Introduction to a Machine Learning Model to Estimate Water Quality Parameters Part 2

0:01 Speaker 1: Hello everyone, Welcome back to this art set training on

0:05 Speaker 1: monitoring water quality in lakes and coastal regions using Stream.

0:10 Speaker 1: Today it's part two on introduction to a machine learning

0:14 Speaker 1: model to estimate water quality parameters based on satellite observations.

0:19 Speaker 1: We have a guest speaker today, doctor Ryan O'she and

0:22 Speaker 1: he will be talking about the machine learning model. Last

0:27 Speaker 1: week in part one, we saw background overview and demonstration

0:32 Speaker 1: of stream webtool and API, where Stream is satellite based

0:36 Speaker 1: TWOL for rapid evaluation of aquatic environment. We had a

0:40 Speaker 1: guest speaker mister William Wainwright. He showed us that Stream

0:44 Speaker 1: uses Landset eight and nine and Sentinel two A, B

0:48 Speaker 1: and C to obtain water quality parameters which include chlorophylea concentration,

0:53 Speaker 1: total suspended solids and seeky disk depths. The area covered

0:58 Speaker 1: our coastales and inland lakes in the US. The data

1:03 Speaker 1: are available at twenty to thirty meters special resolution, and

1:07 Speaker 1: we saw that stream API allows search and download of

1:11 Speaker 1: multiple images for an area of interest. We saw example

1:18 Speaker 1: of selection of water quality parameters using stream map and

1:23 Speaker 1: make time series of water quality parameters using qgis example

1:29 Speaker 1: shown here. We saw Pyramid lake and this map shows

1:32 Speaker 1: chlorophil a concentration. High resolution data that shows pattern and

1:39 Speaker 1: how chlorophyll a is distributed within the lake and in

1:42 Speaker 1: different parts that can be seen here. We downloaded multiple

1:45 Speaker 1: images of chlorophil a concentration as well as TSS into

1:50 Speaker 1: supeag bay and made time series using QGIS. Today now

1:57 Speaker 1: we will focus on introduction to machine learning model to

2:00 Speaker 1: estimate water quality parameters based on satellite observations. This is

2:05 Speaker 1: the model that's used to derive what quality parameters in stream.

2:11 Speaker 1: Just a note here there is a homework posted today

2:14 Speaker 1: on our website and it will be due on tenth

2:17 Speaker 1: of March and the certificate of completion will be ordered

2:20 Speaker 1: to those who attend all live sessions and complete the

2:24 Speaker 1: homework assignment before the given due date. So we'll start

2:31 Speaker 1: with today's session. Our objectives today are that by the

2:38 Speaker 1: end of part two, you'll be able to become familiar

2:42 Speaker 1: with a mixture density network or MDN based model used

2:46 Speaker 1: for deriving water quality parameters from satellite observations in stream.

2:52 Speaker 1: Recognize how to select and process satellite images for deriving

2:56 Speaker 1: water quality parameters in a water body of interest. Recognize

3:00 Speaker 1: how to access the md and model code for deriving

3:03 Speaker 1: water quality parameters in the water body of interest. Here's

3:07 Speaker 1: the outline for today. We will start with a very

3:10 Speaker 1: brief review of preprocessing satellite images for inputs to the

3:14 Speaker 1: MDN based model. Then we will have an overview of

3:18 Speaker 1: the md AND based model which is used in stream,

3:22 Speaker 1: and we'll have a demonstration of access and application of

3:25 Speaker 1: the md AND model code for deriving water quality parameters

3:29 Speaker 1: from satellite data. Our guestpicker dotor Osha will use examples

3:34 Speaker 1: to demonstrate how to use MDN code. He will use

3:38 Speaker 1: Sentinel three OGI images and Paysoci images to show water

3:43 Speaker 1: quality parameters er from this model in San Francisco Bay

3:47 Speaker 1: and Lake. Erie will have a brief review of preprocessing

3:52 Speaker 1: of satellite images. Now, satellite sensors measure top of atmosphere

3:58 Speaker 1: or TA radiance, and these radiances result from a combination

4:04 Speaker 1: of surface and atmospheric conditions, including the effects of clouds

4:07 Speaker 1: and aerosolt particles as seen in this figure. Toa radiances

4:13 Speaker 1: as a function of wavength are shown here and contribution

4:17 Speaker 1: is pretty large from atmosphere and also from surface, including

4:22 Speaker 1: this green curve that is contribution from water surface. So

4:27 Speaker 1: this water living reflectance that depends on BAX scattering and

4:31 Speaker 1: absorption of radiation due to materials in water that is

4:34 Speaker 1: optically active. As shown here that the water living reflectancies.

4:39 Speaker 1: They depend on what is going on in the water

4:43 Speaker 1: and these are the reflectancies they are used to estimate

4:48 Speaker 1: water quality parameters. To get water living remote sensing reflectancies

4:54 Speaker 1: or rrs at different wavelengths, we have to start with

4:57 Speaker 1: satellite observations of TOA radiances and apply atmospheric correction schemes

5:03 Speaker 1: to get rrs's. Now Here, there are a number of

5:06 Speaker 1: schemes or atmospheric correction algorithms listed. NASA has a couple

5:12 Speaker 1: NASA Ocean Biology Processing Group algorithm and NASA AKMA Verse

5:16 Speaker 1: algorithm which is used for stream. There is echo, light

5:20 Speaker 1: and polymer. These are some of the schemes generally used

5:23 Speaker 1: by the community to do atmospheric correction and get remote

5:28 Speaker 1: sensing reflectancies. So we would start with satellite images which

5:32 Speaker 1: is level one top of atmosphere reflectance and then find

5:37 Speaker 1: level two remote sensing reflectancies. Then they're used to estimate

5:42 Speaker 1: water quality parameters. So to summarize to get remote sensing reflectancies,

5:51 Speaker 1: will select a satellite and sensor and start with level

5:55 Speaker 1: one satellite data, then use appropriate atmospheric correction algorithm to

6:01 Speaker 1: derive level to water living reflectancies and then those can

6:06 Speaker 1: be input to the md and model that we will

6:09 Speaker 1: see in a few minutes. Here we're talking about NASA

6:14 Speaker 1: c Earth Atmosphere Data Analysis System or CEDS has Ocean

6:18 Speaker 1: Color Science Software or oCSS W that can help. That

6:23 Speaker 1: has algorithm for atmospheric correction and can be used to

6:26 Speaker 1: get level two remote sensing reflectancies. There are a couple

6:31 Speaker 1: of HORSET trainings they demonstrate how to use CIDs O

6:36 Speaker 1: c ss W to do atmospheric correction and get remote

6:39 Speaker 1: sensing reflectancies. So you can refer to these for more information.

6:44 Speaker 1: And c dos A is an open source software from

6:48 Speaker 1: NASA and can be downloaded and you can find information

6:52 Speaker 1: from NASA Earth Data and search with CDs and you

6:56 Speaker 1: can get more information or you can refer to these

7:00 Speaker 1: trainings for more details. We saw this table last week

7:06 Speaker 1: all the satellites and sensors which are used for water

7:10 Speaker 1: quality monitoring, and here are the sources where you can

7:15 Speaker 1: start with level one data. You can obtain level one

7:19 Speaker 1: satellite images from these satellites and so corresponding links are

7:24 Speaker 1: provided here. These are the data portals where you can

7:27 Speaker 1: download images from these satellites and sensors. With that, we'll

7:35 Speaker 1: start with today's session on overview and demonstration of MDN

7:39 Speaker 1: model for stream Just a note here, please put your

7:44 Speaker 1: questions in the questions box and we will letice them

7:47 Speaker 1: at the end of the webinar, and feel free to

7:50 Speaker 1: enter your questions as we go. We will try to

7:52 Speaker 1: get to all the questions during the Q and A session.

7:55 Speaker 1: After the webinar, the remainder of the questions will be

7:59 Speaker 1: answered in the Q and A document which will be

8:02 Speaker 1: posted on the training website about a week after the training.

8:08 Speaker 1: With that, I'll introduce our speaker for today, doctor Ryan O'she.

8:13 Speaker 1: Doctor Ryan O'she is a senior research scientist at NASA's

8:17 Speaker 1: Goddart Space Flight Center. Ryan received his PhD in Mechanical

8:23 Speaker 1: and Oceanographic engineering from the Massachusetts Institute of Technology and

8:28 Speaker 1: Woodshol Oceanographic Institution joint program in twenty twenty one while

8:33 Speaker 1: he was under a National Defense Science and Engineering Graduate fellowship.

8:38 Speaker 1: Doctor O'she's thesis, titled Computational Approaches for Submitter Ocean Color

8:43 Speaker 1: Remote Sensing, focused on the theoretical limitations, practical considerations, and

8:49 Speaker 1: potential future avenues for deploying lightweight hyperspectral cameras for ocean

8:55 Speaker 1: color remote sensing from low altitude platforms. Currently, Ryan is

8:59 Speaker 1: working on defining freshwater and ocean water products at NASA Garden.

9:06 Speaker 1: So we invite doctor O'sha to present about the md

9:11 Speaker 1: AND model used in Stream. Ryan.

9:17 Speaker 2: Everyone, my name is Ryan O'Shea. I'm a member of

9:19 Speaker 2: the Freshwater Sensing program at Science Systems and Applications, Inc.

9:24 Speaker 2: This program consists of myself, A Rin Sarenathen will Waynewright,

9:29 Speaker 2: and Brandon Smith. It was formerly headed by Nima Polophon,

9:33 Speaker 2: who's moved on to an SHQ. Today, I'll be presenting

9:36 Speaker 2: on monitoring water quality and lakes and coastal regions using Stream.

9:41 Speaker 2: Part two focuses on introduction to a machine learning model

9:44 Speaker 2: to ask me water quality parameters based on satellite observations.

9:49 Speaker 2: On the slide, I'm going to briefly take you through

9:51 Speaker 2: freshwater and littoral imaging from space, showing the mixture density

9:56 Speaker 2: network supported sensors in red or just supported by the

10:01 Speaker 2: MDN in blue, they're supported by stream and ingreen. They're

10:04 Speaker 2: coming student to stream. In two thousand, or actually the

10:10 Speaker 2: end of nineteen ninety nine, Modus was launched. Modus as

10:14 Speaker 2: a multi spectral instrument which can capture a kilometer scale

10:17 Speaker 2: resolution with a near one to two day temporal resolution.

10:24 Speaker 2: This satellite censors lasted for the last twenty six years,

10:28 Speaker 2: so it can produce time series and climatology of given regions.

