NASA ARSET - Drought Prediction Using NASA Sub-Seasonal To Seasonal (S2S) Predictions

0:00 Speaker 1: And showed how SPI and VCI can help in monitoring drought.

0:09 Speaker 1: This brings us to today's session on S2S predictions. Overall objectives for today are that by the end of this session, we will recognize functionality of NASA's sub-seasonal to seasonal forecast system and data. and assess evolving drought conditions using given S2S temperature and precipitation prediction data for a region of interest. So we'll start with description of NASA's S2S forecast system and data, and then we will have a demonstration of analysis of S2S predictions of temperature and precipitation in QGIS.

0:48 Speaker 1: So next, we will start with S2S forecast system and data, And before I introduce our guest speaker, here's how to ask questions. So please put all your questions in the question box and we will address them at the end of the webinar. And feel free to enter your questions as we go. We will try to get to all the questions during the question and answer session. The remainder of the questions will be answered in the question and answer document, which will be posted on the training website about a week after the training.

1:24 Speaker 1: With that, I want to invite Dr. Andrea Mollard to talk about S2S predictions. Dr. Mollard, she is a research physical scientist here at Goddard Space Web Center in Global Modeling and Assimilation Office, or GMAO. She leads Seasonal and Decadal Prediction Group, managing the development of the next-generation atmosphere-ocean-coupled seasonal prediction system. She also is involved in development of GEO's five general circulation model physical parameterizations. And she is GMAO lead of the coupled chemistry model development effort with members of the atmospheric chemistry and dynamics branch here at Goddard.

2:09 Speaker 1: So with that, I ask Dr. Moller to talk about S2S predictions. Thank you. Andrea, take it away.

2:19 Speaker 2: So thank you, Amita, for the introduction.

2:23 Speaker 3: And again, as she just said, I'm Andrea Mollad, and I'm the lead of the NASA Global Modeling and Assimilation Office Subseasonal Prediction Group. I have the list of a lot of names down here, and different people on this list have contributed in one way or another, but deserve recognition for their contributions to this work.

2:47 Speaker 2: That I'm going to talk to you about.

2:52 Speaker 3: So just a little bit of an outline of what I'm going to cover today. I'm going to talk a little bit about what seasonal prediction is and how it fundamentally differs from weather. We're all kind of used to looking at and understanding what a weather forecast is. Seasonal prediction is a little bit.

3:09 Speaker 2: Of a different thing.

3:11 Speaker 3: I'm going to talk a little bit about our user community, who we give data to in a near real-time sense, and what they're doing with it.

3:20 Speaker 2: It's quite a varied list of users.

3:23 Speaker 3: I'm going to talk to you a little bit about the model features, and I don't mean the technical details of physical parameterizations or ocean models or things like that. I'm talking to you about, you know, as I will describe, seasonal is a a fundamentally stochastic animal. And so I'm going to talk a little bit about the ensemble forecast and what that tells us about reliability and for anticipated skill. And I'm going to devote a little time to talking about the output, the details of the output information, what we output, what it looks like, where you can find it, things like that.

4:01 Speaker 3: A note here, because some of the slides are going to talk about different versions of this system. We're about literally next month. We're going to make our transition from a GSS2S2 system that's been running since about 2017 to our new one, which is GSS2S3. It has a lot of upgrades. We are delighted with the improved forecast skill. I'm going to give you a little teeny glimpse of that and a huge extensive retrospective forecast suite, and I'll talk to you a little bit about what the potential benefit for users is of that.

4:39 Speaker 3: So first question, NASA, satellites, satellites measuring the Earth, but why is NASA interested in developing and maintaining a sub-seasonal, a sub-seasonal seasonal prediction project? What's the interest here? So the idea is that we use these models in conjunction with satellite and in-situ observations to study and predict the phenomena.

5:04 Speaker 2: And the idea here is.

5:05 Speaker 3: That our experimentation allows us to explore how to better use NASA or other satellite data to improve forecast skill. So to that end, we maintain a state-of-the-art system, and we're continually doing experimentation.

5:23 Speaker 2: On what happens if we add this data. Does that help the skill?

