the right shows how this quality level

1:03:31 Speaker 3: vary across the globe. As you will notice that the

1:03:36 Speaker 3: flags are higher where the ground monitored density is high.

1:03:45 Speaker 3: Now we will get into some details on how these

1:03:50 Speaker 3: PM to point five products are derived. Marito provides PM

1:03:55 Speaker 3: to point five component at surface such as does sulfate, carbon, etc.

1:04:04 Speaker 3: And this equation is often used to combine these components

1:04:09 Speaker 3: to calculate total surface PM two point five. You'll notice

1:04:13 Speaker 3: that it is missing nitrate, which can be significant contributed

1:04:18 Speaker 3: in certain regions. When we will validate these across the globe,

1:04:25 Speaker 3: we found large biases and comparison with individual station, although

1:04:30 Speaker 3: it shows good consistency over larger areas. Here is the

1:04:40 Speaker 3: map showing its global performance against ground monitors in year

1:04:47 Speaker 3: twenty twenty. What we see is that bias pattern varies

1:04:52 Speaker 3: significantly by region and bi aerosol types. In places like California,

1:05:00 Speaker 3: Eastern China, and part of Middle East, marit tend to

1:05:05 Speaker 3: overestimate PM too point five. In contrast, regions such as

1:05:10 Speaker 3: the indogntetic plane and Southern and South America shows under estimation.

1:05:18 Speaker 3: These regions regionally varying biases highlight y machine learning correction

1:05:25 Speaker 3: is so important for producing PM too point five astimates.

1:05:35 Speaker 3: In this slide, we summarize the key mara to input

1:05:39 Speaker 3: be used to model PM too point five by capturing

1:05:43 Speaker 3: the major process that derives aerosol formation, transport and removal.

1:05:51 Speaker 3: On the aerosol side, we include black carbon, organic carbon, dust, sulfate, ESOTO,

1:05:58 Speaker 3: surface mass concentration as well as total extinction at five

1:06:03 Speaker 3: hundred fifteen nanometer that is our alsoloptical app These variables

1:06:09 Speaker 3: help represent both primary and second reparticle sources. On the

1:06:16 Speaker 3: meteorological site, we use surface pressure, humidity, temperature at multiple

1:06:23 Speaker 3: levels and ten meter wind component to capture boundary dynamics

1:06:28 Speaker 3: and atmospheric mixing. All inputs are taken at their native

1:06:34 Speaker 3: merit to spatial resolution and at hourly timing scale, ensuring

1:06:40 Speaker 3: consistent alignment across the data set. Here we describe how

1:06:47 Speaker 3: we use a reference grade PM two point five measurement

1:06:51 Speaker 3: collected from open eq platform as ground truth for training

1:06:56 Speaker 3: and validating our model. We begin by collecting measurement from

1:07:02 Speaker 3: more than five thousand monitoring station worldwide and collocating each

1:07:08 Speaker 3: site with the nearest Marato grid cell. We then temporarily

1:07:13 Speaker 3: match the ely PM to point failue with the crossmending

1:07:17 Speaker 3: Marato inputs. After collocation, we apply a strict quality control

1:07:23 Speaker 3: to remove outliers and invalid records, resulting in more than

1:07:29 Speaker 3: thirty million varied hourly observations. For model development, we used

1:07:37 Speaker 3: year twenty eighteen and nineteen data and for training, and

1:07:43 Speaker 3: for reserved the data from twenty twenty as an independent

1:07:48 Speaker 3: validation period to assess the performance. Here is a full

1:07:56 Speaker 3: workflow used to produce the bias corrected PM to point

1:08:00 Speaker 3: five data sets. As we described earlier, we start with

1:08:04 Speaker 3: especially and temporarily collocating marat meteorological and aerosol input with

1:08:10 Speaker 3: open EQ surface observation. Then we apply five different aerosol

1:08:15 Speaker 3: aware data processing strategies, each capturing PM two point five

1:08:20 Speaker 3: behavior from a slightly different perspective. The final product is

1:08:25 Speaker 3: a single bias corrected PM to point five estimate for

1:08:29 Speaker 3: every single grid cell every hour, with improved accuracies and

1:08:33 Speaker 3: regional consistency. Here is some more details on five day

1:08:41 Speaker 3: different strategies. Each strategy train its own independent convolution neural network,

1:08:49 Speaker 3: but they differ in how aerosol information is organized. For example,

1:08:55 Speaker 3: M one is our baseline model using the full data

1:08:58 Speaker 3: set with no studyfication. M two classified data by dominating

1:09:05 Speaker 3: aerosol species h each other, while M three assigned dominating

1:09:10 Speaker 3: species species by station for the data period. M four

1:09:18 Speaker 3: assign each mirror to grid cell in dominating species across

1:09:22 Speaker 3: the three years data sets, and M find find that

1:09:26 Speaker 3: by merging sulfate and cars aerosol as a single dominating species. Finally,

1:09:35 Speaker 3: the ensemble model, which is M six, combine output from

1:09:38 Speaker 3: all five approaches to produce a single more robust PM

1:09:43 Speaker 3: to pint five estimates. Here we show how the ensemble

1:09:51 Speaker 3: model M six improve PM to pint five estimates compared

1:09:55 Speaker 3: to the RAWMERA to data products. On on the left

1:10:00 Speaker 3: you see the training data which come which has data

1:10:08 Speaker 3: from year twenty twenty eight to twenty twenty nine. The

1:10:12 Speaker 3: red line shows MARA II and the blue line shows

1:10:15 Speaker 3: the M six output, and on the right is the

1:10:19 Speaker 3: validation period which is an INDEPENDENTATA set are coming from

1:10:24 Speaker 3: year twenty twenty two. So we perform this evaluation by

1:10:28 Speaker 3: binning observed PM two pointween two contiles and comparing model

1:10:33 Speaker 3: values across full concentration range. The blue line represent ensemble

1:10:39 Speaker 3: which closely track the oneine one line in both the

1:10:42 Speaker 3: training years and the independent twenty twenty validation journalists within

1:10:48 Speaker 3: the plus minus two sigma certainty weight. In contrast, the

1:10:53 Speaker 3: red line shows that the raw ERA to systematically overestimate

1:10:56 Speaker 3: low PM two point five values and understimate high concentration.

1:11:03 Speaker 3: These concentrations depend bias are clearly visible, and the results

1:11:08 Speaker 3: demonstrate how the ensemble approach using CNN substantially reviews these biases.

1:11:21 Speaker 3: Here we compare regional PM two point five biases in

1:11:27 Speaker 3: raw MARA to product with CNN corrected data. So as

1:11:31 Speaker 3: we have seen earlier. In the top panel, MARA to

1:11:35 Speaker 3: exhibit large regionally consistent error over estimation by more than

1:11:45 Speaker 3: twenty microgram per cubic meter in places like California, China,

1:11:49 Speaker 3: Medalist and understand entition by more than fifteen microgramperm meter

1:11:53 Speaker 3: over endopendetic plane in parts of South America. The bottom

1:11:58 Speaker 3: panel shows how CNN greatly reduce these biases, producing more

1:12:03 Speaker 3: localized and better calibrated pattern across diverse aerosol regimes. CNN

1:12:12 Speaker 3: correct PM two point five also has some limitation and

1:12:15 Speaker 3: it over predicts in eastern Southeast Asia at high concentrations.

1:12:24 Speaker 3: Here is another performance metric showing the performance of CNN

1:12:30 Speaker 3: versus RA MARA two PM two point five. The bar

1:12:34 Speaker 3: shows the medium index of agreement, which measures how well

1:12:38 Speaker 3: the estimations capture both the magnitude and pattern of observed

1:12:42 Speaker 3: pm too point five closer to one is better. Across

1:12:47 Speaker 3: every region, CNN performs sustentionally better. The improvement are especially

1:12:53 Speaker 3: strong in South Asia and Southeast Asia, where agreement nearly

1:12:57 Speaker 3: doubles compared to MARA two. Europe and the United States

1:13:02 Speaker 3: also shows clear gains. The key TAKEABA is that CNN

1:13:06 Speaker 3: more accurately represents both special and temporal variability a PM

1:13:11 Speaker 3: to point five, providing a more reliable global data sets

1:13:15 Speaker 3: than the raw MARA to PM two point five. So,

1:13:21 Speaker 3: in summary, Mara tow CNN hey cast PM too point

1:13:25 Speaker 3: five data sets provides long term global hourly data improve

1:13:32 Speaker 3: PM TOO point five estimation over Mara I bias corrected

1:13:36 Speaker 3: using ensembles, CNN approach and quality AWAID data sets. Now

1:13:45 Speaker 3: we'll look how to access these data sets. The slide

1:13:51 Speaker 3: kind of summarize the different ways in which a user

1:13:55 Speaker 3: can access the PM to point five data sets. The

1:13:58 Speaker 3: data ending page is the best starting point, offering an

1:14:02 Speaker 3: overview of product along with key links. The read we

1:14:08 Speaker 3: file provide detailed documentation on file structure and variable conditions.

1:14:15 Speaker 3: Users who prefer direct download can use the online archive

1:14:20 Speaker 3: or Earth Data search to retrieve files for their specific

1:14:24 Speaker 3: time period. For cloud based workflow. The data set is

1:14:29 Speaker 3: also available on AWS. Open Deep is another option which

1:14:35 Speaker 3: enables remote access through tools like Python, matlib or R

1:14:40 Speaker 3: and Giovanni for an easy browser based interface for subsetting,

1:14:46 Speaker 3: averaging and visualizing the data sets. Here is just screenshot

1:14:53 Speaker 3: of data lending page at NASA disk which hosts the

1:14:59 Speaker 3: pm to my five data sets from this project. Next

1:15:03 Speaker 3: we'll walk through a quick demo of how to access

1:15:06 Speaker 3: these data in NASA Giovanni tool. So now we are

1:15:13 Speaker 3: going to look at an exercise using NASA's Giovanni tool

1:15:20 Speaker 3: to access the merit to CNN PM to PINT five

1:15:24 Speaker 3: data sets and see a couple of example how we

1:15:28 Speaker 3: can access and map and do the time stage type

1:15:33 Speaker 3: of analysis online. So in order to search Giovanni page,

1:15:40 Speaker 3: I would just go to Google and search Giovanni. NASA

1:15:45 Speaker 3: Giovanni will be much better and you will see the

1:15:48 Speaker 3: first link which is NASA Giovanni. I'll just click on that.