10:34 Speaker 2: This is supported by the MDN VERS was also launched

10:38 Speaker 2: and I believe twenty twelve. It achieves about seven hundred

10:41 Speaker 2: and fifty meters spatial resolution and again one to two

10:44 Speaker 2: day spatial revisit. It has very similar capabilities as Modus,

10:49 Speaker 2: and we have models that support both of these and

10:51 Speaker 2: show generally pretty harmonized retrievals from both. We also support

10:57 Speaker 2: lance at eight and nine. These are the mos that

11:00 Speaker 2: you saw as part of Part one that are included

11:02 Speaker 2: on stream. These sensors have fewer spectral bands but much

11:10 Speaker 2: finer spatial resolution up to about twenty meters spatial scales

11:14 Speaker 2: that we can start to look at fine scale inland

11:17 Speaker 2: water bodies using these sensors. All supported on stream or

11:22 Speaker 2: Sentinel two A and two B with similar spectral resolution

11:25 Speaker 2: as landset and similar spatial resolution However, the costs for

11:30 Speaker 2: both of these sensors is temporal resolution, with Landset having

11:35 Speaker 2: a temporal resolution of about eight days, where Sentinel has

11:39 Speaker 2: a resolution of about five days. However, when you start

11:43 Speaker 2: to combine them together, you can improve the temporal resolution.

11:48 Speaker 2: So we already have Landset eight and Sentinel two supported

11:51 Speaker 2: on stream, but coming sooner stream will also be adding

11:53 Speaker 2: in Sentinel three A and Sentinel three B, for which

11:56 Speaker 2: we already have make sure density network models available. These

12:00 Speaker 2: have temporal resolutions of about one to two days, higher

12:04 Speaker 2: spectral resolutions which allow for improved product accuracy in these

12:08 Speaker 2: optically complex inland and coastal waters that we're sensing in,

12:14 Speaker 2: and spatial resolutions of about three hundred meters. So by

12:17 Speaker 2: combining the lands At eight, Sentinel two and Sentinel three

12:20 Speaker 2: imagery we can get higher temporal, spatial and spectral combinations.

12:26 Speaker 2: As additional sensors are launched, will also be adding these

12:29 Speaker 2: into stream. We also have mature densitly not network models

12:35 Speaker 2: available for commercial satellites like Planets super Dove, which achieves

12:40 Speaker 2: meter scale spatial resolution. We also support hyper spectral instruments,

12:47 Speaker 2: so we develop models for Hiko the hyperspectual imagery for

12:50 Speaker 2: the coastal ocean as well as PRISMA and EMIT. We're

12:55 Speaker 2: currently funded by EMMA as well as the PACE Science

12:58 Speaker 2: and Applications Team to further improve our products for hyperspectral imagery.

13:03 Speaker 2: The PACE model that we have available supports a wide

13:09 Speaker 2: range of products leveraging the full hyperspectral data, so not

13:14 Speaker 2: only does it support the standard color fol, TSS and

13:16 Speaker 2: c DOWN products, but also inherent optical properties like absorption

13:20 Speaker 2: due to phytoplankton, and PACE allows for one to two

13:25 Speaker 2: day temporal resolution and about a kilometer spatial resolution. We're

13:30 Speaker 2: also funded as part of the Lancet Science team to

13:33 Speaker 2: make models in advance of Landset Next, which should be

13:36 Speaker 2: launched around twenty thirty one. So this large set of

13:43 Speaker 2: sensors that we have developed models for are radiometrically capable

13:47 Speaker 2: for sensing in aquatic regions. The combined observations can get

13:52 Speaker 2: us towards daily observations if we're able to leverage estimates

13:56 Speaker 2: from all of these sensors combined. But they have different

13:58 Speaker 2: spatial and spectual cap abilities which allow us to sense

14:02 Speaker 2: in a wide range of spatial resolution region so we

14:06 Speaker 2: can sense an inland as well as coastal waters. The trick, however,

14:11 Speaker 2: is combining these estimates from these different sensors to get

14:14 Speaker 2: a better understanding of global inland and coastal waters. To

14:18 Speaker 2: do that, we need a robust workflow for producing seamless

14:21 Speaker 2: products over fresh and coastal estuaries. To do that, we

14:28 Speaker 2: plan to use machine learning model, and to train our

14:31 Speaker 2: machine learning model, we need a representative data set, so

14:34 Speaker 2: we've assembled a large data set of a variety of

14:38 Speaker 2: different institute products, including chlorophyll, a total suspended sediment, as

14:43 Speaker 2: well as color dissolved organic matter. This data set builds

14:47 Speaker 2: off of Moretz Lehman's Gloria database, which is cited in

14:51 Speaker 2: the bottom right. It has At this point, we have

14:56 Speaker 2: about greater than ten thousand globally distributed in situation measurements.

15:00 Speaker 2: These include those biogeocomical parameters I was talking about before,

15:04 Speaker 2: as well as inherent optical properties from inland and coastal waters.

15:09 Speaker 2: As you can see from the map, about twenty percent

15:12 Speaker 2: about eighty percent of this data set comes from North

15:15 Speaker 2: America as well as Europe, and twenty percent from the

15:18 Speaker 2: rest of the world. Shown here, we have histograms for chlorophyla, PC, TSS,

15:28 Speaker 2: and SDOM from that globally distributed insitue data set, and

15:34 Speaker 2: these histograms show the wide range of different optical conditions

15:38 Speaker 2: we have represented in the data. So if we just

15:41 Speaker 2: look at the chlorophyl a histogram in the top left,

15:44 Speaker 2: we have about six thousand measurements and they span about

15:48 Speaker 2: four orders of magnitude, so it represents anywhere from aligotrophic

15:54 Speaker 2: to utrophic waters quite well. We also have about a

15:58 Speaker 2: thousand measurements of PA see in the top right. PC

16:02 Speaker 2: stands for phychocyanin, which is another pigment, but this pigment

16:05 Speaker 2: specific to cyanobacteria biomass potentially toxic algae. We only have

16:11 Speaker 2: about a thousand measurements of those, which can be a

16:13 Speaker 2: little bit limiting during training. We also have five thousand

16:16 Speaker 2: measurements of toll suspended solids shown on the bottom left,

16:19 Speaker 2: as well as colored dissolved organic matter shown on the

16:22 Speaker 2: bottom right, proxies for nutrient and carbon availability. These again

16:29 Speaker 2: represent a wide range of different values, showing a wide

16:32 Speaker 2: range of different combinations of these different biogeochemical parameters representative

16:36 Speaker 2: in our data set. Okay, so we have this large

16:40 Speaker 2: globally distributed data set. We know we want to use

16:42 Speaker 2: a machine learning model to be able to represent and

16:45 Speaker 2: retrieve estimates in globally distributed inland and coastal waters, so

16:50 Speaker 2: we have to choose a machine learning algorithm. Well, the

16:53 Speaker 2: algorithm that our group chose was termed mixture density networks,

16:57 Speaker 2: and I'll quickly go over that algorithm first. As input,

17:03 Speaker 2: we take the remote sensing reflectance or the ocean color signal,

17:09 Speaker 2: basically just the color of the water at the wavelengths

17:13 Speaker 2: for any specific sensor. Here it's shown for either Hiko

17:16 Speaker 2: or Prisma, so it's hyperspectral. This ocean color signal then

17:21 Speaker 2: goes through band ratio and multi spectral algorithms that typically

17:27 Speaker 2: work well for say, chlorophyl estimation, and these are then

17:30 Speaker 2: input into the standard weights of a neural network after

17:33 Speaker 2: going through a normalization phase where they're scale between negative

17:37 Speaker 2: one and one. What separates mixture density networks from typical

17:41 Speaker 2: machine learning algorithms is the output layer. This output layer

17:45 Speaker 2: is a set of Gaussians which represent the probability distribution

17:50 Speaker 2: function for any given output product, for example, chlorophyl a.

17:59 Speaker 2: These Gaussians have a mean standard deviation and probability associated

18:04 Speaker 2: with them, and when we use a combination function, we

18:07 Speaker 2: can select the most likely value of chlorophyll for the

18:11 Speaker 2: input ocean color signal instead of the average value, which

18:14 Speaker 2: is what typical machine learning algorithms do. We can also

18:19 Speaker 2: simultaneously estimate all of the different products that we're interested in,

18:23 Speaker 2: including chlorophyla if I assign an tss, and seedam, as

18:26 Speaker 2: well as the absorbing IOPs like absorption due to phytoplankton,

18:31 Speaker 2: seed on on and non algo particles. And again this

18:34 Speaker 2: is for those hyper spectral instruments, so product availability depends

18:38 Speaker 2: on the spectral availability. So I'm visually just going to

18:46 Speaker 2: try to show you why mixture density networks are well

18:49 Speaker 2: suited to solve the non unique inverse problem. On the

18:53 Speaker 2: previous page, I showed that we have this sort of

18:55 Speaker 2: typical network structure. The mixture density networks differ because they

19:00 Speaker 2: output a set of gaussians, and these Gaussians can be

19:04 Speaker 2: used to predict let's say the corfy a concentration, when

19:07 Speaker 2: there could be a wide range of different coorfol a

19:09 Speaker 2: concentrations for any input water color spectrum. So let's say

19:14 Speaker 2: we take a water color spectrum like the one shown

19:17 Speaker 2: in this graphic, and we run it through our mixture

19:20 Speaker 2: density network model. What does the output of Gaussians actually

19:23 Speaker 2: look like? So we take this water Coli spectrum in

19:26 Speaker 2: we have let's say a measured in situe corfyle a

19:29 Speaker 2: concentration shown as this black dashed line. We have a

19:33 Speaker 2: prediction from the Ocean color three band or algorithm shown

19:38 Speaker 2: as this dashed blue line. And we have a prediction

19:42 Speaker 2: from a standard machine learning algorithm shown as the dashed

19:46 Speaker 2: red line. Again, these standard machine learning algorithms typically learned

19:50 Speaker 2: to estimate the average of all possible values of coorfal

19:53 Speaker 2: concentration instead of the most likely value. We can then

19:57 Speaker 2: overlay the Gaussians as shown here that our output from

20:03 Speaker 2: the mixture density network. Each of these Gaussians again has

20:07 Speaker 2: a mean, standard deviation and probability associated with it, So

20:12 Speaker 2: we can select the Gaussian with the highest probability, take

20:15 Speaker 2: its median value, and estimate that and for this particular

20:19 Speaker 2: highly eutrophic region, that ends up having a better estimate

20:24 Speaker 2: that matches more closely with the institute chlorophyl a concentration

20:28 Speaker 2: represented by that dash black line than either the standard

20:32 Speaker 2: machine learning algorithm or the ocean color three algorithm. So,

20:37 Speaker 2: in summary, mixture density networks allow us to better solve

20:40 Speaker 2: this non unique inverse problem where there could be many

20:43 Speaker 2: different combinations of chlorophyl a, PC, TSS, and STAM concentrations

20:49 Speaker 2: for the same ocean color signal. Okay, so we have

20:54 Speaker 2: this really large data set and we've picked out a

20:56 Speaker 2: machine learning algorithm that we think is going to work

20:58 Speaker 2: really well for this reason. We can then do a

21:01 Speaker 2: fifty to fifty split from our insitu data, use half

21:05 Speaker 2: of the data for training and half of the data

21:07 Speaker 2: for validation, and then compare the results between our mixture

21:11 Speaker 2: density network and operational algorithms. And that's what's shown here.