5:28 Speaker 3: And so we have a history, a long history of doing this, beginning before the current one. Every one of them had a coupled model, some kind of coupled data assimilation for some kind of reanalysis, and some kind of an ensemble perturbation strategy. The first bullet there, what we call pre-GEOS, that started, the reference is actually here, but this project started quite some time ago. This was actually a model that didn't extend from pole to pole. that fundamentally just covered the tropics, but was still used for seasonal prediction.

6:11 Speaker 3: The GEOS S2S1, 2012. Our version two came out in 2017 at higher spatial resolution, and our newest one is scheduled for release sometime next month.

6:27 Speaker 2: And so, just wanna talk a.

6:30 Speaker 3: Little bit about this phenomenon of seasonal prediction. The first sentence may sound a little technical, but it's not. The climate system is what we called forced dissipative nonlinear dynamical system. It's red. There's red noise all over the system. But we also understand that it's chaotic. And so we've all learned that because there's chaos in the atmosphere, after a certain amount of time, we can't predict things anymore. Okay, so if that's the case, and we're talking about a week to two week predictability, what is this seasonal thing about?

7:06 Speaker 3: Okay, and so the seasonal takes advantage of a few things. The tropics are so strongly determined by the ocean underneath it that it's not that sensitive to the initial conditions, so it's less chaotic. And so basically, The tropical, the tropical ocean also is like an integrated impact of what the overlying atmosphere does. And so basically because they evolve slowly, the ocean and the land too, this also extends this predictability out. So essentially we should be able to predict seasonal scale things in the tropics at least for as long as we can predict the ocean temperature, which.

7:52 Speaker 2: Is up to a year or so. Okay.

7:56 Speaker 3: The key to this whole thing is a concept that's called signal-to-noise ratio. How much of the variability is predictable? That's your signal. How much is not?

8:09 Speaker 2: That's your noise.

8:12 Speaker 3: And so the predictable part of this depends on these what we call sources of predictability.

8:18 Speaker 2: The ocean is one big one.

8:20 Speaker 3: The rest of it is generally unpredictable, and it's noise. So the whole game of seasonal prediction is how to maximize this signal to noise ratio. What do we do to get that signal to stand out from the noise? So in addition to the ocean, there are other sources of predictability, so-called sources of predictability. A lot of these are regular oscillating patterns in the atmosphere and in the ocean. You can think of something like El Nino that we all know that more or less every five, six years, we're seeing variations, okay?

8:57 Speaker 3: So as we see that we're headed into an El Nino phase, which is a warm sea surface temperature in the tropical Pacific, all of a sudden our prediction skill, predictability is better. And there are other of these kinds of oscillating motions. The Manjulian oscillation is one, and each one of this happens on different kinds of timescales. And so this gives us predictability, and then the question is, can we cash in on this to actually get prediction skill out to multi-year time scales? So here's back to this concept of signal-to-noise.

9:35 Speaker 3: What is it actually that's predictable? Can we predict the temperature at 3 o'clock in the afternoon three months from now?

9:43 Speaker 2: The answer is no.

9:44 Speaker 3: Okay?

9:45 Speaker 2: So what is it that's predictable? We're talking about probabilistic measures.

9:51 Speaker 3: How many air quality exceedance days are we going to have next month? We're talking about spatial averages. We're talking about time averages. This is a fundamentally statistical endeavor, and so we must have ensembles of forecasts to predict the probabilities. This also allows us to assess something that we call reliability. I'm going to talk about that briefly. Forecasts also need some kind of calibration. We adjust for model bias and things like that. Calibrated forecasts are more reliable. So the error that we expect them to have is the error that they're going to have.

10:29 Speaker 3: And this time averaging turns out that the longer the lead time, the longer the time averaging in order to get something predictable at these lead times.

10:39 Speaker 2: And so the idea is at weather, we don't do any of the averaging. Yes, we can predict the weather at two o'clock in the afternoon tomorrow. Okay.