1:15:53 Speaker 3: When you first time click on that link, you will

1:15:57 Speaker 3: notice it will ask for a username and passwords. So

1:16:00 Speaker 3: if you do not have Earth data log in user

1:16:04 Speaker 3: name and password, you might want to register and create.

1:16:07 Speaker 3: It's a free If you already have it, you should

1:16:11 Speaker 3: be able to log in. My computer already had so

1:16:14 Speaker 3: it already logged in. And if you do not log

1:16:18 Speaker 3: into the Earth Data log in, then you will have

1:16:22 Speaker 3: limited capabilities from the Giovanni. You wan't to be able

1:16:25 Speaker 3: to do everything. So this is homepage of Giovanni. It's

1:16:31 Speaker 3: a tool which actually allows visualization and analysis of many

1:16:39 Speaker 3: many NASA's data sets. As you can see on the

1:16:43 Speaker 3: left side is the list of all the different platforms,

1:16:48 Speaker 3: variable disciplines the data sets are available. We if you

1:16:55 Speaker 3: actually want to explore, you can click each of these

1:16:59 Speaker 3: taps like measurements and you will see a lot of

1:17:02 Speaker 3: different types of measurements which are listed here. There's like

1:17:06 Speaker 3: five hundred or more. You can also select by platform

1:17:10 Speaker 3: or instruments such as like Modi, is Mara or whatever

1:17:13 Speaker 3: you want. You can also subsidt based on the resolution

1:17:18 Speaker 3: and other things. And then if you go on the

1:17:23 Speaker 3: top that's where the analysis tools are available. So if

1:17:28 Speaker 3: you go to the select part, it will give you

1:17:30 Speaker 3: the list of different type of analysis you can do.

1:17:34 Speaker 3: You can do maps, comparison, time series, histograms, and so

1:17:39 Speaker 3: on and so forth. And this is the selection of

1:17:43 Speaker 3: date starting an end date. And then finally you have

1:17:46 Speaker 3: a geographical region selection and the geographical reasons selection. There

1:17:50 Speaker 3: are various ways we can explore that. First, so since

1:17:55 Speaker 3: we are going to look the MARA to c in

1:17:57 Speaker 3: NPM two point five, instead of going through this whole list,

1:18:01 Speaker 3: I will just try to search here, I'll just say

1:18:05 Speaker 3: CNN and hopefully that should pop up. So here is

1:18:09 Speaker 3: as soon as I do the CNN and do the search.

1:18:12 Speaker 3: The parameter which I'm looking for is called bias corrected

1:18:16 Speaker 3: surface total PM too point five, mass concentration all quality level.

1:18:21 Speaker 3: And then you needs and source and resolutions. And as

1:18:25 Speaker 3: I mentioned earlier, the data is actually now available until

1:18:29 Speaker 3: end of last year twenty twenty five, so we process

1:18:31 Speaker 3: the data frequently. So first thing we are going to

1:18:35 Speaker 3: look we selected the data. Now, first thing we are

1:18:39 Speaker 3: going to do a time averaged map. Okay, so I'm

1:18:43 Speaker 3: going to just select a map type and then I

1:18:47 Speaker 3: will select a date in twenty twenty five last year.

1:18:55 Speaker 3: And for some reason, I'm going to pick August, let's

1:18:59 Speaker 3: say August eight, and that is my starting and then

1:19:06 Speaker 3: same date, I'm going to pick August eighty five, so

1:19:12 Speaker 3: I'm just for quick analysis. I'm just picking one day,

1:19:15 Speaker 3: so it has since this is hourly data, so the

1:19:19 Speaker 3: total data will be twenty four timestamp. And then I

1:19:24 Speaker 3: will select the geographical area. Now, in the geographical area,

1:19:28 Speaker 3: I can actually select a box rectangular box around any

1:19:32 Speaker 3: geographic region, or I can actually select specific countries like

1:19:37 Speaker 3: or many other US states and many other things. So

1:19:41 Speaker 3: for simplicity, we're going to just select a simple box

1:19:47 Speaker 3: over continental northern America. I'm going to select something which

1:19:53 Speaker 3: covers both Canada and US to show a specific case here.

1:19:59 Speaker 3: So and you can if you're walking through with me,

1:20:02 Speaker 3: you can select any region of the world you like

1:20:05 Speaker 3: to select. Once you do that, so now we have

1:20:08 Speaker 3: selected the type of analysis we want to do, which

1:20:13 Speaker 3: is time abridge map. We have selected the dat range,

1:20:15 Speaker 3: we have selected the region of interest, and we have

1:20:18 Speaker 3: selected the variable which we want to plot. Once you've

1:20:21 Speaker 3: done all of that, you just click on that bottom

1:20:24 Speaker 3: right corner plot data. Once you do that, the system

1:20:30 Speaker 3: is going to go and get the data for that

1:20:35 Speaker 3: particular day, all twenty four time of stems and then

1:20:40 Speaker 3: start making plot. So here is the plot. It's very quick.

1:20:47 Speaker 3: You have a color scale here on the right, and

1:20:51 Speaker 3: then you have a map. Now what you see here

1:20:55 Speaker 3: these high values in Canada is actually there huge fire

1:21:02 Speaker 3: during this period. So that is why I selected that

1:21:04 Speaker 3: specific month and day to demonstrate the capture of smoke

1:21:10 Speaker 3: loom within these data sets. So we can make this

1:21:14 Speaker 3: plot look a little bit better. So to do that,

1:21:18 Speaker 3: I will click on the option and under the options

1:21:24 Speaker 3: again there is option I will change the color scales

1:21:29 Speaker 3: to a little bit more, to seventy five, and then

1:21:35 Speaker 3: this palette I can actually select another palette. I'm going

1:21:39 Speaker 3: to select something which is a little bit more. I'm

1:21:46 Speaker 3: going to select this orange from yellow to oranges. And

1:21:51 Speaker 3: then also I'm going to apply some smoothing so that

1:21:55 Speaker 3: the map looks a little bit better, and then the

1:21:59 Speaker 3: projection is fine, and just say replot. Once I do,

1:22:07 Speaker 3: you can see that the color scale has changed the values.

1:22:11 Speaker 3: Since I've only selected the maxim value up to fifty

1:22:14 Speaker 3: of the high values are masked. So let's go back

1:22:17 Speaker 3: and actually revise that and maybe go to hundreds and

1:22:27 Speaker 3: do the replotting again. And you can use the lab

1:22:30 Speaker 3: ramathic scale also if you like to see the plots

1:22:35 Speaker 3: and difference. So now we can see really nice smoke

1:22:40 Speaker 3: prooms from these fires showing high concentration of PM two

1:22:44 Speaker 3: point five here. So this is very simple to use tool.

1:22:49 Speaker 3: You can do this for any parts of the world.

1:22:52 Speaker 3: Now let's try another. We will use the same day

1:22:57 Speaker 3: and we will just do a quick time series over

1:22:59 Speaker 3: this region and see how the PM two point five

1:23:03 Speaker 3: is evolving throughout the day. So to do that, I'll

1:23:08 Speaker 3: just click again bottom right corner which says back to

1:23:12 Speaker 3: the data selection. So once I do that, it goes

1:23:15 Speaker 3: back to my previous selection. Now I can change the

1:23:19 Speaker 3: plot type. I'm going to do time series area averaged, okay,

1:23:25 Speaker 3: and I will keep the day same because I want

1:23:28 Speaker 3: to see the hourly evaluation of the PM two point five.

1:23:32 Speaker 3: But I will just focus on this small area where

1:23:37 Speaker 3: we saw very high concentration of PM too point five.

1:23:40 Speaker 3: Just the small area because we want to see the

1:23:43 Speaker 3: time step evaluation. Right, So again I changed the plot

1:23:48 Speaker 3: type to time series date. I kept the same from

1:23:52 Speaker 3: zero to twenty three hours and the area I have changed. Again,

1:23:56 Speaker 3: I'll say plot the data and since it is averaging

1:24:03 Speaker 3: very small area, so it should be very fast if

1:24:06 Speaker 3: you're doing this over longer a period of time. Many

1:24:10 Speaker 3: many years or larger area, then the tool can take

1:24:15 Speaker 3: much longer time. So now you can see a dinal

1:24:19 Speaker 3: variability on August eight, twenty twenty five, and all the

1:24:27 Speaker 3: times are in UTC, so make sure that you remember

1:24:31 Speaker 3: that you can convert that into local time if you want.

1:24:35 Speaker 3: But you can see the concentrations are very high at

1:24:39 Speaker 3: the start of the day and they started going down actually,

1:24:43 Speaker 3: so again I would suggest you to actually play with

1:24:50 Speaker 3: this tool. It's really nice. I can show you some

1:24:53 Speaker 3: other example which I have run earlier. This is another

1:24:57 Speaker 3: day where we have done the times. There is from

1:25:00 Speaker 3: July first to August July thirty first, and this is

1:25:04 Speaker 3: a different part of the US actually, but it does

1:25:07 Speaker 3: show some increasing trend and again this increase is actually

1:25:12 Speaker 3: resulting from the transport of the smoke in that part

1:25:16 Speaker 3: of the region. This is another one from August. You

1:25:19 Speaker 3: can see from earlier August there were a lot more

1:25:24 Speaker 3: PM to point five concentration and August and this again

1:25:28 Speaker 3: influenced by the those wildfires which we have seen. So

1:25:32 Speaker 3: it's a really nice and easy tool. I would strongly

1:25:38 Speaker 3: recommend you play with different types of analysis you have here.

1:25:42 Speaker 3: You can do the histogram how the PM too point

1:25:46 Speaker 3: five actually looked on that particular area. So it will

1:25:49 Speaker 3: basically take all the data in that box which we

1:25:53 Speaker 3: have selected for that particular day and do a quick

1:25:57 Speaker 3: histogram to understand what kind of values we see. So

1:26:02 Speaker 3: you can see very often the values are very low

1:26:11 Speaker 3: most of the time, but it does shows longer tail

1:26:16 Speaker 3: with values ranging all the two hundred hundred twenty which

1:26:19 Speaker 3: we saw earlier in the map. Also, so you can

1:26:22 Speaker 3: actually do a lot of good things. You can also

1:26:26 Speaker 3: download this as a PNG or net city of files,

1:26:30 Speaker 3: and then there are other options. So I hope this

1:26:34 Speaker 3: tool is useful and next week CALL will actually go

1:26:40 Speaker 3: through this data sets a little bit more and provide

1:26:43 Speaker 3: you more ways in which you can actually analyze this

1:26:46 Speaker 3: data along with the ground data and the satellite based

1:26:49 Speaker 3: PM to point five station. Thank you back to the call.