21:16 Speaker 2: So on the y axis, we have the estimated chlorophyl

21:18 Speaker 2: a concentration from our algorithm applied to in situ remote

21:23 Speaker 2: sensing respect reflectance data that's been resampled with the sensor

21:27 Speaker 2: spectral response function in this case MSI, and we compare

21:31 Speaker 2: that to estimates from standard operational algorithms like the Gillerson

21:36 Speaker 2: two band algorithm or the blend algorithm. As you can

21:40 Speaker 2: see from this graphic, mixture density networks are able to

21:43 Speaker 2: achieve lower error, lower bias, higher slope, and lower root

21:48 Speaker 2: means square log difference over a wide range of different

21:52 Speaker 2: chlorophyl values. But this is in the idealized scenario where

21:57 Speaker 2: we're estimating based on institute data, where the MP has

22:00 Speaker 2: seen half that data, the actual accuracy when we apply

22:06 Speaker 2: it to satellite instruments is reduced. So shown here we

22:10 Speaker 2: have MSI matchups, so their same day matchups where MSI

22:15 Speaker 2: the sentinel to satellite or the imager on the Sentinel

22:19 Speaker 2: two satellite overpassed while the insitue chlorophyla concentration was being measured,

22:25 Speaker 2: and on the y axis you can see the MDN

22:28 Speaker 2: retrieve chlorpyla from MSI. We have very similar aerror metrics,

22:34 Speaker 2: where MDSA is the error that we saw previously. Previously,

22:37 Speaker 2: it was about twenty eight percent when we were looking

22:39 Speaker 2: on institute data, but once we start looking on satellite

22:42 Speaker 2: retrieve data, when we had to remove that atmospheric signal,

22:47 Speaker 2: the accuracy drops down to about ninety four percent. So

22:50 Speaker 2: that means that the residuals from atmospheric correction substantially reduced

22:53 Speaker 2: the accuracy. However, what we can see from this plot

22:57 Speaker 2: is that the MDN matchups still well represent really wide

23:00 Speaker 2: ynamic range from zero point one to about one hundred

23:03 Speaker 2: milligrams per meter cubed of chlorophyla, So this means it

23:08 Speaker 2: represents anywhere from very illigotrophic waters to very eutrophic waters,

23:13 Speaker 2: so do a good job capturing the full bloom life cycle.

23:17 Speaker 2: Shown on this page, we have MDM derived products for

23:21 Speaker 2: Lake Erie and we can see that they're spatially consistent.

23:26 Speaker 2: So the products you see in the bottom left are

23:28 Speaker 2: actually from OLCIE instead of MSI, so OLGA has more bands.

23:32 Speaker 2: We expected to do a little bit better in terms

23:35 Speaker 2: of product accuracy as well as atmosphere correction for these regions,

23:39 Speaker 2: and we're using it here just to look at the

23:41 Speaker 2: spatial consistency of our mixture density networks when applied to

23:45 Speaker 2: atmospherically corrected data. In the top left you can see

23:50 Speaker 2: the chlorophyl a concentration. Next to it the toll suspended sediments,

23:53 Speaker 2: and the bottom left absorption due to seedam and the

23:56 Speaker 2: bomb right like a cyanin, which again is a proxy

23:59 Speaker 2: for potentially toxic cyano bacteria, specially in the Let's see

24:06 Speaker 2: if we look at the chlorophyl a map in the

24:08 Speaker 2: top left, you can see the Detroit River plum which

24:10 Speaker 2: has really low concentrations of chlorophyl a. In the western

24:15 Speaker 2: section of the image is mau Mi Bay. You can

24:18 Speaker 2: see there's higher concentrations there. In the southernmost section you

24:22 Speaker 2: can see really high concentrations in Sandusky Bay. And specially

24:26 Speaker 2: this lines up with what we've seen historically from this region.

24:30 Speaker 2: We can also look at matchups with in situe measurements

24:36 Speaker 2: that were taken, so I won't cover that here, and

24:38 Speaker 2: in general they line up quite well. Not only can

24:41 Speaker 2: we produce products with our algorithms, we can actually also

24:44 Speaker 2: produce uncertainties. And these uncertainties demonstrate the mall's confidence in

24:48 Speaker 2: its own estimates. So spatially they're very consistent, they don't

24:54 Speaker 2: have dramatic artifacts, and they line up with what we've

24:57 Speaker 2: seen previously. But what happens when we take a temporal

25:01 Speaker 2: time series of these products? So I did that leveraging

25:05 Speaker 2: again ULCI, which has this fine spatial, spectral and temporal

25:08 Speaker 2: resolution combination, and shown here you can see a time

25:13 Speaker 2: series of a harmful elgal bloom which starts in Malmi

25:16 Speaker 2: Bay and interacts with the Detroit River plume. So right

25:19 Speaker 2: now it's march as it turns into April, May and June.

25:23 Speaker 2: This bloom starts to develop here it's interacting with the

25:26 Speaker 2: Detroit River plume, and then we can see it sort

25:29 Speaker 2: of reset again. So temporally we can see strong consistency

25:34 Speaker 2: in our products and actually see some hydrological interactions in

25:39 Speaker 2: these regions. So I was focusing on chlorophyl a, but

25:42 Speaker 2: we can also look at phyic a sign in which

25:44 Speaker 2: again is this marker for potentially toxic cyanobacteria, And you

25:49 Speaker 2: can see really high PC in the Sandusky Bay and

25:53 Speaker 2: Malmy Bay in the months of July and August, suggesting

25:56 Speaker 2: there might be cyanobacteria there. So now I'm going to

26:00 Speaker 2: jump to how stakeholders or users like yourselves can apply

26:06 Speaker 2: our models to preprocessed satellite imagery. The tutorials can be

26:11 Speaker 2: found in the QR code linked below. If you have

26:14 Speaker 2: any issues with it, you can also email me at

26:16 Speaker 2: the blow email. But just to give you a quick

26:19 Speaker 2: overview of what's going to be included in this demonstration, first,

26:24 Speaker 2: we will import some packages as well as functions from

26:29 Speaker 2: those packages, including specific functions from the MDN itself. We

26:33 Speaker 2: will then visualize pre corrected using Acolyte Oulchi imagery and

26:38 Speaker 2: learn how to plot those with some of our own functions.

26:45 Speaker 2: Will then retrieve the RS from this pre processed atmospherically

26:50 Speaker 2: corrected tile. These functions can work with acolyte or LtGen

26:55 Speaker 2: or even polymer. Will then generate predictions from the R,

27:00 Speaker 2: so we'll generate predictions for say tss or chlorophyl a.

27:06 Speaker 2: Then we can also generate uncertainties from these rs and

27:09 Speaker 2: visually look at them and analyze the imagery. And finally

27:12 Speaker 2: we'll do the same thing but for L two gen

27:15 Speaker 2: corrected pace imagery so that we can see some of

27:18 Speaker 2: the more advanced products. All right, now we'll jump into

27:21 Speaker 2: that tutorial. So now we'll present demonstration of our Jupiter

27:32 Speaker 2: Hub notebook called Leveraging Mature Dencity Networks to generate biogeochemical

27:37 Speaker 2: parameter and IOP maps. This is one of three that

27:42 Speaker 2: we have available. The first two cover generating products from

27:46 Speaker 2: institu measurements and generating uncertainties, and they take a little

27:50 Speaker 2: bit of a deeper dive into how these MDMs work.

27:55 Speaker 2: But today I thought I would present on how to

27:57 Speaker 2: generate product and uncertainty maps. This is often what users

28:01 Speaker 2: are most interested in. If you have any questions on these,

28:05 Speaker 2: you can email me a run or a cash directly.

28:10 Speaker 2: So the first thing we're going to do in section

28:13 Speaker 2: one is import the required packages and environment settings so

28:19 Speaker 2: we can run this first block of code. This imports

28:22 Speaker 2: the Python packages needed for the notebook. They're pretty standard

28:25 Speaker 2: packages like numpos, pathlib, zip file, and map plotlib for plotting.

28:32 Speaker 2: And then there's some packages that we've written ourselves as

28:36 Speaker 2: part of the MDM package, which is installed via the

28:40 Speaker 2: instructions at the QR code that you saw during the presentation.

28:44 Speaker 2: I won't go through those instructions here, but if you

28:47 Speaker 2: have any issues again, please email me. These specific MDN

28:51 Speaker 2: functions I'll briefly go through here, and I'll dive into

28:54 Speaker 2: their exact uses a little bit more as we get

28:57 Speaker 2: into the actual code. These functions include get sensor bands,

29:02 Speaker 2: which when you provide it with the sensor name for example,

29:05 Speaker 2: OLGI will get you the required bands for the MDN sensor.