10:47 Speaker 3: By the time we get to the one to two week averaging, the sub-seasonal leaves three, four weeks out, one to two weeks, seasonal is monthly averages. And if we're looking at decadal prediction, we have to do annual mean averages. So is next year's annual mean temperature gonna be higher or lower than normal? Those are the kinds of metrics that we're talking about here. So here are some of the users of our current system that we produce systems in near real time, which means essentially on the sixth of every month, we're producing the forecast from that month to go forward, nine months out.

11:29 Speaker 3: So there are some multi-model forecasts and intercomparisons. These multi-model intercomparisons and multi-model forecasts are always more accurate They're even more accurate than the best contributing model to the multimodal ensemble, always. We send things to a Korean intercomparison. There's some stuff at the Columbia University. We send them the El Nino indices. There's an atmospheric river intercomparison project that actually California water resources managers are using to predict precipitation.

12:07 Speaker 3: Sea ice prediction, drought briefings, okay, and sea level prediction, okay? Now, we don't have the capacity right now to just take our entire forecast and dump them out, and so these are collections that are available, data sets that are available upon request that we are now generating in near real time for these groups, and we post them on our FTP site. We also have a few operational forecasts that are taking our output from our seasonal forecast and using it as input. And so we've got the drought forecast for the Fusenet program.

12:45 Speaker 3: We've got the folks at Tennessee Valley Authority using it for water resources forecasting. And there's a group in Mekong River Basin, also water resources. And then we have a bunch of people using, running, trying to get experimental forecasts done with our system. We are the only near real-time system right now anywhere in the world that predict aerosol, and so the folks are using it for controlling dust sources.

13:12 Speaker 2: In the US.

13:15 Speaker 3: We have a landslide prediction group within Goddard. There's a group based on the connection between dust and meningitis outbreaks that are taking our forecast and making forecasts of those outbreaks, ecological foresting, pest population models, crop models, etc. We also, in addition to the ones that we're already in touch with, we also envision that we could be using this to reach a very much larger different user communities. Certainly applications in the healthcare field that we have not been able to make connections.

13:51 Speaker 3: And what I said before, we can predict seasonal anomalies of air quality, number of exceedance days, or other metrics that we would work with potential users to figure out what to send them. And then Obviously, spread of disease associated with dust and rainfall and things like that. Municipal planners, winter snow anomaly forecasts. Right now, they're basing it off some climatology in a lot of municipalities and making rather expensive decisions about snowplow trucks at the beginning of the winter.

14:23 Speaker 3: Forecasts of rainfall, soil moisture anomalies for more applications of water resources. And because of the large ensemble size, we have a lot of interest in this and expanding interest from the scientific community to use this to study predictability. And just a brief note about what I said before about the interactive aerosol model. We're the only game in town for this, for interactive aerosol model. There are other centers that are testing this and have seen some potential benefits, but right now the near real time is us.

14:58 Speaker 2: In addition, we.

15:01 Speaker 3: Include an aspect in our model that not only gets us the aerosol prediction, but it lets us also understand how the aerosols may be affecting the clouds and the rainfall. And so under certain circumstances, this really helps us with prediction of cloudiness and rainfall.

15:20 Speaker 2: An example would.

15:21 Speaker 3: Be the dust that comes off of Africa impacts the development of tropical cyclones in the Atlantic that season. And we are properly able to capture that. So we use an aerosol model called GoCard, the so-called interaction with cloud interaction. And basically, this is shown to result in credible forecasts of air quality, aerosol optical depth, and an aerosol-based particulate matter 2.5 micron. And in some regions, basically, if we're dominated by smoke, we're out of luck. But anyplace else other than that, we're good.

16:00 Speaker 3: We've also seen an increase in the skill of two-meter temperature and cloud water ice at moments when these effects are important. And so there's a group here that are studying this. Certainly, if anybody's interested in details, please reach out. I'll put you in touch with Dhoni Phan Varahana. And just a word a little bit about what I talked about with the reliability and this ensemble business. We're going to evaluate the confidence. I'm calling it confidence here. The word is reliability. We're going to evaluate the reliability of the forecast by comparing the ensemble spread and the mean error.