1:26:58 Speaker 1: So thank you very much doctor Goop and doctor von

1:27:01 Speaker 1: Donklar for those overviews of the data products and their

1:27:04 Speaker 1: associated methodologies. To summarize what we discussed in this part,

1:27:09 Speaker 1: the two data sets we discussed are the Washington University

1:27:13 Speaker 1: in Saint Louis SAT PM two point five data set

1:27:15 Speaker 1: and the NASA bias corrected Merit two PM two point

1:27:18 Speaker 1: five data set. For the SATPM data set, the methodology

1:27:23 Speaker 1: it employs is a hybrid approach where a geophysical estimate

1:27:27 Speaker 1: of PM two point five from model simulations and aerosolo

1:27:30 Speaker 1: optical depth is corrected for biases using ground based measurements

1:27:34 Speaker 1: of PM two point five on a global scale. The

1:27:37 Speaker 1: key strengths of this data set are its fine spatial resolution.

1:27:41 Speaker 1: Gridded products are available at point zero one degree roughly

1:27:44 Speaker 1: one kilometer spatial resolution, as well as having a long

1:27:48 Speaker 1: record going back to nineteen ninety eight and continuing till

1:27:51 Speaker 1: twenty twenty four. The data set has consistent performance even

1:27:57 Speaker 1: far from the location of ground based monitors, as assessed

1:27:59 Speaker 1: through robust spatial cross validation. A possible limitation to consider

1:28:05 Speaker 1: when using this data set are the monthly temporal resolution

1:28:09 Speaker 1: of the products, so this would be unable to resolve

1:28:12 Speaker 1: variability at daily or hourly scales. The data are free

1:28:17 Speaker 1: and openly accessible via the SATPM website. For the bias

1:28:22 Speaker 1: corrected Merit II data set, the methodology used by that

1:28:26 Speaker 1: data set is an ensemble approach where a series of

1:28:30 Speaker 1: convolutional neural network models are developed and calibrated for different

1:28:34 Speaker 1: dominant aerosol conditions, and then a final ensemble model is

1:28:38 Speaker 1: used to select the appropriate model depending on the dominant

1:28:41 Speaker 1: conditions for any location. Key strengths of this data set

1:28:46 Speaker 1: are its high temporal resolution. It's available at hourly temporal resolution.

1:28:51 Speaker 1: It also has a rather long data record from twenty

1:28:54 Speaker 1: to twenty twenty four, and this data set represents PM

1:28:58 Speaker 1: two point five spatial and ten p poral variability much

1:29:01 Speaker 1: better compared to the Merit II Global reanalysis data set

1:29:05 Speaker 1: on which it was based. A possible limitation to consider

1:29:10 Speaker 1: when using this data set is its relatively coarse spatial resolution.

1:29:14 Speaker 1: It's available on the original MERRITI grid, which is a

1:29:17 Speaker 1: point five zero point six two five degree or roughly

1:29:20 Speaker 1: fifty kilometer spatial resolution grid. These data are also free

1:29:25 Speaker 1: and openly accessible via NASA Earth Data and for example,

1:29:28 Speaker 1: you can use tools like Giovanni to do online analysis

1:29:31 Speaker 1: with this data product. In Part three, we'll be looking

1:29:36 Speaker 1: at a case study where we will download and compare

1:29:40 Speaker 1: the sat PM the bias correct Merit two data sets

1:29:43 Speaker 1: to ground based PM two point five measurements for specific

1:29:46 Speaker 1: region and time period of interest. This will demonstrate both

1:29:50 Speaker 1: how to access these data sets. And how to evaluate

1:29:53 Speaker 1: them in a real world setting for practical applications. To

1:29:59 Speaker 1: be sure that you're ready for Part three, we're asking

1:30:02 Speaker 1: you to prepare it advance three different things. First, ensure

1:30:06 Speaker 1: that you have access to Google Collab. This is accessible

1:30:10 Speaker 1: to anyone with a Google account such as a Gmail

1:30:12 Speaker 1: address if you have that. If not, those are free

1:30:16 Speaker 1: to sign up for, and accessing a Gmail address or

1:30:20 Speaker 1: having a Google account will enable you to access Google

1:30:23 Speaker 1: Collab and use it during the training. Second, you'll need

1:30:27 Speaker 1: a account with NASA Earth Data. Again, this is a

1:30:30 Speaker 1: free account. You can sign up for one if you

1:30:32 Speaker 1: don't already have one, and for the training, you should

1:30:35 Speaker 1: have your username and password ready to input into the

1:30:41 Speaker 1: code we're using to allow you to access the NASA data.

1:30:44 Speaker 1: And finally, we ask you to have an open aq

1:30:48 Speaker 1: account to access an open aq API key, which will

1:30:51 Speaker 1: also be using to download data from that global data

1:30:56 Speaker 1: aggregation service that we discussed in part one. If you

1:30:59 Speaker 1: don't have an account, again, that's free to sign up for,

1:31:01 Speaker 1: so you can go to the open Aq website and

1:31:02 Speaker 1: register for an account and they will provide you with

1:31:05 Speaker 1: an API key as a reminder, the homework for this

1:31:11 Speaker 1: training will be issued after the last part of the training,

1:31:13 Speaker 1: which will be Part three next week, and in order

1:31:16 Speaker 1: to receive a certificate for a completion of the training,

1:31:18 Speaker 1: you should attend all three parts of the training through

1:31:22 Speaker 1: the webinars and also complete the homework assignment by the deadline,

1:31:25 Speaker 1: which will be two weeks after the homework is issued.

1:31:30 Speaker 1: This is the contact information for myself and the trainers

1:31:34 Speaker 1: who contribute to this part of the training, as well

1:31:36 Speaker 1: as for the RSET program in general. And here are

1:31:40 Speaker 1: the references included in the material. So thank you all

1:31:48 Speaker 1: very much for your attention, and we'll move on to

1:31:50 Speaker 1: the question and answer portion of the training. All right,

1:31:57 Speaker 1: So as we transition to the Q and A, I

1:32:00 Speaker 1: just want to remind everyone we'll try to answer as

1:32:02 Speaker 1: many questions as possible in the rest of this session.

1:32:06 Speaker 1: We'll be recording all our answers in a document, including

1:32:10 Speaker 1: the questions we don't get to in the live session,

1:32:13 Speaker 1: and we'll be posting that document to the to the

1:32:19 Speaker 1: training website as soon as we've gone through and completed

1:32:23 Speaker 1: answering all the questions. So let me just check. Okay,

1:32:29 Speaker 1: here we go. So the first question is related given

1:32:39 Speaker 1: the non unique relationship between satellite observed AOD and PMT

1:32:42 Speaker 1: one five, Can conditional density estimation models such as mixed

1:32:46 Speaker 1: density networks represent which of the uncertainty threshod exceed the

1:32:50 Speaker 1: probability early warning and so forth? And what validation framework

1:32:54 Speaker 1: is needed to demonstrate spatial transferability to monitor sparse regions.

1:32:59 Speaker 1: We had a couple of questions similar to this, so

1:33:02 Speaker 1: I'll just answer sort of generally here.

1:33:05 Speaker 3: Uh.

1:33:05 Speaker 1: You know, the techniques we've talked about in this training

1:33:08 Speaker 1: are are two of the techniques, but there are you know,

1:33:12 Speaker 1: other techniques such as the one you mentioned, can also

1:33:14 Speaker 1: be applicable, and they may have their own strengths and

1:33:17 Speaker 1: weaknesses which would have to be determined for a given application.

1:33:21 Speaker 1: But in general, the validation framework needed to demonstrate the

1:33:27 Speaker 1: transferability of any methodology, including the methodologies we talked about today.

1:33:32 Speaker 1: Were the methodologies which which Aaron von Donkel are covered

1:33:35 Speaker 1: in his portion. Uh. Those include buffer leave, cluster out

1:33:39 Speaker 1: cross validation to assess spatial generalizability and temporal hold back

1:33:45 Speaker 1: to general to assess the temporal uh generalizability or transferability.

1:33:50 Speaker 1: So I'll just use again, I'll just use that as

1:33:54 Speaker 1: a general opportunity general opportunity to answer questions about UH validation,

1:34:01 Speaker 1: and similarly for a question too. This question is specifically

1:34:03 Speaker 1: about how models trained in a country can be transferred

1:34:06 Speaker 1: to Korea, but I'll use this as an opportunity to

1:34:09 Speaker 1: ask kind of the general question of generalizability and transferability

1:34:14 Speaker 1: of models, including the models presented here. So in this case,

1:34:20 Speaker 1: maybe i'll I'll ask Aaron and then maybe pol On

1:34:23 Speaker 1: to comment on this on the more general question of transferability.

1:34:28 Speaker 1: So Aaron, let's start with you.

1:34:32 Speaker 2: Fuir, So, I mean, in general I would not advise

1:34:35 Speaker 2: using a statistically trained model that was trying to want

1:34:38 Speaker 2: region to go to another. It really depends on the structure.

1:34:40 Speaker 2: Using One of the strengths you know when I discussed

1:34:42 Speaker 2: the geophysical method was it is really designed to try

1:34:46 Speaker 2: to be as globally obflitable as possible. But really, in general,

1:34:53 Speaker 2: you don't want to train a model one addict another,

1:34:54 Speaker 2: meaning in either event. Really, as you know, Carl, answer

1:34:57 Speaker 2: the first question, I mean, good robust validation of any

1:35:01 Speaker 2: model that you want to use is really helpful. In

1:35:03 Speaker 2: such cases. You want to make sure that if you're

1:35:05 Speaker 2: trying to extend the model beyond where it's been trained

1:35:08 Speaker 2: you have some sort of independent data sets to try

1:35:10 Speaker 2: to validate that.

1:35:13 Speaker 1: Okay, thank you, Paul. On anything to add to that or.

1:35:20 Speaker 3: No, I think that is true. Only thing I would

1:35:23 Speaker 3: add is it really depends. There are many things which

1:35:28 Speaker 3: you can consider. One is what kind of PM two

1:35:31 Speaker 3: pint five concentrations and aerosols types are in different regions.