29:09 Speaker 2: Get tile data pulls the bands as well as RRS

29:13 Speaker 2: data from a given tile path so that we can

29:17 Speaker 2: get the rs from a preprocessed net CDF download example

29:23 Speaker 2: imagery we wrote specifically for the Jupiter Hub notebook. It

29:26 Speaker 2: pulls data that we have already atmospherically corrected, so that

29:32 Speaker 2: users can run this notebook without having to do that

29:35 Speaker 2: atmosphere correction themselves. We have fined RGB image which pulls

29:40 Speaker 2: the red, green, and blue bands and outputs that which

29:43 Speaker 2: we can then use as input to the display set RGB,

29:47 Speaker 2: which displays the satellite's RGB image so that we can

29:50 Speaker 2: look at the image itself identify the aquatic regions and

29:54 Speaker 2: cloudy pixels and land pixels. We can also get tiled

29:59 Speaker 2: gear graphic info. This pulls the extent like Latin longitude

30:05 Speaker 2: for the given image. We can then run the RRS

30:11 Speaker 2: data that we pulled using gettile data through map cube

30:14 Speaker 2: MDN full, and then from those predictions that it gives

30:18 Speaker 2: from the MDN, we can overlay the RGB image and

30:22 Speaker 2: the DM products to get final maps for both uncertainties

30:26 Speaker 2: as well as the products. And this will make more

30:28 Speaker 2: sense as I jump into the examples. The next block

30:31 Speaker 2: of code that we run right here just sets the

30:34 Speaker 2: display parameters for map plotlib. I won't go into that

30:37 Speaker 2: in detail. In section two, we just download an ULGA

30:42 Speaker 2: image that again I've already atmospherically corrected. If you were

30:45 Speaker 2: to do this yourself. You'll have to provide your own

30:48 Speaker 2: atmospherically corrected data and just switch out the tile paths below.

30:54 Speaker 2: So here we're setting the sensor to OULCI and the

30:57 Speaker 2: date location just allow us to pull the direct example imagery.

31:02 Speaker 2: I've already downloaded that tile locally, so and it's displayed

31:08 Speaker 2: at this tile path right here. After running that, we

31:14 Speaker 2: can then display the RGB image using displays at RGB,

31:19 Speaker 2: and that image looks like this, which I'll go through

31:22 Speaker 2: in a second. It just takes as input the tile

31:25 Speaker 2: path and sensor. You can change the size of the

31:28 Speaker 2: figure as well as the title which we pull from

31:31 Speaker 2: the sensor and location and date that we specify previously,

31:35 Speaker 2: and it outputs the image as an RGB image which

31:39 Speaker 2: we can use for future algorithms. So this can serve

31:42 Speaker 2: as the base map for future functions that we've written.

31:47 Speaker 2: So if we just like a look at the RGB

31:49 Speaker 2: image itself, you can see that this is an image

31:52 Speaker 2: from OLHI of San Francisco Bay on March sixteenth, twenty nineteen.

31:58 Speaker 2: Let's see, this is the Pacific Ocean. You can see

32:01 Speaker 2: it's pretty blue, meaning it's probably dominated mostly by chlorophyll,

32:05 Speaker 2: probably has low concentrations of kss and sediment. As it

32:10 Speaker 2: moves into the San Francisco Bay, you can see it

32:13 Speaker 2: turns maybe greener, maybe there's higher concentrations of chlorophyll there

32:17 Speaker 2: and higher phytoplankton biomass. As you move up to the

32:21 Speaker 2: northern section of San Francisco Bay, a turns sort of

32:24 Speaker 2: a yellowy color, suggesting maybe there's higher concentrations of suspended

32:28 Speaker 2: sediment there from the Sacramento River in San Juaquan Rivers.

32:34 Speaker 2: And if we look at the southern section of the bay,

32:36 Speaker 2: you can also see these little areas which are either

32:41 Speaker 2: bright green or bright orange or bright yellow. And these

32:44 Speaker 2: are the salt flats which are separated from the bay.

32:48 Speaker 2: So there's a wide range of different optically distinct regions

32:50 Speaker 2: represented here in this image. You can also see some

32:53 Speaker 2: clouds in the bottom left. Okay, so we have the

33:00 Speaker 2: tile data, we have the RGB image. Now we can

33:03 Speaker 2: pull out the remote sensing reflectance data with this block

33:05 Speaker 2: of code using get tile data with the tile path

33:10 Speaker 2: specified and the sensor specified to OLG, it'll pull out

33:13 Speaker 2: the bands and r rs required by our MD prediction function.

33:20 Speaker 2: We can also use get tile geographic info from the

33:23 Speaker 2: tile path to get the lat, long and extent for

33:26 Speaker 2: the image, and then we can plot the RS data

33:30 Speaker 2: for this specific band, just for the first band, just

33:33 Speaker 2: to check that we plot the RS correctly. Indeed, we

33:36 Speaker 2: can see the RS is pulled out for all the

33:38 Speaker 2: aquatic regions and there aren't any RS pixels for land

33:42 Speaker 2: or clouds. Now that we have the RS data, we

33:48 Speaker 2: can move on to generating predictions from the NDN using

33:52 Speaker 2: our MDN cube or map cube MDN full function. So

33:58 Speaker 2: I'll run that while I just go through this funk.

34:00 Speaker 2: You can see it running in the bottom. You can

34:02 Speaker 2: mute this output though here we've included it just so

34:05 Speaker 2: you can see that it is running. So from the

34:09 Speaker 2: MDN package we can pull map cube MDN full as

34:13 Speaker 2: well as get ARGs. These allow us to map the

34:17 Speaker 2: model predictions and uncertainties and to find the arguments that

34:21 Speaker 2: select the correct model. So, for example, in keywords, here

34:24 Speaker 2: we set the sensor to OLCI, the products to chlorophyl TSS,

34:28 Speaker 2: and c doom, and we set the model UID, which

34:32 Speaker 2: is the name of the model weights that we actually use.

34:35 Speaker 2: You don't have to change that for now. If you

34:37 Speaker 2: have questions about setting the model UID for specific sensors.

34:41 Speaker 2: You can again message me after we run get arguments,

34:46 Speaker 2: which pulls the correct arguments for this set of keyword arguments.

34:49 Speaker 2: We can then run map cube MDN full on the

34:52 Speaker 2: arguments and rrs that we've already pulled, as well as

34:56 Speaker 2: the bands. There's a few different that we can set,

35:01 Speaker 2: like land mask, which we set to false year as

35:03 Speaker 2: we just use Acolytes land mask. There's also figure subsample,

35:07 Speaker 2: which allows us to subsample the image so we could

35:12 Speaker 2: reduce the spatial resolution. Basically, this might be useful if

35:16 Speaker 2: you want to plot an MSI image quickly and it

35:18 Speaker 2: has too high spatial resolution and it takes a little

35:21 Speaker 2: too long on your computer to run. We also have

35:24 Speaker 2: scaler mode set to invert. This doesn't need to be changed,

35:28 Speaker 2: but just basically puts the model uncertainty values into the

35:31 Speaker 2: same scale as the model predictions. Block size is set

35:35 Speaker 2: to ten thousand. This is one that you might want

35:37 Speaker 2: to change as this controls the size of the or

35:42 Speaker 2: the number of the rs that are processed at a

35:45 Speaker 2: single time, so reducing it allows you to run this

35:50 Speaker 2: on a personal laptop and increasing it allows you to

35:57 Speaker 2: run things faster if you have more RAM available, and

36:02 Speaker 2: the uncertainty mode is set to composite, so instead of

36:05 Speaker 2: providing uncertainty bounds like low and high uncertainty estimates, it

36:09 Speaker 2: just produces a single model uncertainty. All right, it looks

36:13 Speaker 2: like this ran in the time I took to explain it,

36:16 Speaker 2: and you can see the outputs for each block of

36:18 Speaker 2: ten thousand rs. Again, you can mute this if necessary,

36:23 Speaker 2: but it looks like it finished running. We can then

36:26 Speaker 2: from the model predictions as to plot graphics for chlorophyll

36:31 Speaker 2: as well as for kss and for seed on, and

36:37 Speaker 2: then we can start to visually sort of analyze these products.

36:41 Speaker 2: So these plots are created using overlay RGBMDM products, where

36:45 Speaker 2: we take the RGB image as input the model predictions

36:49 Speaker 2: which were output by the previous algorithm that ran our

36:53 Speaker 2: mature density network model. We pull all the two D

36:57 Speaker 2: information and we can pull this LCE that corresponds to chlorophyl,

37:03 Speaker 2: and how we change it is we just change from

37:05 Speaker 2: chlorophyl to TSS or to seedom depending on the product

37:08 Speaker 2: that we want to plot. For this chlorophyll product, we

37:15 Speaker 2: have this string here that gets set that we can

37:17 Speaker 2: use as the title, and we can put prediction ticks

37:21 Speaker 2: from ten to the zero to ten to the two.

37:24 Speaker 2: So these are in the log scale and we can

37:26 Speaker 2: also set the figure size and that produces these plots

37:30 Speaker 2: with the RGB image background and the MDM products overlaid. Okay,

37:35 Speaker 2: so we saw darker waters in the Pacific and that

37:38 Speaker 2: corresponds to lower concentrations of chlorophyll. That checks out. We

37:42 Speaker 2: saw these really bright salt flats again which have high

37:45 Speaker 2: concentrations of chlorophyl. That also sort of makes sense. And

37:49 Speaker 2: we see a little bit of slightly higher concentrations of

37:52 Speaker 2: chlorophyl in the northern section of San Francisco Bay and

37:56 Speaker 2: the Sacramento and San Jake and rivers generally tracks with

38:01 Speaker 2: hope we saw visually in the RGB imagery. We can

38:05 Speaker 2: also look at the total suspended solids map that we

38:08 Speaker 2: produced from that same image. If we follow this from

38:12 Speaker 2: the Pacific Ocean, you can again see lower concentrations in

38:15 Speaker 2: the Pacific, which makes sense with how we understand the region.

38:19 Speaker 2: Higher concentrations seeming like you know, coming out of the

38:22 Speaker 2: San Francisco Bay, and as you move into the northern

38:25 Speaker 2: section you see really high concentrations which came from these

38:28 Speaker 2: two rivers, which again tracks with our understanding of the region,

38:35 Speaker 2: and we can do the same thing for SEAEDOM, where

38:37 Speaker 2: again you see lower concentrations in the Pacific Ocean and

38:40 Speaker 2: higher concentrations in the northern section of this bay. Not

38:46 Speaker 2: only can we get the products, but we can also

38:49 Speaker 2: visualize uncertainties again using this overlay RGB MDM products function.

38:56 Speaker 2: But now we just also set the image uncertainties using

39:00 Speaker 2: the output model uncertainties for the corresponding slice, for example,

39:04 Speaker 2: chlorophyll in this example, so we can run that for

39:10 Speaker 2: chlorophyl TSS as well as SEDOM, and we can start

39:14 Speaker 2: to see the model's confidence in its own predictions in

39:18 Speaker 2: the same space as the chlorophyl a predictions itself, so

39:22 Speaker 2: you can see the regions where there's higher uncertainties. Often

39:25 Speaker 2: you'll see some of these higher uncertainties in your cloud

39:28 Speaker 2: or maybe some of the atmosphere correction failed and in

39:31 Speaker 2: the really high concentration chlorophyl estimates in these salt ponds,

39:36 Speaker 2: which generally tracks with what we would expect for this region. Finally,

39:43 Speaker 2: we can move on to section six, where we're generating

39:45 Speaker 2: predictions from hyper spectral pace slash oci imagery. We can

39:51 Speaker 2: again just download some example imagery for a day where

39:54 Speaker 2: I've already corrected this imagery from an image of Lake Erie.