16:41 Speaker 3: What we would hope is that the ensemble spread matches the mean error. So if we make a forecast and the ensembles are all over the place, we're going to tell you we're not so certain about that forecast. If we make an ensemble forecast and the ensemble spread is tight and all of our ensemble members are telling us the same thing, we better be pretty confident that that's the case. For S2S2, we found that over North American land in particular, that was spot on. If you look at the curves on the right, if you look at the two on the right, precipitation North America and two meter temperature Europe, that curves that represent the error the spread is in the blue line and the green line is the mean error.

17:31 Speaker 3: Those curves are very close. That's telling us that in general, when the spread is small, the error is small. Over the tropics, version two did not do so well with this. And we've made some upgrades for version three that this is now spot on as well. And on the left, I have a little schematic of what we want our ensemble to do. If we look on the left of that schematic below under the heading of good ensemble, we're basically talking about an ensemble spread that, whoops, we're looking about a mean, the green line is the mean error and the blue line is the ensemble spread and the black lines are the sort of probabilities, the largest and smallest probabilities of the observations, okay?

18:26 Speaker 3: we need our ensemble spread, the truth is that red dotted line. And so what we need is we need our ensembles to stay within the truth, within a.

18:38 Speaker 2: Certain distance of the truth.

18:41 Speaker 3: As opposed to the one on the right that I would call a bad ensemble that has the truth somewhere far away, the mean error is huge, the ensemble spread is small.

18:51 Speaker 2: We regard that as a bad ensemble.

18:53 Speaker 3: So we're just trying to make the characteristics of the ensemble.

18:57 Speaker 2: Do a better job here. Okay?

19:02 Speaker 3: So what did we do for ensembles? The motivation for the change was what I showed you before. Our tropical Pacific, we're talking about, we called it too confident because our error was bigger.

19:18 Speaker 2: Than our spread.

19:19 Speaker 3: Okay, that says we're too confident. Our spread was too small, we made a bigger mistake, and we.

19:25 Speaker 2: Can't have that.

19:26 Speaker 3: And then later in the forecast, it turns out that they were underconfident. We were saying the spread was going to be huge, but our error was small. We changed our entire ensemble perturbation strategy.

19:37 Speaker 2: In order to get this better.

19:38 Speaker 3: Okay, in addition to that, in the extratropics, high latitudes, Europe mostly, but other places too, We were lower than some of the best models out there for S2S2 because we had too small an ensemble size. We had 10-member ensembles per seasonal prediction. And there's a clear evidence from this paper by Scafidol 2018 and others that the skill in these regions at these lead times increases with ensemble size. So we increased our ensemble size to 40, okay? Also, this dependence of ensemble size doesn't extend to the tropics.

20:19 Speaker 3: In the tropics, we're talking about the sea surface temperature, and the ensemble size doesn't help us. So beyond a certain month, we're reducing the ensemble, okay? And so what do we want these perturbations to do? The ensemble perturbations, we want them to capture the things that are going to have the error grow quickly.

20:44 Speaker 2: So if we have a structure like the ones that we're looking down at the bottom. So this is in the ocean.

20:53 Speaker 3: The horizontal axis is longitude at the numbers that we see. The vertical axis is depth of the ocean starting at zero and going down. And so what we're looking here is the depth of the ocean mix layer, the thermocline, that we know is deeper in the West Pacific and shallower in the East Pacific. That tilted structure looks like that. And so if we can make ensemble perturbations that are tickling something that looks like that kind of a structure, we know that we're going to capture the variations in that thermocline in our.

21:31 Speaker 2: Ensemble spread, which is what we want.

21:33 Speaker 3: Okay, if the variations are going to be big, we're going to make a big mistake, small, small mistake. And so we basically have these fast growing structures that we're looking to have those be the ones that we tickle. And we have a mix of them running throughout the forecasts. And this we found gave us a much better reliability estimate, especially in the tropics, than what we had before. Okay, so with all of that in place, our seasonal prediction suite we run forecasts out for nine months. We issue forecasts every five days for the sub-seasonal, for the out-to-a-monthly time.

22:15 Speaker 3: For monthlies, we're issuing them once a month. We have 40 ensemble members for the first three months, and then we sub-select from there and run out for the remaining six with only 10. And this is, you know, basically we're looking to get the biggest bang for our buck for the computations that we have. the computational space that we have. And so the 40 gets us a lot of what we want beyond the first couple of months, the first three months. The predictability in the exotropics is gone, and we stay with only 10 members for the tropics, for El Nino forecasts, and that's all we need.