1:35:36 Speaker 3: So let's say you trend a model over US and

1:35:40 Speaker 3: you want to apply over Europe. In some cases it

1:35:43 Speaker 3: will work. But then, as Aaron says, if you have

1:35:46 Speaker 3: if you do not have any ground measurement to validate,

1:35:49 Speaker 3: then it is going to be somebody's guests.

1:35:52 Speaker 1: Right.

1:35:53 Speaker 3: We don't know what is the quality of that. So

1:35:56 Speaker 3: depending on specific situation type of ASO concentration range, meteorological condition,

1:36:04 Speaker 3: geographical terrain, and other kind of look constraint, the model

1:36:10 Speaker 3: may or may not work. So it's it's a really

1:36:13 Speaker 3: case by case determination whether it should, whether it will

1:36:17 Speaker 3: work or not.

1:36:20 Speaker 1: Okay, thank you. So moving on to question three, was

1:36:24 Speaker 1: the minimum exploitable spatial and temporal resolution for estimating PM

1:36:28 Speaker 1: two point five at the scale of a Sephlian city

1:36:31 Speaker 1: like Waga, dougu or Kaya using mayak or Veer's products,

1:36:35 Speaker 1: So the Mayak, the Modus Mayak, and the Veers MAAC

1:36:39 Speaker 1: products specifically are available at a one kilometer spatial resolution,

1:36:43 Speaker 1: so that would be the maximum possible, the best possible

1:36:46 Speaker 1: spatial resolution, and these would be these are both polar

1:36:49 Speaker 1: orbiting satellites, so those would be available at a one

1:36:53 Speaker 1: day temporal resolution. So if you restricted yourself to just

1:36:57 Speaker 1: those sources of information, that would be the the best

1:37:00 Speaker 1: possible spatial and temporal resolution you can get. As discussed

1:37:05 Speaker 1: in the training, there are ways to bring in other

1:37:07 Speaker 1: types of information to get for the biace correcting Mara

1:37:12 Speaker 1: two product, for example, to get the higher tempore resolution.

1:37:17 Speaker 1: So moving on to question four, I think we basically

1:37:21 Speaker 1: answered this. This is kind of a similar to question one,

1:37:23 Speaker 1: So we'll just to sort of reiterate yes, it may

1:37:28 Speaker 1: be applicable and robust validation strategy would be needed to

1:37:32 Speaker 1: assess that. Question five. Are the AODPM two point five

1:37:36 Speaker 1: conversion coefficients already calibrated for West Africa or the the hell?

1:37:40 Speaker 1: Or is it necessary to estimate them locally? Yeah, I'll

1:37:45 Speaker 1: just kind of quickly We talked about that a little

1:37:47 Speaker 1: bit before as well, in the sort of general question

1:37:50 Speaker 1: about generalizability. So, yes, these data sets, in the data

1:37:54 Speaker 1: sets we discussed today include localized calibrations that would make

1:37:59 Speaker 1: these locally relevant. But if you want to create your

1:38:02 Speaker 1: own data set you would have to perform those local

1:38:06 Speaker 1: calibrations yourself. Anything else to add on that from Aaron

1:38:11 Speaker 1: or on?

1:38:14 Speaker 2: No, not, not really, I mean that's yeah. I mean

1:38:16 Speaker 2: one of the reasons we precee these, that's just to

1:38:18 Speaker 2: avoid having to your own calibrations. But there are cases

1:38:21 Speaker 2: string that aren't covered by what we do. In those cases, yes,

1:38:24 Speaker 2: some local calibration can be can be helpful.

1:38:28 Speaker 1: Well I don't have anything, okay, okay, thank you. So

1:38:33 Speaker 1: question six, how can surface PM two point five be

1:38:36 Speaker 1: distinguished from elevated smoke dust or transported aerosol layers when

1:38:40 Speaker 1: they produce similar AOD values? This is a good question.

1:38:43 Speaker 1: We alluded to this a little bit in Part one,

1:38:46 Speaker 1: and then in Part two you kind of saw the

1:38:47 Speaker 1: specific strategies that the two data sets we talked about use. So,

1:38:53 Speaker 1: in particular, the SATPM product uses the geos Kim simulation

1:38:57 Speaker 1: to simulate the vertical profiles of aerosols, and MERIT two

1:39:01 Speaker 1: uses the go Kart simulation. The Aerosol Simulation module, and

1:39:07 Speaker 1: there are also observational based methods to look at, especially

1:39:12 Speaker 1: the vertical distribution of aerosol from either ground or space

1:39:15 Speaker 1: based light ar observations which can provide the vertical distribution

1:39:19 Speaker 1: of atmospheric particles. Question seven, how is the correlation between

1:39:24 Speaker 1: AOD and maybe PM two point five at shorter temporal

1:39:29 Speaker 1: scales such as days or weeks. I'd say there's no

1:39:33 Speaker 1: sort of general answer to this question. It can depend

1:39:35 Speaker 1: on a lot of factors. For example, you know, averaging

1:39:38 Speaker 1: over longer periods of time could average out some noise

1:39:43 Speaker 1: in the signal, But then you're also looking at changing

1:39:46 Speaker 1: meteorological conditions, and as we discussed in part one, those

1:39:49 Speaker 1: meteorological conditions such as humidity of mounted boundary layer height

1:39:54 Speaker 1: can have a big impact on the AOD to PM

1:39:56 Speaker 1: two point five relationships. So looking at a longer period

1:40:00 Speaker 1: of time or a shorter period of time may may

1:40:03 Speaker 1: not help. Yeah, I'll maybe leave it there. Question eight,

1:40:11 Speaker 1: how can I obtain MYCAOD data to estimate PM two

1:40:14 Speaker 1: point five? We put some links there for you to

1:40:17 Speaker 1: access those data. They're available through NASA's Earth Data site,

1:40:20 Speaker 1: and once we post this document, to the website, you'll

1:40:23 Speaker 1: be able to access those links directly, but in general,

1:40:26 Speaker 1: if you go to the NASA Earth Data website and

1:40:30 Speaker 1: search for the data product, you should be able to

1:40:32 Speaker 1: find information on how to download it. To the method

1:40:36 Speaker 1: question nine, to the methodologies in estimating pm slash aerosol

1:40:40 Speaker 1: apply to all other air pollutants, So that's a good question.

1:40:46 Speaker 1: While this specific methodology wouldn't directly apply, there are certainly

1:40:50 Speaker 1: ways to modify the methodology or use similar methodologies for

1:40:54 Speaker 1: other pollutants. Aaron, maybe just briefly comment on your group's

1:40:59 Speaker 1: work there, sure so.

1:41:01 Speaker 2: I mean, as Carl mentioned, a lot of the concepts

1:41:03 Speaker 2: become the same, a lot of the details become slightly different.

1:41:06 Speaker 2: So N two is when our group has worked with

1:41:07 Speaker 2: a little bit. You know, the idea is very similar.

1:41:10 Speaker 2: You have a satellite retrieval of total column N two,

1:41:14 Speaker 2: you can relate that to near service N two using

1:41:18 Speaker 2: a chemical transfer model, and you can use a statistical

1:41:22 Speaker 2: framework to further refine those estimates. So by broad strokes,

1:41:25 Speaker 2: it can be very similar. But a lot of the

1:41:28 Speaker 2: details are often quite different and the unique challenges that

1:41:31 Speaker 2: come with that, so there can be a lot of

1:41:33 Speaker 2: work to just to copy that those concepts over to

1:41:36 Speaker 2: a different species, but the concepts to broadly hold.

1:41:41 Speaker 1: Okay, thank you. Question ten. Can MLR model for PM

1:41:46 Speaker 1: two point five bias correction be applied in regions with

1:41:49 Speaker 1: limited ground based monitoring stations like North Africa? How do

1:41:52 Speaker 1: you handle missing data? Again? I think, Aaron, if you

1:41:58 Speaker 1: want to comment on this, but we've sort of touched

1:42:00 Speaker 1: on this already with some of the previous questions about generalizability.

1:42:03 Speaker 1: But if you have anything else to add or Pollen,

1:42:05 Speaker 1: if you have anything else to add.

1:42:10 Speaker 2: Not dramatically beyond what was covered in the presentation. I

1:42:13 Speaker 2: think this question might have come in a bit earlier

1:42:15 Speaker 2: on in the presentation, So hopefully we've covered some of

1:42:17 Speaker 2: this during our talk. But I mean probably a lot

1:42:20 Speaker 2: of the methods that we're covering here were designed specifically

1:42:23 Speaker 2: to be more globally applicable trying to handle missing data,

1:42:26 Speaker 2: fill in missing data, relate those relationships or feel those

1:42:30 Speaker 2: relationships over broader regions as well, So they're inherently trying

1:42:34 Speaker 2: to account for a lot of those challenges. But again,

1:42:38 Speaker 2: each region is slightly unique.

1:42:43 Speaker 1: Okay, thank you. Question eleven, which is recommended for bright

1:42:49 Speaker 1: sparsely vegetated surfaces. I interpreted this as a question about

1:42:54 Speaker 1: the AOD products, and in general, we tend to think

1:42:57 Speaker 1: the deep blue AOD algorithm is better for those types

1:43:00 Speaker 1: of surfaces. And I believe, Aaron, you did cover that

1:43:04 Speaker 1: in slides twenty eight through thirty four the presentation, where

1:43:07 Speaker 1: there's some examples of comparing the performance of the different

1:43:11 Speaker 1: algorithms in different regions. I believe this question did come in,

1:43:14 Speaker 1: you know, before you got to those slides, So hopefully

1:43:18 Speaker 1: by referring to those slides, this person can get an

1:43:20 Speaker 1: answer to their question question twelve. Okay, this is pretty specific.

1:43:26 Speaker 1: When evaluating interurban environmental exposure inequality using the version six

1:43:31 Speaker 1: Global product TATPM product in dense megacities, we notice the

1:43:36 Speaker 1: gridded product heavily flattened spatial gradients, i e. Very low

1:43:40 Speaker 1: spatial coefficient of variation across districts compared to localized ground

1:43:46 Speaker 1: monitors that show very large microclimatic and seasonal peaks. How

1:43:51 Speaker 1: do you recommend researchers to account for or bias correct

1:43:54 Speaker 1: this localized spatial smoothing when conducting fine scale subdistrict exposure

1:43:59 Speaker 1: and Genie coefficient analyses. So, Aaron, I think that's a

1:44:03 Speaker 1: question for you.