39:59 Speaker 2: Then we can display this RGB image from the pace

40:05 Speaker 2: image that we've again already atmospherically corrected. You can see

40:09 Speaker 2: it has darker waters in the eastern section, brighter waters

40:12 Speaker 2: in the western section in Sandusky Bay, which again tracks

40:16 Speaker 2: with our understanding of this region, and again lower brightness

40:20 Speaker 2: from the Detroit River plumb. We can then do all

40:25 Speaker 2: of the same functions or use all the same functions

40:29 Speaker 2: that we used previously, like get tile data, get the

40:32 Speaker 2: tile geographic info to get the r RS and lat

40:35 Speaker 2: long plot the RRS, and then use map cube MDN

40:39 Speaker 2: full on that RRS to get model predictions uncertainty as

40:44 Speaker 2: well as the slices. So I'll run that in the

40:51 Speaker 2: background here and again you can see it running with

40:55 Speaker 2: this process bar that'll run in the backround. But we

41:01 Speaker 2: can already look at some of the visual imagery that

41:03 Speaker 2: I've produced previously. In the top here you can see

41:07 Speaker 2: mgns made for chlorophyl A, PC, TSS and c DOM.

41:14 Speaker 2: So this hyperspectral algorithm allows us to pull out additional

41:20 Speaker 2: products like fecacyanin as well as more advanced products like IOPs,

41:25 Speaker 2: which you can see in the bottom, but spatially, I

41:28 Speaker 2: can just quickly go through what we expected to see,

41:31 Speaker 2: where you can see lower concentrations in the Detroit River

41:34 Speaker 2: plume and the eastern section of Erie, and higher concentrations

41:38 Speaker 2: in Malmi and Sandusky Bays. For PC, this is in May,

41:46 Speaker 2: so you'd expect there to be lower concentrations of pyc

41:50 Speaker 2: asyanin as these harmful albablims if not started yet, and

41:54 Speaker 2: you can again see that in most of this image,

41:56 Speaker 2: though there's a little bit of higher concentrations in Malmi Bay.

42:05 Speaker 2: You can see similar spatial distributions for TSS as well

42:08 Speaker 2: as SEEDM. We can also look at the more advanced

42:13 Speaker 2: products like absorption due to phytoplankton non algo particles as

42:18 Speaker 2: well as seed DOM. Shown here is mdnesimate at aph

42:24 Speaker 2: at four hundred and fifteen animeters four pace. You can

42:29 Speaker 2: also see a different way of length like four ninety

42:32 Speaker 2: five point fifty six twenty This one often corresponds to

42:36 Speaker 2: cyanobacteria and six hundred and seventy and we can also

42:41 Speaker 2: look spectually at absorption due to seedam or we've just

42:44 Speaker 2: plotted fourteen animeters four ninety five, sixty six fifty and

42:49 Speaker 2: six seventy five, and we did the same thing for

42:51 Speaker 2: non algal particles. So you can see that with this

42:54 Speaker 2: hyperspectral data we're able to represent a wide range of

42:57 Speaker 2: additional products. Thanks for listening to my demonstration. I'll be

43:05 Speaker 2: happy to field any questions you might have during the

43:08 Speaker 2: question answer period after a NTA provide some summary information. Thanks.

43:21 Speaker 1: Thank you very much Ryan for your excellent presentation and

43:25 Speaker 1: demonstration of the code, and also for sharing the code

43:28 Speaker 1: with everyone. That brings us to the close off today's session.

43:35 Speaker 1: To summarize, we started with an overview of preprocessing top

43:39 Speaker 1: of atmosphere radiances from satellite images. We would start with

43:45 Speaker 1: level one satellite data. We saw that there are several

43:49 Speaker 1: data portals where different satellite and sensor level one data

43:53 Speaker 1: are available. Then get atmospherically corrected level two remote sensing

43:58 Speaker 1: deflectancies from the data, and these remote sensing reflectancies can

44:04 Speaker 1: be used as inputs to the MDN models. Then OSHA

44:10 Speaker 1: provided an overview and demonstration of using md AND models

44:14 Speaker 1: to derive water quality parameters. Specifically, the models are trained

44:19 Speaker 1: using a large globally distributed data set of INC two

44:23 Speaker 1: water quality measurements and colocated hyperspectral measurements from a wide

44:28 Speaker 1: range of optically complex regions and sharing many of these

44:31 Speaker 1: samples with Gloria data. The MDN approach is well suited

44:37 Speaker 1: for addressing non unique inverse problems like ocean color remote

44:41 Speaker 1: sensing due to its output layer. The models retrieve key

44:46 Speaker 1: water quality parameters depending on spectral availability, including chlorophyl a,

44:52 Speaker 1: total suspended solids, colored dissolved organic matter, and phycosyanin, along

44:58 Speaker 1: with their associated uncertain Similar models were developed to support

45:04 Speaker 1: multiple multi and hyperspectral satellite images including MOTIS Weirs, Lancet

45:10 Speaker 1: eight and nine, ONLY, Sentinel two, MSI, Sentinel three, OLG,

45:16 Speaker 1: and base OCI. This brings us to the conclusion of

45:20 Speaker 1: this webinar series. To summarize, we started with an overview

45:25 Speaker 1: and demonstration of the satellite based Tool for Rapid Evaluation

45:30 Speaker 1: of Aquatic Environments or stream Webtool and API. In part one,

45:35 Speaker 1: we saw how to map water quality parameters including chlorophylate concentration,

45:40 Speaker 1: total suspended solids, and siky disk depth. The data are

45:44 Speaker 1: available from twenty twenty four to near real time in

45:48 Speaker 1: coastal estuaries and inland lakes across the US at twenty

45:52 Speaker 1: to thirty meters special resolution and these data are derived

45:56 Speaker 1: from Lancet eight and nine and Sentinel two A, two

45:59 Speaker 1: B and to C imagery. We saw how to search

46:04 Speaker 1: and download water quality products using the stream webtool and API,

46:08 Speaker 1: and then we examine time series of water quality parameters

46:12 Speaker 1: in the region of interest using QGIS. Today in part two,

46:16 Speaker 1: we had an overview and demonstration of the aquower's algorithm.

46:21 Speaker 1: So introduction to the md and models to derive water

46:24 Speaker 1: quality parameters and access an application of md AND based

46:29 Speaker 1: model code to retree water quality parameters. Finally, stream webtool

46:36 Speaker 1: and API and the MDEN code are all open source

46:40 Speaker 1: and you can download and use the code as we

46:43 Speaker 1: saw today from doctor Roche. There is one homework assignment

46:49 Speaker 1: and it's posted on the training webpage today. Answers must

46:54 Speaker 1: be submitted by Google Forms and the homework is due

46:57 Speaker 1: on tenth March. Certificate of completion will be awarded to

47:03 Speaker 1: those who attended all live webinar sessions. Would complete the

47:07 Speaker 1: homework assignment by the deadline and then you will receive

47:11 Speaker 1: a certificate via email approximately two months after completion of

47:15 Speaker 1: the course. Once again, we thank our guest speakers, William

47:21 Speaker 1: Wainwright and Ryan O'Shea for their presentations and demonstrations describing

47:28 Speaker 1: how what quality parameters are derived from remote sensing and

47:33 Speaker 1: then making these data available through stream website. Trainer's contact

47:40 Speaker 1: information is provided here and you can always contact our set.

47:46 Speaker 1: Our set website and our set YouTube link are given here.

47:50 Speaker 1: All the training materials this training and previous trainings material

47:54 Speaker 1: can be found here. For questions, comments, or to share

47:58 Speaker 1: how you have applied our trainings to your work or studies,

48:02 Speaker 1: email NASA dot r set at gmail dot com. Join

48:06 Speaker 1: our quarterly newsletter to stay up to date on our

48:09 Speaker 1: latest trainings and the steps are given here. Just send

48:12 Speaker 1: an email with no subject line to this email address

48:16 Speaker 1: and follow the instructions sent in response. Here are some

48:21 Speaker 1: useful resources. And with that we thank you all for

48:25 Speaker 1: attending today's session and we will go to the question

48:28 Speaker 1: and answer session. Now we'll start with the question and

48:39 Speaker 1: answer session, and once again we thank our guest speakers,

48:44 Speaker 1: William Wainwright and doctor Ryan O'she. We'll start with question one,

48:52 Speaker 1: how much does special resolution affect the processing time of

48:56 Speaker 1: the MTN? And you can mute and answer the question, all.

49:03 Speaker 2: Right, thanks. I'd say most of the variability comes from

49:08 Speaker 2: the number of aquatic pixels in the given image. So

49:12 Speaker 2: for the same area, MSI, if you're using it at

49:15 Speaker 2: the ten meter spatial resolution, would have almost you know,

49:19 Speaker 2: nine hundred times more pixels than say, OCHI at three

49:23 Speaker 2: hundred meter resolution and therefore take that much longer to process.

49:28 Speaker 2: But really it's a function of the number of available

49:30 Speaker 2: aquatic pixels.

49:34 Speaker 1: Great, thank you. Next question is what is the difference

49:37 Speaker 1: between PC and TSS. Yeah.

49:41 Speaker 2: Sure, PC is defined on slide twenty in the presentation.

49:46 Speaker 2: It stands for phytocyanin fyight. A sigin is a pigment

49:50 Speaker 2: that is specific to cyanobacterial biomass, which is a potentially

49:55 Speaker 2: toxic algae, and TSS stands for total suspended solids. These

50:01 Speaker 2: can be things like sand which has been suspended in

50:03 Speaker 2: the water column.

50:08 Speaker 1: Next question, is there any reason to select fifty to

50:11 Speaker 1: fifty training and validation data set? We usually considered eighty

50:15 Speaker 1: twenty or seventy thirty.

50:19 Speaker 2: Sure. We have a long response to this question in

50:23 Speaker 2: the Q and A document, but the short answers to

50:27 Speaker 2: ensure statistically robust and representative validation across the range of

50:31 Speaker 2: observed conditions, so we adopted this fifty to fifty split.