22:52 Speaker 2: Okay?

22:54 Speaker 3: We have a coupled assimilation called GI Ocean that we have our retrospective initial forecasts from. Retrospective forecasts ran from 91 to 2024. The near real-time conditions are coming from our coupled reanalysis with this name, GI Ocean NRT.

23:11 Speaker 2: And again, once a month, these are issued more or less the fifth or sixth of the month. And now just a.

23:17 Speaker 3: Little peek at some of the improvements that we saw in the forecast skill that we're really liking. So this is a metric that we call the anomaly correlation. So we're measuring the month-to-month variations of the temperature. Higher, lower than normal, etc. And this anomaly correlation measures the correlation of the predicted and the forecasted anomalies for each month. And this is over the span of our retrospective forecast periods that we use for this. Across the top, we have June on the left and January global view on the right.

23:58 Speaker 3: And across the bottom, We have the new system, again, June on the left and global view of January on the right. And the deeper reds that we're seeing in that lower left picture, especially in some of the Western US regions, that very small region of really crappy skill with the.

24:17 Speaker 2: Blues and the whites is smaller. And we have deeper reds also on the East Coast. And we're liking that. In the Northern.

24:26 Speaker 3: Hemisphere winter, the skill over the US is substantially better. We're seeing those reds in there and we're seeing whites and dark pinks for the January from S2S2. So we've looked at this, we've looked at many, many, many other metrics. Anybody interested in any specific metric can come and talk to me, but we're very happy with the skill. And so let me switch gears a little bit now and talk about our output. And this includes how the output is organized, what's available, the file specification document, what that is, and when that's going to be released for S2S3, which is not yet, and where to find files that we stage for users.

25:14 Speaker 2: Okay?

25:15 Speaker 3: So S2S2 forecast is available if anybody's still interested in a period of retrospective forecast from 81 to 2020, and near real time from actually 2018 onward. S2S3 is available from 91 to 2024. The World Meteorological Organization has changed their base climatology period from 91 to 2020. And so going back earlier than 91 was very difficult.

25:45 Speaker 2: We have an extensive amount of output.

25:48 Speaker 3: There's no way that we can just post it all and have people just grab it. And so the current practice is that users request data to be staged, into specific collections and specific data sets for them, and the person to request this is by sending.

26:05 Speaker 2: Me an email, okay? So, collections.

26:10 Speaker 3: We have a concept in GeoS that we call collections, and here's an example of what a collection entails. Collection for us entails a name on the left, ATM is a three-digit identifier that says that these are atmospheric fields. INST says instantaneous, six-hour, six-hour frequency, global in scope, a lat-long grid at essentially half degree, and 49 pressure levels in the vertical. The specification for each collection basically has what I said there, that the name is indicating that gives you the list of what those 49 pressure levels are that these atmospheric fields are available at, and then a list of the fields themselves.

27:00 Speaker 3: So for each one of our collections, we have a specification that's like this, that will appear with a little bit more detail even than this, but this is the way that we tell our model what to output. Our model output is totally flexible, we can configure this however we want, This is how it's configured for S2S3. And just for you to see, here's a whole list of our collections. And except for that first digit three-letter identifier, we have prime average or instantaneous. We have the temporal frequency of the available output.

27:40 Speaker 3: Right now, everything is global. And a lot of things are output on that half-degree grid. But our ocean... runs on a quarter-degree grid, and so we have some things from the ocean that are output on that native grid as well. And the three-letter identifier is our fantasy about what things mean. They may or may not be self-evident. SFC, surface, RAD, radiation, MST, moist, AER, aerosol, OCN, ocean, that kind of thing, ice, sea ice, et cetera. But so the sample file name is going to have that three-digit identifier, and then will include an experiment number and the month and day, for instance, for a monthly average.

28:26 Speaker 2: Okay? So, GMAO S2S2, the file specification document.