1:44:04 Speaker 2: Sure. So, I mean there can be a number of

1:44:06 Speaker 2: factors affecting the scatter that you're seeing between the local monitors.

1:44:10 Speaker 2: I think Carl showed a slide last week that that

1:44:12 Speaker 2: really highlighted the amount of variability within monitors within a

1:44:15 Speaker 2: given city, how much monitors can vary even within a

1:44:18 Speaker 2: one kilometer distance. So it's worth braining mind that the

1:44:21 Speaker 2: estimates we're providing our average over one kilometer. Beyond that,

1:44:25 Speaker 2: of course, it is challenging to capture these fine scale features.

1:44:30 Speaker 2: One of the things within the soldering my presentation is

1:44:33 Speaker 2: a lot of different resolutions of data sets to go

1:44:35 Speaker 2: into it.

1:44:36 Speaker 3: And so.

1:44:38 Speaker 2: As much as these statistcal models are attempting to account

1:44:41 Speaker 2: for this fine scale of variability, there are a lot

1:44:45 Speaker 2: of resolutions going on, and so we tend to think

1:44:47 Speaker 2: of it as the one klumeter product may not fully

1:44:51 Speaker 2: capture those one those subparometer features, even one klmeter features fully,

1:44:55 Speaker 2: and so we typically recommend averaging over slightly larger areas. Dramatically,

1:45:01 Speaker 2: you know, up to ten colmers you can see some

1:45:03 Speaker 2: pretty dramatic improvement, although in my experience a lot of

1:45:05 Speaker 2: health studies, people tend to move around, and so I

1:45:09 Speaker 2: tend to think this is not necessarily for most applications

1:45:13 Speaker 2: a severe limitation, but that may depend on the particular

1:45:16 Speaker 2: study you have in mind.

1:45:19 Speaker 1: Okay, thank you. Question thirteen. When we build local machine

1:45:25 Speaker 1: learning models or downscale or validate satellite estimates with limited

1:45:30 Speaker 1: ground data, our leave one station out spatial cross validation

1:45:34 Speaker 1: yields a near zero spatial R squared even if temporal

1:45:38 Speaker 1: validation is high. What are your recommended best practices for

1:45:42 Speaker 1: validating spatial models and establishing local confidence in satellite derive

1:45:46 Speaker 1: PM two point five when ground stations are so severely limited.

1:45:50 Speaker 1: So I would actually say that you're probably doing pretty

1:45:55 Speaker 1: much the best you can. So the leave one station

1:45:57 Speaker 1: out is probably a good cross validation method for you,

1:46:01 Speaker 1: especially considering a condition of of sparse ground monitoring stations.

1:46:06 Speaker 1: If you had more of a dense network, then probably

1:46:08 Speaker 1: a buffered leave one station out or a buffered clustered

1:46:14 Speaker 1: leave a cluster out basically method would be probably more recommended.

1:46:18 Speaker 1: But in the case of not so dense ground measurements

1:46:21 Speaker 1: that there's probably not much statistical difference between these different

1:46:25 Speaker 1: cross validation techniques so yeah, I would say, there's there's

1:46:29 Speaker 1: no really no simple answer here. You're I think the

1:46:33 Speaker 1: believe one station out strategy is probably a good strategy

1:46:36 Speaker 1: for your case. And the fact that may just be

1:46:39 Speaker 1: that the methodology you're looking at is not able to

1:46:42 Speaker 1: give you a good spatial generalization. There may just be

1:46:46 Speaker 1: too much variability, or you're you're you're missing an information

1:46:50 Speaker 1: on which would explain that variability. And then my my

1:46:56 Speaker 1: other recommendation here would be, in such situations where ground

1:47:01 Speaker 1: monitoring is very sparse, it may actually be better to

1:47:04 Speaker 1: use a pre existing product like the saven PM product

1:47:08 Speaker 1: or the bias corrected Barriti product that we talked about today,

1:47:11 Speaker 1: and use your limited ground based monitoring information purely for

1:47:16 Speaker 1: evaluation of the quality of that existing product. Oh and sorry,

1:47:21 Speaker 1: I see polland stands up.

1:47:22 Speaker 3: Go ahead, Well, no, go ahead, finish your question. Answer.

1:47:25 Speaker 3: I wanted to just make an announcement.

1:47:27 Speaker 1: Okay, yeah, so yeah, I think that's basically what I

1:47:31 Speaker 1: want to say. So I would say the most efficient

1:47:34 Speaker 1: use of such sparse data is it would really be

1:47:36 Speaker 1: instead of trying to develop your own model, use that

1:47:39 Speaker 1: data to evaluate the models that are already out there

1:47:42 Speaker 1: and see which one is the most appropriate for your setting. Okay,

1:47:46 Speaker 1: go ahead, Polan, Yeah.

1:47:47 Speaker 3: I just want to, I think, put a node out there.

1:47:51 Speaker 3: I know several of you are trying to access the

1:47:54 Speaker 3: merit to CNN using geo Wanni and you're not able

1:47:59 Speaker 3: to find the data. And I was not aware, but

1:48:03 Speaker 3: I just found out that Giovanni team is moving that

1:48:08 Speaker 3: data sets to the cloud. So for time being, they

1:48:12 Speaker 3: have actually taken it off, so it will take about

1:48:17 Speaker 3: ten days for them to move it. So after July

1:48:20 Speaker 3: twenty fourth, the data should be available back in the

1:48:24 Speaker 3: Giovanna So we apologize for this inconvenience. I think we

1:48:30 Speaker 3: didn't plant ahead in time, but hopefully the data sets

1:48:34 Speaker 3: will be available soon back in the system and you

1:48:38 Speaker 3: will be accessed.

1:48:41 Speaker 1: Okay, yeah, so thank you for that announcement, Pollen. We'll

1:48:46 Speaker 1: maybe see I don't think that should affect your the

1:48:51 Speaker 1: homework that we're assigning for this. We're not using Giovanni

1:48:53 Speaker 1: specifically for the homework assignment, but maybe we'll just double

1:48:58 Speaker 1: check that that won't affect people's ability to complete the

1:49:02 Speaker 1: homework assignment. But yeah, thank you for that announcement. All right,

1:49:07 Speaker 1: So question fourteen, what proportions of nitrate dominated urb in

1:49:11 Speaker 1: PM two point five are formed from local emissions generated

1:49:14 Speaker 1: and surrounding regionals or transported across national boundaries, and how

1:49:17 Speaker 1: do these change under different meteorological and pollution conditions. So

1:49:21 Speaker 1: that's a very complicated question. You may be able to

1:49:24 Speaker 1: use something like a chemical transport model to get an

1:49:27 Speaker 1: answer to that question. It's really a little bit beyond

1:49:30 Speaker 1: the scope of today's training, so we won't go into

1:49:33 Speaker 1: more detail than that, but feel free to maybe contact

1:49:36 Speaker 1: us or first explore that training that we put a

1:49:39 Speaker 1: link to, and if that doesn't answer your question, you

1:49:41 Speaker 1: might need to contact us for further discussions. Question fifteen,

1:49:47 Speaker 1: how well does the vertical scaling of GEOSCM capture highly

1:49:50 Speaker 1: localized boundary layer dynamics and low altitude micro level emission

1:49:55 Speaker 1: sources in extremely dense megacities And are their future efforts

1:49:59 Speaker 1: to integrate local high resolution meteorology to improve this? So, Aaron,

1:50:03 Speaker 1: I think you spoke to this a little bit in

1:50:05 Speaker 1: your previous answers, but if you want to add anything

1:50:07 Speaker 1: onto this, go ahead, sure.

1:50:09 Speaker 2: I mean, so it's a good question, and I mean

1:50:12 Speaker 2: I like the idea of trying to integrate higher resolution

1:50:14 Speaker 2: meteorology to improve local conditions. There has been some work

1:50:21 Speaker 2: done on like ship emissions representing subgrid variation stuff. I'm

1:50:24 Speaker 2: not aware of any of that specifically looks at megacities

1:50:27 Speaker 2: for now. Really, what we're relying on is more of

1:50:30 Speaker 2: the statistical models to capture those any shortcomings within just

1:50:35 Speaker 2: came to capture those really high resolution features that maybe

1:50:39 Speaker 2: may come about because of those features such as megacities.

1:50:43 Speaker 2: It's it's pretty good, we think, but it's certainly an

1:50:47 Speaker 2: actual physically based model would be even better if we

1:50:50 Speaker 2: could to pull it off.

1:50:53 Speaker 1: Okay, thank you. Question sixteen. For developing countries which cannot

1:50:58 Speaker 1: afford ground measurements and they have a little to no

1:51:00 Speaker 1: ground stations and there's limited technical capacity for advanced machine learning,

1:51:05 Speaker 1: what would you recommend as the ideal data and method

1:51:07 Speaker 1: to analyze air pollution there? So, I think that's, you know,

1:51:11 Speaker 1: partial part of the motivation for this training is the

1:51:14 Speaker 1: fact that both the SATPM and the bias corrected MERRIT

1:51:17 Speaker 1: two data sets are global, provide global coverage, and are

1:51:21 Speaker 1: publicly freely available everywhere in the world. The reason, but

1:51:26 Speaker 1: you know, the reasoning behind that is exactly to provide

1:51:29 Speaker 1: this kind of information to those who have more limited resources.

1:51:33 Speaker 1: So using these data sets that are already available to you,

1:51:38 Speaker 1: and wherever possible, comparing them to whatever local information you

1:51:41 Speaker 1: do have to establish their data quality would be kind

1:51:46 Speaker 1: of my recommendation for the best place to start your analysis. Okay,

1:51:54 Speaker 1: So for question seventeen, given random cafolds, spatially buffered and

1:51:58 Speaker 1: temporal holdout cross byalidation, which of these gives the most

1:52:01 Speaker 1: reliable estimate of error when the model is applied to

1:52:04 Speaker 1: unmonitored locations in future data? So I think the recommendation

1:52:08 Speaker 1: of best practice here is to use both a spatially

1:52:12 Speaker 1: buffered cross validation to assess the spatial generalizability, that is,

1:52:16 Speaker 1: the applicability to unmonitored locations, and a temporal holdout strategy

1:52:21 Speaker 1: to assess the temporal generalizability, so the applicability to future data.