50:35 Speaker 2: We have a pretty large data set, so we're able

50:38 Speaker 2: to split it and get you know, robust validation metrics. Additionally,

50:44 Speaker 2: this is really are idealized aerometric I guess that we're

50:51 Speaker 2: reporting here. We have multiple other ways of testing these algorithms.

50:56 Speaker 2: One of them is a regional split where we'll split

50:58 Speaker 2: the data regionally, so we'll train on everything except for

51:02 Speaker 2: one region and test it in that region and then

51:04 Speaker 2: report the aerometrics there to get an out of training

51:07 Speaker 2: data set estimate that usually has like fifty uncertainty or residuals.

51:13 Speaker 2: And then the second way, which I showed here today

51:16 Speaker 2: was we'll also try to get matchups with the satellite

51:19 Speaker 2: sensor for the probably most realistic error metric of what

51:23 Speaker 2: you'll actually see from satellite imagery.

51:28 Speaker 1: Thank you, Rian. Question for is just to be sure

51:31 Speaker 1: stream uses day radiances for satellites as input to the

51:35 Speaker 1: md and model to derive water quality parameters.

51:42 Speaker 2: Sure, so stream actually has two different mDNS. One MDN

51:48 Speaker 2: retrieves the remote sensing reflectance or ocean color from the

51:52 Speaker 2: tay radiance, and then we actually have a second MDM

51:55 Speaker 2: that retrieves what our quality parameters from the remote sensor

51:58 Speaker 2: reflectants or ocean colors. So most of what we've published

52:02 Speaker 2: so far has been on that second MBM where we

52:05 Speaker 2: try to retrieve these water quality parameters, you know, things

52:08 Speaker 2: like corphyl ATSs SECU to step from the RS. But

52:12 Speaker 2: we are rapidly trying to get out a publication to

52:16 Speaker 2: retrieve the RS from the ta radiances as well, and

52:21 Speaker 2: we expect in the next I guess six months to

52:23 Speaker 2: have that MDN available online. But we'll keep you posting.

52:30 Speaker 1: Great thank you for that reply, and if you go

52:33 Speaker 1: back and see some trainings that our set did. In

52:38 Speaker 1: those trainings see dos orcss W have been used. That's

52:42 Speaker 1: also NASA software and soon a COOWERS will be available

52:47 Speaker 1: as well. Question five. The ms I scatter plots show

52:52 Speaker 1: a classic pteroctcit pattern of residual Sorry, there are stepatistical

53:00 Speaker 1: corrections that can account for this and provide better predictions.

53:05 Speaker 1: Has anyone tried that residuals are clearly a function of

53:10 Speaker 1: the X variable.

53:13 Speaker 2: So this isn't exactly my area of expertise, but somebody

53:16 Speaker 2: else in my group wrote a response which I will

53:19 Speaker 2: read here. The MSI scatterplot should not be interpreted in

53:23 Speaker 2: the classical regression sense. In this case, the x axis

53:27 Speaker 2: does not represent an independent predictor variable upon which the

53:31 Speaker 2: y axis depends. Mal predictions are generated from the input rs,

53:36 Speaker 2: whereas the x axis represents independent in situe measurements solely

53:40 Speaker 2: for validation, we acknowledge the presence of systematic bias and

53:45 Speaker 2: range dependent error behavior in certain portions of the dynamic

53:48 Speaker 2: range or the water quality variables. These trends are consistent

53:52 Speaker 2: with the headerosydastic characteristics of the data set and are

53:58 Speaker 2: an active area of ongoing mind refinement.

54:05 Speaker 1: Thank you. My question six is as what is temporal

54:09 Speaker 1: resolution is higher than lensor sentinel. Is there a way

54:13 Speaker 1: to use it as a support tool that works whenever

54:17 Speaker 1: there are clouds or glint effects and getting a higher

54:20 Speaker 1: amount of images.

54:23 Speaker 2: Yeah, So this is again another like active area of research.

54:27 Speaker 2: How do we harmonize between these different satellite sensors and

54:31 Speaker 2: combined together their outputs. So one thing we tried to

54:36 Speaker 2: do with these mixture density networks was apply them to

54:39 Speaker 2: the available satellites that we had access to. And I

54:42 Speaker 2: showed in one of the earlier slides all of the

54:45 Speaker 2: different satellites that we do have MDMs that we that

54:49 Speaker 2: support these different satellite sensors. These MDMs are trained on

54:53 Speaker 2: similar data sets, but you know, each of these sensors

54:56 Speaker 2: does have different spectralaracteristics. So what we end up seeing

55:03 Speaker 2: is that these mDNS generally produce pretty harmonized retrievals for

55:08 Speaker 2: the different products, including chlorophyll, between the different sensors. However,

55:15 Speaker 2: long term, we're trying to attempt to further improve harmonization

55:19 Speaker 2: between these sensors, specifically Landsat, Sentinel two and Sentinel three

55:23 Speaker 2: by leveraging even more advanced machine learning models. So I

55:26 Speaker 2: guess the short answer to that question is yes, you

55:29 Speaker 2: certainly can use our models to fill in and build

55:32 Speaker 2: time series, and we're trying to improve that in the future.

55:39 Speaker 2: Thank you.

55:40 Speaker 1: Question seven, Can planet data be used for MDN or

55:44 Speaker 1: is among the coming soon?

55:47 Speaker 2: Yeah? Actually, we do have an MDN available for chlorophyla

55:51 Speaker 2: and psyche de step from Planets super Dove. I provide

55:54 Speaker 2: a link in the Q and a text. However, it's

55:58 Speaker 2: not currently supported as part of the tutorials are GitHub.

56:01 Speaker 2: We do plan to addit soon, probably in the next

56:04 Speaker 2: month or two. As we are rapidly adding in these

56:08 Speaker 2: additional sensors that we have models for but aren't quite

56:12 Speaker 2: supported by the tutorials just yet. So feel free to

56:15 Speaker 2: email me and when we push the updates to our

56:17 Speaker 2: GitHub I can, I can send an email.

56:19 Speaker 1: Back wonderful, Thank you Rian. Our next question is have

56:24 Speaker 1: the md and results been compared with the detard products

56:27 Speaker 1: from Echolite.

56:30 Speaker 2: I'm not sure exactly what algorithm Accolte uses, but I

56:34 Speaker 2: guess the short answer to this question is no, we

56:36 Speaker 2: haven't done that direct comparison. We have done comparisons to

56:39 Speaker 2: other operationally available algorithms where we download the products from websites,

56:46 Speaker 2: like the Ocean Color V four algorithm, as well as

56:48 Speaker 2: cyan four regions where we have a lot of insitue

56:51 Speaker 2: data like Lake Erie and the Chesapeake Bay. And in

56:55 Speaker 2: general we've seen the mdn's outperform these algorithms across the

57:00 Speaker 2: wide dynamic range that's represented, particularly in Lake Erie. So

57:04 Speaker 2: for bluem formation in particular, they're they're quite robust.

57:11 Speaker 1: Next question is can this code be used anywhere? Do

57:15 Speaker 1: we have to change the input which depends on the

57:18 Speaker 1: zone that we study.

57:21 Speaker 2: Short answers, Yes, it can be used anywhere. The model

57:26 Speaker 2: only relies on r rests or the remote sensing reflectants

57:30 Speaker 2: the ocean color signal as input, so it can be

57:33 Speaker 2: used in any region. There aren't any other inputs.

57:40 Speaker 1: Question ten. Currently, I'm working on estimating TSS using surface

57:45 Speaker 1: reflectance data since high TSS in institute data are not

57:50 Speaker 1: available for training. My model predicts low TSS well used

57:54 Speaker 1: during extreme events, even though I know the actual TSS

57:58 Speaker 1: consultations should be much higher than the estimates. Could you

58:01 Speaker 1: please suggest how to address this issue.

58:08 Speaker 2: Let me think about that one for a second. In

58:12 Speaker 2: this case, or the TSS being estimated from the MDN

58:15 Speaker 2: and I guess how high is high for the TSS values.

58:20 Speaker 2: The short answer I guess is that if you did

58:24 Speaker 2: have high TSS values with RS available, you could potentially

58:31 Speaker 2: either retrain the MDN or apply a transfer learning approach

58:36 Speaker 2: where you take the MDN that we have in the

58:37 Speaker 2: industry retrain some of the later layers using this institute data. Yeah,

58:46 Speaker 2: I think that would probably be the best approach.

58:51 Speaker 1: So Ran, would you say that because this was trained

58:55 Speaker 1: MDN was trained with like a large range of values,

59:01 Speaker 1: we may perform better than other ritual schemes for DSS.

59:07 Speaker 2: I would expect to perform pretty well for the region

59:13 Speaker 2: or for the concentration ranges that we have available, but

59:18 Speaker 2: not necessarily outside of that. We do have limited data

59:22 Speaker 2: at the higher ends, so that might be a limitation

59:28 Speaker 2: for accuracy.

59:30 Speaker 1: Thank you. Question eleven, is it possible to find tune

59:34 Speaker 1: MD and using incito data?

59:36 Speaker 2: Yeah, that's what it builds off that last question. The

59:40 Speaker 2: short answer is yes, it is possible, particularly using transfer learning.

59:46 Speaker 2: I think we've tried that in our lab before. I

59:49 Speaker 2: do think we have some code for it. We don't

59:52 Speaker 2: currently support that via tutorials, but if you send me

59:55 Speaker 2: an email, we can set up a video conference or

59:59 Speaker 2: something and we can discuss how to do this yourself.

1:00:02 Speaker 1: Great. Thank you Ryan. Question twelve, the uncertainty map seem

1:00:07 Speaker 1: to look very similar to the water quality maps. High

1:00:11 Speaker 1: not wqu estimates equal high uncertainty. Does that mean that

1:00:15 Speaker 1: the model can be less reliable at higher values?

1:00:20 Speaker 2: I guess I'd say that the bulk value of uncertainty

1:00:23 Speaker 2: is higher for higher values, and we can use relative

1:00:26 Speaker 2: uncertainty or percentages to identify outliers in terms of prediction performance.

1:00:36 Speaker 1: Next question is can uranium contamination in surface water be estimated.

1:00:43 Speaker 2: I don't have any experience with mapping uranium contamination, and yeah,

1:00:49 Speaker 2: I don't think we've used that previously or our seen

1:00:52 Speaker 2: data from that.