1:52:28 Speaker 1: Question eighteen, will the retrievals be biased over airnet locations?

1:52:33 Speaker 1: I'm not sure if this refers to the aerosoloptical depth

1:52:36 Speaker 1: retrievals or the PM two point five data sets, but

1:52:39 Speaker 1: I think in either case there really shouldn't be any

1:52:43 Speaker 1: kind of bias specific to the airnet stations. In either

1:52:48 Speaker 1: of these methodologies. Question nineteen for deep investigations about the

1:52:55 Speaker 1: problems and possible solutions for Shalian countries and the Sahelian

1:53:01 Speaker 1: Alliance countries, what kind of partnership is possible for NASA

1:53:06 Speaker 1: R SET if the University of Thomas Sunkar University. So

1:53:10 Speaker 1: two points here. First, our set is a training program.

1:53:13 Speaker 1: We don't really do our own research and investigations. That's

1:53:17 Speaker 1: sort of something that's done separately by others at our set.

1:53:19 Speaker 1: But with that being said, you know, the people who

1:53:22 Speaker 1: participate in this training, myself and doctor von Donkler and

1:53:25 Speaker 1: doctor Gupta are involved in research around the world. So

1:53:30 Speaker 1: if you have specific questions or you know specific research

1:53:34 Speaker 1: topics which you think you want to collaborate on, you

1:53:37 Speaker 1: can reach out to us directly about that. Question twenty.

1:53:43 Speaker 1: Since process based constraints are interpolated smoothly from the course

1:53:48 Speaker 1: fifty kilometer model, what limitations or what are the limitations

1:53:52 Speaker 1: of using the downscaled one kilometer product for subgrid, micro

1:53:55 Speaker 1: environment or local buffer analyses in a highly complex urban

1:53:59 Speaker 1: area is Again I guess this is a question for

1:54:02 Speaker 1: Aaron and anything to add, you know, beyond what you've

1:54:05 Speaker 1: already said basically.

1:54:09 Speaker 2: No, I mean, it's a great question, as I mentioned before,

1:54:12 Speaker 2: I mean, actually one thing, I'm really glad to hear

1:54:14 Speaker 2: that you're thinking this way, whoever asked us questions, is

1:54:16 Speaker 2: a great way to think about the limitations of the model,

1:54:19 Speaker 2: and it really highlights why we talk about the value

1:54:22 Speaker 2: you know, as much as possible average to a larger area.

1:54:24 Speaker 2: So as much as we do have stysical models and

1:54:26 Speaker 2: things trying to capture that fine scale resolution down to

1:54:29 Speaker 2: one kilometer, it's quite reasonable to assume that those really

1:54:32 Speaker 2: really fine scale features will not be fully represented and

1:54:35 Speaker 2: they'll higher uncertainties. So it's a really good question we

1:54:40 Speaker 2: are relying on the statistical models to capture that really

1:54:43 Speaker 2: fine scale and how things will vary over that really

1:54:46 Speaker 2: really fine scale at least, how that relationship will over

1:54:48 Speaker 2: fine scale.

1:54:52 Speaker 1: Okay, thank you. Question twenty one, Again, this was a

1:54:57 Speaker 1: pretty detailed specific question and a bit beyond the scope

1:55:00 Speaker 1: of this training. I think this is sort of a

1:55:01 Speaker 1: whole research topic in and of itself, so we can't

1:55:04 Speaker 1: really provide a simple answer to it, unfortunately. So question

1:55:09 Speaker 1: twenty two, to what extent does visitor traffic inside and

1:55:12 Speaker 1: closed museum environments contribute to indoor PM two point five. Again,

1:55:16 Speaker 1: indoor PM concentrations is not really the be on the

1:55:19 Speaker 1: scope of this training, and I don't think it's really

1:55:21 Speaker 1: in any of our expertise unfortunately, so we can't really

1:55:26 Speaker 1: comment on that in a very informed way. So let's

1:55:29 Speaker 1: go to question twenty three next, scroll down a little bit.

1:55:37 Speaker 1: There we go, due to the lack of ground based

1:55:40 Speaker 1: measurements in Africa, would you expect there to be significant

1:55:43 Speaker 1: uncertainties in the derived PM two point five concentrations there

1:55:46 Speaker 1: from these data sets? Would you expect similar performance or

1:55:50 Speaker 1: performance to be similar, better or worse than merit to

1:55:53 Speaker 1: over Africa? So I think, yeah, maybe to Aaron and

1:55:57 Speaker 1: then to poll on about regionally specific differences in uncertainties

1:56:04 Speaker 1: and the products.

1:56:06 Speaker 2: Sure, so yeah, sorry, As the question you know quite

1:56:10 Speaker 2: rightly implies Africa is a certainly going to be a

1:56:13 Speaker 2: more uncertain region. There are some really there's some challenging

1:56:17 Speaker 2: conditions for models to capture both from major dust sources,

1:56:20 Speaker 2: major byomass bringing sources, and both these methods and painals

1:56:23 Speaker 2: weak because I think we are trying our best over

1:56:25 Speaker 2: those regions to speak to some of the strengths of

1:56:27 Speaker 2: the set PM data set. We really have that observational

1:56:31 Speaker 2: constraint from satellite, very just hardwired in what we're doing.

1:56:35 Speaker 2: The methods are designed in our by design using sparse

1:56:41 Speaker 2: ground based measurements as much as we have available to

1:56:43 Speaker 2: try and calibrate that better. It's definitely a region of

1:56:47 Speaker 2: more uncertainties, but I would still argue that what we're

1:56:50 Speaker 2: producing is probably one of the better estimates that is

1:56:53 Speaker 2: available out there.

1:56:57 Speaker 1: Okay, I just want to.

1:57:01 Speaker 3: Add some the Africa, as Aaron mentioned, is always very

1:57:07 Speaker 3: interesting reason and there's a lake up ground measurement. We

1:57:10 Speaker 3: all know about it, right, so it's very very hard

1:57:13 Speaker 3: to actually assess what kind of uncertaintives we might have.

1:57:17 Speaker 3: But I just want to let people know that there

1:57:19 Speaker 3: have been a lot of efforts going on in Africa

1:57:23 Speaker 3: in deploying low cost census although we know they are

1:57:27 Speaker 3: not best to use for this kind of validation, but

1:57:31 Speaker 3: they do provide useful information, and there are several groups

1:57:35 Speaker 3: in Africa and from other parts of the world who

1:57:39 Speaker 3: are actually trying and deploying and collecting the data. So

1:57:42 Speaker 3: we are hoping to actually collaborate with these groups to

1:57:46 Speaker 3: validate our products in future and provide more accurate assessment

1:57:52 Speaker 3: as we move forward in processing more data.

1:57:58 Speaker 1: Okay, thank you. So question twenty four was a question

1:58:03 Speaker 1: about actually like future funding opportunities, which is again not

1:58:06 Speaker 1: really something we're qualified to speak to, but I encourage

1:58:09 Speaker 1: you to take a look at the nspire's website, which

1:58:12 Speaker 1: is a place where all NASA funding opportunities are posted

1:58:16 Speaker 1: for you to look at. So question twenty five, in

1:58:19 Speaker 1: my own PM two point five machine learning framework SHAP

1:58:22 Speaker 1: analysis ranked my separate satellite AOD retrieval as the weakest

1:58:26 Speaker 1: predictor compared to continuous meteorological and spatial features. Have you

1:58:31 Speaker 1: observed similar trends where the model downweights physical AOD in

1:58:36 Speaker 1: favor of net data proxies. So I'll just speak maybe

1:58:41 Speaker 1: to my own experience, I would say that this is

1:58:44 Speaker 1: certainly possible in specific conditions that especially where PM two

1:58:50 Speaker 1: point five might be highly meteorology dependent. There's also the

1:58:54 Speaker 1: possibility that there are multiple layers of aerosols in the

1:58:57 Speaker 1: atmosphere and so the total AOD is really not reflective

1:59:00 Speaker 1: of the surface concentration very well. And also especially if

1:59:04 Speaker 1: you smoke, if you focus your analysis on a very

1:59:06 Speaker 1: small region where the meteorology is similar, and the PM

1:59:09 Speaker 1: two point five is also similar. Statistical methods might just

1:59:13 Speaker 1: be able to pick out on similar patterns and meteorology

1:59:16 Speaker 1: and AOD and made the connection between them much more

1:59:20 Speaker 1: easily than if you tried the same methodology on a

1:59:23 Speaker 1: larger region or even on a global scale, where the

1:59:26 Speaker 1: relationship between AOD and PM two point five is is

1:59:29 Speaker 1: much less consistent, and therefore other information sources like AOD

1:59:33 Speaker 1: would have a bigger role to play. Paul On or

1:59:37 Speaker 1: Aaron if you want to add on to that.

1:59:41 Speaker 2: No, no, I think you covered it. Sorry, go ahead,

1:59:43 Speaker 2: paul On.

1:59:44 Speaker 3: No, No, I was just saying, good thing.

1:59:46 Speaker 2: Well, you covered it very nicely, Carl.

1:59:48 Speaker 1: Okay, thanks, Okay. Question twenty six, how do we ensure

1:59:52 Speaker 1: the data is used accurately? Okay, this is the pretty

1:59:56 Speaker 1: broad question, but at a simple level, I'd say, you know,

1:59:59 Speaker 1: the this course, for one, is trying to convey to

2:00:04 Speaker 1: you basically what the limitations and the capabilities are of

2:00:09 Speaker 1: the data. So keeping those in mind as you use

2:00:12 Speaker 1: the data will help you to use it accurately. And

2:00:16 Speaker 1: as I think we've said several times, wherever possible, performing

2:00:19 Speaker 1: your own local assessments with locally collected data is really

2:00:24 Speaker 1: the best way to establish the quality of these data

2:00:26 Speaker 1: products for your specific region of interest. So anything to

2:00:32 Speaker 1: add on to that, I'll ho or eron.