1:00:55 Speaker 1: Question four is do you have any guidance or recommended

1:00:59 Speaker 1: sources resources for reprocessing the RS inputs? Specifically, I'm interested

1:01:05 Speaker 1: in MSI and only.

1:01:08 Speaker 2: Yes, so the I guess atmosphere corrections that from our

1:01:13 Speaker 2: experience we've seen work well in neutrophic waters is ACCOLAI

1:01:17 Speaker 2: and more oligotrophic waters is ltgener CDs. Also coming soon,

1:01:22 Speaker 2: we are producing our own in house aquaverse atmosphereic correction method.

1:01:29 Speaker 2: We are working on paper for that and should have

1:01:31 Speaker 2: that out in the next sort of six months, so

1:01:34 Speaker 2: stay turned.

1:01:39 Speaker 1: Next question is the simple dataset histogram shown earlier. What

1:01:45 Speaker 1: is the temporal range and the special extent of the data?

1:01:52 Speaker 2: Sure, it's similar to that of the Gloria database. We've

1:01:56 Speaker 2: taken Gloria database and sort of extended it with our

1:01:59 Speaker 2: own situ data as well. I'm not sure exactly on

1:02:03 Speaker 2: the temporal range, but much of the data is from

1:02:07 Speaker 2: the last twenty years and spacially most of the data

1:02:11 Speaker 2: from that map I showed. I don't know the slide number,

1:02:15 Speaker 2: but I do have that graphic where it shows about

1:02:17 Speaker 2: eighty percent of the data is from North America and

1:02:19 Speaker 2: Europe and maybe twenty percent is from the rest of

1:02:22 Speaker 2: the globe. We are trying to build that out. If

1:02:25 Speaker 2: if you have data set from underrepresented regions, we'd be

1:02:29 Speaker 2: happy to incorporate that into future versions of the MDN model,

1:02:34 Speaker 2: and theoretically that would improve the mdn's retrievals in those regions.

1:02:38 Speaker 2: So again, yeah, feel free to send me an email

1:02:40 Speaker 2: message if you're interested in that.

1:02:44 Speaker 1: Thank you. Next question, is there a new atmospheric correction

1:02:48 Speaker 1: embedded in the MDN. Do you suggest just starting with

1:02:51 Speaker 1: L two data.

1:02:54 Speaker 2: Yeah, that's correct, So the mdians that are available on

1:02:57 Speaker 2: the GitHub tutorial right now just start from the level

1:03:01 Speaker 2: to data, so i'd recommend something like L two gen

1:03:04 Speaker 2: or acolyte. Again, we are working on developing in house

1:03:07 Speaker 2: atmospheric correction models. We have some that work really well,

1:03:11 Speaker 2: but we have to get those publications out and then

1:03:13 Speaker 2: we can add those to the GitHub tutorials. We're planning

1:03:18 Speaker 2: to do this in advance of Ocean Optics, which is

1:03:22 Speaker 2: in I believe September, where we'll host a tutorial there.

1:03:28 Speaker 1: Next question, is it possible to apply these models to

1:03:32 Speaker 1: earlier versions of landsets, such as landset five or seven.

1:03:36 Speaker 1: What would be the limitations?

1:03:45 Speaker 2: Let me think about that for a second. I think

1:03:47 Speaker 2: we and somebody from my group can correct me if

1:03:51 Speaker 2: I'm wrong. I believe we have applied some of these

1:03:54 Speaker 2: models for SECI distep retrieval from landset five and them,

1:04:01 Speaker 2: but I don't think we've done that for corpyl A

1:04:04 Speaker 2: or TSS, so I'm not sure if we've explored that fully.

1:04:12 Speaker 2: I'm not sure how the SNR for these sensors would

1:04:15 Speaker 2: support retrieval of corpolate products, but that's definitely something we'd

1:04:20 Speaker 2: be interested in exploring.

1:04:23 Speaker 1: Thank you. Next question is are there any matchup studies

1:04:27 Speaker 1: for MD and base predicted, TSS and HL.

1:04:34 Speaker 2: Not published? However, we are doing that in house. We

1:04:37 Speaker 2: are funded as part of the PACE Science and Applications team.

1:04:43 Speaker 2: We are working on one such publication where we have

1:04:46 Speaker 2: matchups for corfy A and fykasionin across oligotrophic waters in

1:04:54 Speaker 2: the US that I don't know, hopefully will be out

1:04:57 Speaker 2: in the next year. And then also we have an

1:05:01 Speaker 2: in house validation data set from the PACE validation team,

1:05:09 Speaker 2: and those seem to do really well when compared to

1:05:11 Speaker 2: OCV four in more again utrophic inlandic coastal waters. So

1:05:21 Speaker 2: short answer is, yes, we've done these studies, but they're

1:05:24 Speaker 2: not published yet.

1:05:28 Speaker 1: Thank you. Next question, what is the accuracy of the

1:05:31 Speaker 1: md AND network?

1:05:36 Speaker 2: You know, that's a difficult question to answer because there's

1:05:41 Speaker 2: a few different accuracies that you can report, and for

1:05:44 Speaker 2: multiple different satellite sensors, so it really depends on the

1:05:48 Speaker 2: satellite sensors spectral capabilities as well as it's SNR the

1:05:53 Speaker 2: accuracy of the atmospheric correction, and you know, if you're

1:05:59 Speaker 2: applying with this algorithm to institute rs without any atmospheric correction,

1:06:04 Speaker 2: or if you're applying it to satellite imagery, if you

1:06:09 Speaker 2: are interested in the accuracy for specific sensors, we have

1:06:12 Speaker 2: a wide range of publications that are cited in the

1:06:16 Speaker 2: tutorials at the bottom. You can look through those, and

1:06:21 Speaker 2: in general we try to publish the accuracies in each

1:06:24 Speaker 2: of those different scenarios for the different sensors that we support.

1:06:30 Speaker 1: Thank you. Next question, is there a reason why you

1:06:33 Speaker 1: did not compare with other available machine learning algorithms?

1:06:40 Speaker 2: Yeah, I guess typically we try to compare with the

1:06:45 Speaker 2: sort of more operationally available algorithms, and we've generally focused

1:06:54 Speaker 2: on those. We'd be happy to compare against the other

1:06:59 Speaker 2: machine learning out agorithms if there's any specific one that

1:07:02 Speaker 2: you're interested in with any data so we might have.

1:07:08 Speaker 1: Thank you, Brian. Question twenty one is the same network

1:07:12 Speaker 1: used for all sensors.

1:07:17 Speaker 2: So for the multi spectral instruments, we typically try to

1:07:22 Speaker 2: use the same MDN network architecture with the same number

1:07:28 Speaker 2: of layers and neurons. However, when we move to hyper

1:07:33 Speaker 2: spectral sensors like EMIT, Prisma, and pace, we had to

1:07:38 Speaker 2: increase the number of neurons for the larger amount of

1:07:42 Speaker 2: inputs as well as outputs because now we were inputting

1:07:46 Speaker 2: hyper spectral data and we're outputting more complex products like

1:07:51 Speaker 2: absorption due to phytoplankton, which is spectually variable. So we

1:07:54 Speaker 2: have something like fifty inputs and seventy outputs. So the

1:07:59 Speaker 2: short answer is for the multi special instruments, yes, it's

1:08:02 Speaker 2: very similar. For hyper spectral we had to increase the

1:08:05 Speaker 2: number of neurons in the network. We use MDS for both. Sorry, yeah, yeah.

1:08:14 Speaker 1: So next question is I didn't understand that one MD

1:08:18 Speaker 1: in redevese rrs for iop ocean color properties, and the

1:08:22 Speaker 1: second MD and retrieves water quality parameters. Isn't it support

1:08:26 Speaker 1: supposed to be the same.

1:08:28 Speaker 2: Yeah, So for each sensor, there's they have different spectra

1:08:36 Speaker 2: that are supported. So for the multi spectral instruments, typically

1:08:41 Speaker 2: what we retrieve ischlorophy a, TSS, and c DOM, also

1:08:46 Speaker 2: pechey dis deep sometimes for the models that we support.

1:08:50 Speaker 2: But for the hyper spectral instruments, we started adding in

1:08:55 Speaker 2: the IOPs as well. So not only do we retrieve

1:08:58 Speaker 2: chlorophy a, TSS, and on, but we also simultaneously retrieve

1:09:03 Speaker 2: of absorption due to phytoplankton, color, dissolved organic matter, and

1:09:08 Speaker 2: non alcohol particles. Yeah. So it depends on the sensor

1:09:14 Speaker 2: which products we currently support if they do retree of

1:09:18 Speaker 2: them simultaneously.

1:09:22 Speaker 1: Next question is which bands are the network using for

1:09:26 Speaker 1: chlorophylla and the other parameters.

1:09:30 Speaker 2: Yeah, again, it depends very much on the sensor that

1:09:32 Speaker 2: we're talking about. Typically we try to keep the bands

1:09:36 Speaker 2: in the I believe it's four hundred and fifteen nanimeter

1:09:41 Speaker 2: to about seven hundred and fifteen animeter range that varies

1:09:45 Speaker 2: a little bit. Sensor to sensor. The exact way of

1:09:50 Speaker 2: lengths can be found in the tutorials using the get

1:09:56 Speaker 2: sensored bands function for the specific sensor, but we try

1:10:00 Speaker 2: to use all the bands in that range typically.

1:10:06 Speaker 1: Next question is is there a document that lists the

1:10:09 Speaker 1: water bodies where the India is trained and validated.

1:10:16 Speaker 2: There's an individual document. However, we do have again a

1:10:22 Speaker 2: wide set of publications available. These are cited at the

1:10:27 Speaker 2: bottom of our tutorials and in each one of those

1:10:34 Speaker 2: publications we go through very similar training where we do

1:10:37 Speaker 2: the fifty to fifty split. Sometimes we do a regional

1:10:40 Speaker 2: split as well, and then finally we try to do

1:10:44 Speaker 2: a matchup analysis for each one of the sensors, and

1:10:49 Speaker 2: in that we shall also talk about we do also

1:10:51 Speaker 2: talk about the sort of spatial distribution of that data.

1:10:56 Speaker 1: The next question is is there can it distinguish between

1:11:00 Speaker 1: oil spills and habs or oceans.

1:11:04 Speaker 2: We don't have experience using these MDMs for distinguishing between

1:11:08 Speaker 2: oil spills and habs. Okay.

1:11:14 Speaker 1: The next question is how to NASA Ocean Color and

1:11:18 Speaker 1: inland water Quality algorithms manage seasonal spectral shifts caused by

1:11:23 Speaker 1: monsoon driven trabid t seedom variations and changing solar geometry

1:11:28 Speaker 1: without introducing systematic bias in long term time series.

1:11:34 Speaker 2: Let me think about that one for a second. Yeah,

1:11:45 Speaker 2: I think this is a really good question. I think

1:11:47 Speaker 2: I'd have to think on it a little bit before

1:11:50 Speaker 2: providing an answer. I'll try to do that over the

1:11:52 Speaker 2: next week and get back to it.

1:11:56 Speaker 1: Great, Thank you so much. The next question is why

1:12:01 Speaker 1: was the NC two chlorophyl A data from USGS, which

1:12:05 Speaker 1: covers the United States not used.

1:12:10 Speaker 2: Let's see does that data set include r rs as

1:12:15 Speaker 2: well as colorful A. If so, we'd be happy to

1:12:19 Speaker 2: add it in. I'm not entirely sure if yeah, If

1:12:23 Speaker 2: it does include rs, we'd be happy to add it

1:12:25 Speaker 2: in when we retrain these models. If it is just

1:12:29 Speaker 2: coreful A, probably have pulled some of that data for validation,

1:12:36 Speaker 2: but we can't use it for training with the MDM

1:12:38 Speaker 2: since they require the RS as input.

1:12:44 Speaker 1: Question eight, does stream recommend regional recalibration of retrieval algorithms

1:12:51 Speaker 1: or does it rely on globally parameterized models with deviation

1:12:56 Speaker 1: correction layers.

1:12:59 Speaker 2: We do not currently recommend regional recalibration of these algorithms. Usually,

1:13:06 Speaker 2: what we would like to do is just keep adding

1:13:09 Speaker 2: in more and more data globally. So again, if you

1:13:13 Speaker 2: have in siture data with RS and the parameter of

1:13:16 Speaker 2: interest and you're willing to share it, please send me

1:13:18 Speaker 2: an email and we'd be happy on future model updates

1:13:22 Speaker 2: to include that data, which theoretically should increase the accuracy

1:13:27 Speaker 2: of the model for that given region. Yeah, the model

1:13:32 Speaker 2: is globally trained. We don't deviate at all based on

1:13:38 Speaker 2: the region, so the same model gets applied just to

1:13:40 Speaker 2: the RS no matter what region we're in.

1:13:46 Speaker 1: Thank you. Next question, Can the NDN models be used

1:13:50 Speaker 1: for volatile and non volatile compounds as part of water

1:13:54 Speaker 1: quality predictions in the case of nitrates, phosphorates, am monco,

1:14:01 Speaker 1: nitrates and other ions, could MDN be used if yes?

1:14:06 Speaker 1: In general in general practice, how many incituo observations are

1:14:11 Speaker 1: we looking here to train md IN models?

1:14:17 Speaker 2: Oh? I see the MDN can only be used for well,

1:14:24 Speaker 2: I guess the md and from the satellites with multi

1:14:27 Speaker 2: spectral or hyperspectual data and visible and urinef red range

1:14:31 Speaker 2: can only be used for products that change the ocean

1:14:35 Speaker 2: color signal if you have RS with different concentrating. I

1:14:42 Speaker 2: don't know specifically how these different compounds change the ocean

1:14:46 Speaker 2: color signal or if they do interact. If you do

1:14:49 Speaker 2: have a data set of RS with different concentrations of

1:14:52 Speaker 2: these different products and they do change it, then we

1:14:55 Speaker 2: could try to train an MDM to do that for

1:15:00 Speaker 2: how many institue observations? Again, that sort of depends, but

1:15:05 Speaker 2: I guess we've used in practice as little as a

1:15:10 Speaker 2: thousand samples to train our PC models, so that represented

1:15:14 Speaker 2: a wide range of different flectasign and concentrations pretty well,

1:15:19 Speaker 2: but I don't know if we've looked at an exact minimum.

1:15:25 Speaker 1: Next question, can mbian model run on wetland water quality monitoring?

1:15:34 Speaker 2: Let me think about that for one second. If the

1:15:38 Speaker 2: wetlands are optically deep, meaning that we can't see the

1:15:42 Speaker 2: bottom reflectance, then theoretically I believe they should work just

1:15:48 Speaker 2: as well for retrieving CHLOROPHYLA or TSS. However, if you

1:15:52 Speaker 2: can see the bottom reflectance, that is going to change

1:15:56 Speaker 2: the ocean color signal or the signal by these atmosphere

1:16:01 Speaker 2: correction algorithms, And we really haven't tested the model in

1:16:05 Speaker 2: these regions, thank you.

1:16:10 Speaker 1: Our next question is should input LTO data be gritted

1:16:13 Speaker 1: before using the MDN or does it not matter.

1:16:21 Speaker 2: One second? Now I think the answer is no, it

1:16:25 Speaker 2: doesn't matter. It doesn't have to be gritted. Yeah, the

1:16:33 Speaker 2: MDN just takes as input the remote sense and reflectance

1:16:37 Speaker 2: in basically an array. So as long as you can

1:16:39 Speaker 2: get it into the format similar to that that's applied

1:16:44 Speaker 2: in the tutorials, you should be good.

1:16:50 Speaker 1: Thank you so much. It looks like there's a lot

1:16:52 Speaker 1: of interest and we had a lot of great questions.

1:16:56 Speaker 1: We will be posting our Q and A documents on

1:17:00 Speaker 1: the training website in a week or so for both

1:17:04 Speaker 1: the sessions. Okay, I think there's one more question. How

1:17:09 Speaker 1: was the RS obtained for gloria and other data sets

1:17:13 Speaker 1: used where they downscale from any satellite data to the

1:17:16 Speaker 1: location of the sample site? If so, can I do

1:17:19 Speaker 1: similarly for my institute data sets, which only have glorophyll a.

1:17:26 Speaker 2: Sure, so the Gloria data set and we actually use it,

1:17:29 Speaker 2: you know, an augmented version of that, but it has

1:17:32 Speaker 2: very similar properties, has hyperspectral RRS and a few different

1:17:40 Speaker 2: in situ products, including chlorophyla. We take that hyperspectral RS

1:17:45 Speaker 2: and resample it with the sensors relative spectral response function

1:17:51 Speaker 2: and then we train the models using that resampled RS.

1:17:58 Speaker 2: You could do that with your own insert your data

1:18:00 Speaker 2: set if you have chlorophyla with RRS. If you only

1:18:05 Speaker 2: have chlorophyll a, you wouldn't be able to reach OT.

1:18:11 Speaker 1: I'm currently working on a coastal water quality study area.

1:18:17 Speaker 1: This area covers a small region and I want to

1:18:19 Speaker 1: use Stream, but Stream does not support this area with data.

1:18:23 Speaker 1: What can I do?

1:18:26 Speaker 2: You can email us. We are trying to support users

1:18:31 Speaker 2: that are interested in using our products. We do have

1:18:35 Speaker 2: some processing limitations that we're trying to work through, but

1:18:38 Speaker 2: if you email us, we can try to add a

1:18:42 Speaker 2: specific tile in that you might be interested in.

1:18:47 Speaker 1: That's fantastic, Thank you so much. Are these data exsible

1:18:54 Speaker 1: and Google attentionIn?

1:18:58 Speaker 2: No, we have not included these data and Google or

1:19:01 Speaker 2: if engineer. We do not currently have plans to.

1:19:08 Speaker 1: Hello, I'm from Mexico and I would like to use

1:19:11 Speaker 1: this tool in my country, but I saw it's only

1:19:13 Speaker 1: available in the US. Can be the possibility that we

1:19:18 Speaker 1: can use in Latin America? Sure?

1:19:22 Speaker 2: Again? Yeah, I feel free to email me if there's

1:19:24 Speaker 2: a specific site that you're interested in, and we can

1:19:27 Speaker 2: try to add that to Stream. Again. We do have

1:19:31 Speaker 2: some processing limitations, but if it's only one side should

1:19:35 Speaker 2: be able to do that if you're talking about stream,

1:19:38 Speaker 2: if you're talking about the mDNS themselves.

1:19:40 Speaker 1: We do have data.

1:19:41 Speaker 2: From again around the globe and we've used that in training.

1:19:47 Speaker 2: As long as you have RS data from satellite, using

1:19:50 Speaker 2: the tutorials that I showed here today, you should be

1:19:53 Speaker 2: able to produce your own product maps for that region.

1:19:58 Speaker 2: We're also as part of the land sat science team,

1:20:02 Speaker 2: we're also adding in data from some Latin American countries,

1:20:08 Speaker 2: including Mexico, and trying to validate our algorithms in these regions.

1:20:16 Speaker 1: This is one of the last questions we can take.

1:20:19 Speaker 1: In highly turbid Indonesian tropical river systems such as the

1:20:24 Speaker 1: Cupuas and Mahakam in Calimantan, where seed dom and suspended

1:20:30 Speaker 1: sediments are both elevated, how does the md AND handle

1:20:34 Speaker 1: spectral ambiguity between chlorophy a and si dom absorption in

1:20:38 Speaker 1: blue bends.

1:20:41 Speaker 2: Sure, so, we typically treat the MDN more as a

1:20:48 Speaker 2: black box. We don't really go into like the layers

1:20:52 Speaker 2: to see how it is span chlorophyll and seed dom.

1:20:57 Speaker 2: But you know the way it works is it's a

1:21:00 Speaker 2: empirical algorithms. So we have again a large data set

1:21:03 Speaker 2: with a wide range of chlorophy a, c DOM TSS

1:21:10 Speaker 2: values represented. So if we have similar chlorophyl a and

1:21:14 Speaker 2: seed dom combinations, we would expect it to perform relatively well.

1:21:20 Speaker 2: If we don't, it might not. But then you could again,

1:21:25 Speaker 2: if you have in such a data you're willing to share,

1:21:27 Speaker 2: you could email me and we could add it in

1:21:28 Speaker 2: to future trainings in the model.

1:21:34 Speaker 1: Great, thank you. I think we're almost at the end

1:21:36 Speaker 1: of today's session. Thank you all for attending this training,

1:21:41 Speaker 1: and once again we thank our guest speakers for their

1:21:46 Speaker 1: valuable time and information that they're provided. Also, thanks they

1:21:52 Speaker 1: are the team for their coordination and helping putting this

1:21:55 Speaker 1: training together. And we hope to see you at our

1:22:00 Speaker 1: next training and look forward to get suby results from

1:22:05 Speaker 1: you and feedback from you. Thank you,