2:00:36 Speaker 3: I can add one thing about specifically related to the

2:00:41 Speaker 3: statistical modeling or machine learning method. One thing which is

2:00:46 Speaker 3: very very critical whenever we do the statistical or machine

2:00:49 Speaker 3: learning method is the data distribution. So if you have

2:00:52 Speaker 3: several input parameter and you're trying to map with the

2:00:55 Speaker 3: output PM two point five, do check how the data

2:01:00 Speaker 3: has distributed. Because machine learning algorithm and statistical methods often

2:01:07 Speaker 3: try to learn what is around the mean and the

2:01:12 Speaker 3: extreme values become actually has less impact on the relationship

2:01:18 Speaker 3: which are which is formed, and often you will notice

2:01:21 Speaker 3: that these methods either underestimate the very high values because

2:01:26 Speaker 3: there are very few data points in your training and

2:01:29 Speaker 3: overestimate the low values because that's again very small values

2:01:34 Speaker 3: in your OWLD data set. So the data distribution, checking

2:01:37 Speaker 3: and adjusting your validation data versus the training data can

2:01:42 Speaker 3: be an important aspect of preparing your data before you

2:01:48 Speaker 3: start modeling. Whether you use a multiple linear regression or

2:01:52 Speaker 3: linear regression or machine learning method, it applies to all

2:01:56 Speaker 3: of those.

2:01:57 Speaker 1: Yeah, think that's a yeah, that's a really good point

2:02:01 Speaker 1: of the extreme values, especially if you have an a

2:02:05 Speaker 1: special interest in the extreme values as opposed to the mean,

2:02:08 Speaker 1: then that's that's an important caveat to keep in mind. Okay, yeah,

2:02:14 Speaker 1: let's maybe move on to the next question. Let's scroll

2:02:18 Speaker 1: down a little bit to question twenty seven. Will the

2:02:23 Speaker 1: mL based estimate capture events such as dust storms and

2:02:28 Speaker 1: smoke from fires? So, powin, do you want to speak

2:02:31 Speaker 1: to this a little bit?

2:02:37 Speaker 3: Sorry I was muted. I think it will capture the

2:02:45 Speaker 3: mara to as written here in the response, does assimilate

2:02:52 Speaker 3: satellite based aerosol optical depth and the fire emissions. And

2:02:58 Speaker 3: if that event is captured by the satellite is specifically

2:03:02 Speaker 3: from the modies, then the merit should have it captured

2:03:06 Speaker 3: and the machine learning based bias corrections should actually capture it.

2:03:10 Speaker 3: In some cases it might if the machine learning training

2:03:17 Speaker 3: didn't have dust or a smoke type of cases, then

2:03:22 Speaker 3: it might miss. So there are certain cases where smoke

2:03:26 Speaker 3: is very thick or dust plume is very thick, and

2:03:29 Speaker 3: the satellite sometimes get confused whether it is dust or

2:03:33 Speaker 3: cloud and mask it as a cloud and it does

2:03:35 Speaker 3: not get propagated into the assimilation stream and that can

2:03:40 Speaker 3: actually be not seen in the MARI too. So I

2:03:43 Speaker 3: would say it veries case by case, but in general

2:03:47 Speaker 3: it should be able to capture qualitatively. Quantitatively things can

2:03:52 Speaker 3: be very different, and depending on which event and how

2:03:56 Speaker 3: far we are looking from the location of the source,

2:04:00 Speaker 3: the quantitative numbers can be different. Qualitative lids should be captured.

2:04:06 Speaker 1: Okay, thank you. Question twenty eight, what are the main

2:04:11 Speaker 1: challenges in making PM two point five products like the

2:04:15 Speaker 1: Questionington University SAT PM two point five or the Meritus

2:04:18 Speaker 1: and then hay Cass product from retrospective processing to near

2:04:23 Speaker 1: real time to delivery. So this is a question about

2:04:25 Speaker 1: like the data latency and how maybe how that might

2:04:29 Speaker 1: be reduced in the future. So I guess yeah, Aaron, Aaron,

2:04:32 Speaker 1: go ahead, sure.

2:04:34 Speaker 2: So, as I wrote, I mean, that's a great question.

2:04:36 Speaker 2: We'd love to be able to get data sets, data

2:04:38 Speaker 2: sets out even faster. It's unforced, its often not practical.

2:04:42 Speaker 2: There is you know, as I've written here, as you

2:04:44 Speaker 2: gather from the presentation, is an awful lot of different

2:04:46 Speaker 2: data sets that come together to make these final PM

2:04:49 Speaker 2: twoint five data sets possible. From biomass burning emission inventories,

2:04:54 Speaker 2: which seems to be brought you brought into the kepital

2:04:56 Speaker 2: trust models models needs to be run. One very practical

2:05:01 Speaker 2: example of a delay that we face is to get

2:05:03 Speaker 2: the highest quality ground based monitors. Often governmental agencies will

2:05:08 Speaker 2: perform their own internal checks for usually about six months.

2:05:11 Speaker 2: So really for next year's release, you know which we're

2:05:14 Speaker 2: playing later this year, we're just starting to work with

2:05:17 Speaker 2: the ground based monitors now, which are the highest reference

2:05:20 Speaker 2: grade model that as well as our own internal validations

2:05:26 Speaker 2: take time. It's really multifaceted why it takes us so long,

2:05:29 Speaker 2: but I really feel that the time gives us an

2:05:33 Speaker 2: opportunity to reach the highest quality data product that we can.

2:05:38 Speaker 1: Okay, thank you, I on it.

2:05:41 Speaker 3: I think it applies to the matra TO or any

2:05:45 Speaker 3: other satellite product in the same ways because they all

2:05:48 Speaker 3: depends on some ancile data sets. Specifically in case of

2:05:53 Speaker 3: matter TO, it assimilates to so many data sets, so

2:05:56 Speaker 3: to get all those data prepared that for assimilation some time,

2:06:01 Speaker 3: and MERI to is re analysis is always I think

2:06:05 Speaker 3: several weeks behind because it's a re analysis. The idea

2:06:10 Speaker 3: is to retrospective analysis. So I think that is why

2:06:14 Speaker 3: we want to produce the best availabled data sets, and

2:06:18 Speaker 3: there are some real time and forecasting more data products

2:06:22 Speaker 3: which have a different way of processing.

2:06:27 Speaker 1: Okay, thank you, so question twenty nine this refers to

2:06:32 Speaker 1: a specific slide where we're talking about the index of

2:06:34 Speaker 1: agreement and asking about some of the other metrics. So

2:06:37 Speaker 1: some of the other slides did have different metrics, especially bias.

2:06:41 Speaker 1: And then also the paper describing the data set itself,

2:06:45 Speaker 1: which we'll link here obviously has a much more detailed

2:06:49 Speaker 1: description of the performance. Okay. Question thirty is this at

2:06:58 Speaker 1: PM data set available for North Africa? Yes, it is,

2:07:00 Speaker 1: and was the spatial and temporal resolution it's point zero

2:07:03 Speaker 1: one degree spatial end monthly ten or a resolution. We

2:07:06 Speaker 1: covered that in the presentations. Question thirty one, okay. Another

2:07:10 Speaker 1: regional question. In Indonesia, the days with the worst air pollution,

2:07:15 Speaker 1: for example hayes from fires are often the same days

2:07:18 Speaker 1: covered by clouds, so these days get removed from the

2:07:23 Speaker 1: data the AOD data. Specifically, how do we fix PM

2:07:26 Speaker 1: two point five estimates so they don't end up too

2:07:28 Speaker 1: low because of this? Aaron, maybe I think you answered

2:07:32 Speaker 1: this question here, so.

2:07:34 Speaker 2: Sure, I mean, I'll speak to how we address that

2:07:36 Speaker 2: within the SATPM data set, and maybe Poland has more

2:07:39 Speaker 2: to add after within sat PM. This is we account

2:07:43 Speaker 2: for these sampling bias in a few different ways. First off,

2:07:47 Speaker 2: the ground based monitors themselves we okay you said. First off,

2:07:52 Speaker 2: the sat AOD we directed to a monthly mean based

2:07:55 Speaker 2: upon sampling adjustment factors provided by the CTM, so again

2:07:59 Speaker 2: using that chepical transt model to infer how the days

2:08:03 Speaker 2: we observe relate to the days we haven't observed, and

2:08:05 Speaker 2: then providing a more representative overall monthly mean. Subsequent to that,

2:08:10 Speaker 2: the ground based monitors themselves will sample irrespective of that

2:08:13 Speaker 2: cloud cover. So when we apply those machine learning models

2:08:17 Speaker 2: or GBR models as a subsequent calibration to the geophysical estimates,

2:08:24 Speaker 2: that inherently also corrects for any sampling biases that may

2:08:28 Speaker 2: be from limitations to the satellite retrievals themselves. So there's

2:08:31 Speaker 2: a few ways in which this is hopefully addressed. None

2:08:35 Speaker 2: of them are obviously perfect, but there should not be

2:08:38 Speaker 2: inherently a sampling bias due to cloud covered days in

2:08:41 Speaker 2: our product.

2:08:44 Speaker 3: Okay, yeah, for Merit too, I think as it is

2:08:49 Speaker 3: a model produced data, so the data should be available

2:08:52 Speaker 3: even if there is our cloud. Now, in terms of

2:08:55 Speaker 3: the bias, I think the bias can arise is because

2:08:58 Speaker 3: if the satellite a woods are missing, then the model

2:09:03 Speaker 3: is relying completely on its own capability. I don't think

2:09:08 Speaker 3: we do any specific correction to actually account for that

2:09:15 Speaker 3: sampling bias there, but we haven't seen any clear evident

2:09:21 Speaker 3: in all validation studies that it should it should have

2:09:25 Speaker 3: a biases.

2:09:28 Speaker 1: Okay, thank you, So question two question thirty two, rather

2:09:33 Speaker 1: as a follow on, I think asking about which of

2:09:36 Speaker 1: these satellites would work best for Indonesia where there are

2:09:40 Speaker 1: cloud covered So just in terms of the this is

2:09:44 Speaker 1: maybe a little bit more general, but a geostationary satellite

2:09:47 Speaker 1: like Kimawari offering more frequent observations would give you more

2:09:52 Speaker 1: opportunities for cloud free observations. However, you know, any of

2:09:57 Speaker 1: these instruments would be I think similarly affected by the

2:10:01 Speaker 1: presence of clouds. And I also just want to note

2:10:06 Speaker 1: we've gone a little bit over our allotted time. If

2:10:10 Speaker 1: it's all right with with Aaron and Pallen, we might

2:10:13 Speaker 1: keep going for maybe ten more minutes and answer some

2:10:18 Speaker 1: more questions and then everything anything. We don't get to

2:10:21 Speaker 1: live we can answer in this document offline. Yeah, so

2:10:28 Speaker 1: let's keep going for a little bit longer. Question thirty four,

2:10:33 Speaker 1: you mentioned that multiliner regression is used to predict you

2:10:36 Speaker 1: a physical bias in PM two point five. Could you

2:10:38 Speaker 1: specify which independent variables or predictors are typically used in

2:10:42 Speaker 1: the set pm MLR model. So, I think to save time,

2:10:47 Speaker 1: we'll just sort of refer you to the publication where

2:10:49 Speaker 1: these are listed, and then in general they fall into

2:10:52 Speaker 1: the categories of physical and process based parameters, emissions, meteorology,

2:10:58 Speaker 1: and site location. Believe Aaron covered this in the slides

2:11:02 Speaker 1: as well. So question thirty five, within the workflow of

2:11:07 Speaker 1: the six ensemble predictive model developed for the Maritu sine

2:11:11 Speaker 1: n Heyksta, is that what is the scientific rationale for

2:11:15 Speaker 1: training five distinct convolitional neural network strategies rather than relying

2:11:20 Speaker 1: on a single baseline strategy basically the M one model. So,

2:11:26 Speaker 1: poland if you want to, yeah, speak to that freequently.

2:11:28 Speaker 3: Sure. Sure. So. I think, as we have discussed and

2:11:35 Speaker 3: seen in many cases, that the relationship between input parameter

2:11:42 Speaker 3: and the output target variable PM too point five is

2:11:45 Speaker 3: not linear and it varies as a function of aerosols type, meteorology,

2:11:52 Speaker 3: range of Pm two point five and many other parameters.

2:11:56 Speaker 3: And our goal was to what best we can do right,

2:12:00 Speaker 3: So in order to actually get the best of the

2:12:04 Speaker 3: data sets, we divided the data into different nominating aerosol

2:12:11 Speaker 3: component time so that we can get a better sense

2:12:14 Speaker 3: of how the input to output mapping can be done

2:12:18 Speaker 3: with more reliability. And that is why we actually try

2:12:24 Speaker 3: to play with different schemes and then we came up

2:12:28 Speaker 3: with these five different model which provide us an optimized solution.

2:12:35 Speaker 3: It may not be the best, but it provided an

2:12:37 Speaker 3: optimal solution which worked globally. And then the ensemble kind

2:12:42 Speaker 3: of bring them together in a more unified consistent data

2:12:45 Speaker 3: sets across the globe. So to use the strength of

2:12:51 Speaker 3: narrow too and their aerosol component, use those different approach

2:12:56 Speaker 3: to get an optimized global Team nine five data sient.

2:13:04 Speaker 1: All right, thank you. Question thirty six Quick clarification Slide

2:13:10 Speaker 1: sixty three notes that the mL generalization across region and

2:13:14 Speaker 1: time isn't guaranteed well. Slide seventy eight cosmeritive product robust

2:13:18 Speaker 1: onscing regions in years, given that it's trained on twenty

2:13:22 Speaker 1: eighteen to twenty nineteen dow date on twenty twenty. Could

2:13:24 Speaker 1: you help me understand what a bust onceing readings and

2:13:27 Speaker 1: years is based on Powen if you want to, Yeah.

2:13:31 Speaker 3: So, I think we continue to evaluate this data, although

2:13:36 Speaker 3: in those years which we have mentioned in the those

2:13:40 Speaker 3: are based on the published work which we did, but

2:13:44 Speaker 3: we continue to evaluate and I think we haven't seen

2:13:47 Speaker 3: any degression in the model performance over even beyond twenty

2:13:53 Speaker 3: twenty one and twenty twenty four, which we have done recently,

2:13:56 Speaker 3: which is not part of this paper. So I think

2:14:02 Speaker 3: what we can say is robust over time now over

2:14:07 Speaker 3: the region. The problem is the same as we have

2:14:10 Speaker 3: been discussing in earlier question is that if they're not

2:14:13 Speaker 3: ground stations available over certain regions, then it's very very

2:14:16 Speaker 3: hard for us to actually even predict what kind of

2:14:21 Speaker 3: uncertainties or error might be there. And that is why

2:14:25 Speaker 3: we try to provide a quality assurance flat with the

2:14:28 Speaker 3: data sets which basically rely on the ground density of

2:14:35 Speaker 3: the ground measurements and few other parameters which we are

2:14:39 Speaker 3: important in obour training algorithm, and that gives you an

2:14:43 Speaker 3: idea about whether a product is reliable in certain revision

2:14:47 Speaker 3: or not. So I think the quality assurance is one

2:14:51 Speaker 3: way to kind of see whether the product is robost

2:14:55 Speaker 3: or not.

2:14:58 Speaker 1: Okay, thank you. Question thirty seven, how do do you

2:15:02 Speaker 1: phittic physical models handle sudden extreme local emissions like crop

2:15:06 Speaker 1: burning or wildfires, especially when satellite AAD observations are missing

2:15:11 Speaker 1: due to cloud cover, Aaron, if you want to add

2:15:14 Speaker 1: anything on that.

2:15:16 Speaker 2: Sure mean this probably ties a little bit back into

2:15:18 Speaker 2: the earlier question why it takes so long to put

2:15:21 Speaker 2: these estimates together each time. So within the geophyscal model,

2:15:25 Speaker 2: it's obviously driven by a chemical transport model, and so

2:15:29 Speaker 2: in our case, and we use daily biomass burning emissions,

2:15:33 Speaker 2: and so those would capture within those daily inventories. They

2:15:38 Speaker 2: specific events like a sudden residual burning or wildfires, And

2:15:42 Speaker 2: those emission inventories themselves aren't necessarily dependent on the same limiting,

2:15:47 Speaker 2: don't have the same dependence sort of limitations, don't have

2:15:49 Speaker 2: the same limitations as the AOD retrieval. So a number

2:15:53 Speaker 2: of those daily biomass burning inventories can work as fine

2:15:57 Speaker 2: even when there is some cloud cover. Now that's not

2:16:00 Speaker 2: to say they're perfect, of course, I know crop resume

2:16:02 Speaker 2: burning has received a little bit more attention in a

2:16:04 Speaker 2: few papers I've read recently, thing that can be a

2:16:06 Speaker 2: higher uncertainty. But for the geophysical lestments, they're trying to

2:16:10 Speaker 2: within those daily inventories. And one of the reasons we

2:16:13 Speaker 2: of course enhance the geophysical lestments using the hybrid approach

2:16:17 Speaker 2: is to try to account for any misunderstandings or underrepresentations

2:16:22 Speaker 2: that occur within those inventories. So that's kind of how

2:16:25 Speaker 2: we we account for it.

2:16:30 Speaker 1: Okay, thank you. And question thirty eight another question about

2:16:35 Speaker 1: Indonesia due to the persistent cloud cover, which for the

2:16:39 Speaker 1: machine learning methods, which meteorological variables or predictors are the

2:16:44 Speaker 1: most crucial to include so the model doesn't become biased

2:16:47 Speaker 1: in a human tropical tropical region that experiences extreme humidity

2:16:50 Speaker 1: fluctuations during hotspot seasons. I guess Fallen, if you want

2:16:54 Speaker 1: to offer suggestions there or this might just be something

2:17:01 Speaker 1: that would need to be studied, but go ahead, go ahead, Yeah,

2:17:05 Speaker 1: let me.

2:17:06 Speaker 3: Just read the tropical system cloud co throat the US

2:17:14 Speaker 3: most moder Yeah. This this is a real problem, I guess.

2:17:20 Speaker 3: And it is not only over Indonesia.

2:17:23 Speaker 1: Uh.

2:17:24 Speaker 3: This is if you look across the ITCZ, which is

2:17:29 Speaker 3: has a consistent clouder cover even over South America or Africa.

2:17:35 Speaker 3: Part of that, and satellite the current passive satellite observation

2:17:42 Speaker 3: of aerosols will always have that problem because if there

2:17:47 Speaker 3: are cloud, the signal is so dominated by the cloud

2:17:50 Speaker 3: we cannot really get anything any information about aerosols. So

2:17:54 Speaker 3: I don't think there is any observational solution there. We

2:17:58 Speaker 3: will have to rely on the model for that, and

2:18:02 Speaker 3: more we can improve on the model, I think the

2:18:04 Speaker 3: better answer we can get. The other option is rely

2:18:08 Speaker 3: on ground measurement and active satellite based observation, which are

2:18:14 Speaker 3: a few of them right now in orbit and the

2:18:16 Speaker 3: few morees coming in the future which should be able

2:18:19 Speaker 3: to provide some more details or aerosol information in presence

2:18:24 Speaker 3: of cloud, and once they get assimilated into models, we

2:18:28 Speaker 3: will have a better idea what is happening in those reasons.

2:18:34 Speaker 1: Okay, thank you, And then I also saw another question

2:18:38 Speaker 1: just came in which was also referring to Indonesia. I

2:18:42 Speaker 1: think you have probably answered answered this in your response

2:18:46 Speaker 1: as well. About the general challenges so well, I think we'll,

2:18:52 Speaker 1: you know, we'll read through these and make sure that

2:18:54 Speaker 1: we've captured all of them. But I think for the

2:18:56 Speaker 1: sake of time, we're gonna I think call it call

2:19:01 Speaker 1: an end to the Q and A here. So thank

2:19:04 Speaker 1: you very much doctor Owen Dunkelar and doctor Kupta for

2:19:08 Speaker 1: contributing your time and your expertise to this training. And

2:19:12 Speaker 1: thank you very much to our attendees for your interest

2:19:16 Speaker 1: and all these very interesting and insightful questions about how

2:19:20 Speaker 1: to use these data sets appropriately. As a reminder, the

2:19:25 Speaker 1: next and final part of the training will be next week,

2:19:29 Speaker 1: and to prepare yourself with that training, you should test

2:19:33 Speaker 1: out that you're able to access Google Collab the sorry

2:19:40 Speaker 1: the have an Earth Data login password for accessing the

2:19:44 Speaker 1: NASA Earth data and also have an account with open

2:19:47 Speaker 1: Aq to access the open aq data set via their API,

2:19:53 Speaker 1: and there'll be some links to those. The slides will

2:19:57 Speaker 1: have links to those when they're posts the training website.

2:20:02 Speaker 1: So yeah, again, thank you everyone, and we will see

2:20:05 Speaker 1: you next week.