28:32 Speaker 3: Is online and available. And we do have an FTP site for files that have already been staged. And I think Amita, in her demonstration, has downloaded some files from here and done some calculations that she'll show you how to do. GSS2S3, the file specification document is in preparation. We are scurrying around to get that thing done. And we are gonna have FTP access for files that have been staged in a slightly different location. And so now I talked about this file specification document that we're gonna prepare.

29:11 Speaker 3: In addition to general information about the grid, the file naming convention, For every one of our collections, it contains this entry that includes the variable name, the dimensions, and the units. And so this is a little bit more detailed than what you would get from that collection specification that I talked about before. So again, S2S2, all of this is out there. and available for people to have a look at. In general, S2S3 is outputting more, many more collections and fields than what S2S2 did.

29:50 Speaker 3: In particular, the prognostic fields are all the same, you know, what we're predicting, the temperature, the humidity, things like that. We have a lot of diagnostic fields that are present in this system, in the S2S3 system. Okay, so that's about all for my part of the talk. Thank you very much for your attention. And so the GMAO and the GS S2S forecast output availability is basically, we have different assorted ways of reaching out, of getting stuff staged for users.

30:25 Speaker 2: Get in touch.

30:26 Speaker 3: We're also sort of a nimble system, and so specialized requests for specialized fields are, you know, we're perfectly happy to do that. For a user that's interested in what our retrospective forecast skill of a particular field that they're interested in, we're also more than happily willing to provide that. So we can answer some questions later. In the meantime, let's pass this back.

30:52 Speaker 2: To Amita so she can do the demo using the S2S data.

31:00 Speaker 1: Thank you so much, Andrea, for the excellent presentation and information about the S2S prediction system. Next, we are going to have a brief demonstration of how to access and analyze S2S data. So to get S2S data for a specific period, it's either historical or present, the data are available, but you have to request them. You can send an email to Dr. Mollard with the period of your interest, and the data will be extracted from the GMO repository and put on an FTP site from where you can download the data.

31:45 Speaker 1: So this FTP site, if you go there now, you will see some data sets staged there in response to users' requests. These are also all open source data. So if you are interested in using them, if happen to be the same data set or same period, you can download and use this data. Otherwise, you can request data. So next, There are two demonstrations that we want to show. The first one is that we will examine maps of atmospheric anomalies to identify areas of dry or wet and warm and cold conditions. So for next nine months, how things are going to evolve is what is available in these pre-calculated maps that we will look at.

32:35 Speaker 1: And then next, we will examine surface temperature and precipitation predictions for next three months based on the most recent forecast made available by Dr. Mollard from S2S data repository. So let's start with this atmospheric anomalies. This page has a lot of information about As Dr. Mollard mentioned, S2S version 3 will be out soon and this website might be even easier to navigate, but we're going to focus on this S2S 2.1 system that's available here. If you can see, this starts in June 2024 ensemble mean and goes all the way to February of next year.

33:30 Speaker 1: The nine months of forecast based on May initial condition. Here you will see variable. So that two meter atmospheric temperature is available and precipitation. So anomalies are available and here you can pick regions, either global, North and South America, Southeast Asia, Australia and Europe. So you can zoom into any of these and to pick, let's just stay with temperature anomaly and look at the global maps first. You can submit here. We already have it. You can, let's look at the July ensemble mean.

34:15 Speaker 1: If you click on it, you can zoom in. Here's the color bar with anomalies. So all blues are negative and reds are positive anomalies, as you can see in degree Celsius. Clearly, in July, we see that here in southern and eastern US, it's quite warm. Warmer temperature anomalies are there in Mexico, in Central America, all the way to Brazil, you can see warmer anomalies. Similarly, in the southern part of Africa, something that Sean was showing last session, here you have positive warmer temperatures in higher latitudes in Western India, in Australia, everywhere you see warmer than normal T2M or 2 meter temperature.

35:16 Speaker 1: If you go to August, it persists here almost everywhere.

35:23 Speaker 1: Here also, if you go to Western India, I think after monsoon rain, it may cool down and maybe anomalies are not that strong. If you go down all the way to, say, January, February of next year, you can see that wintertime temperatures are much warmer over the US in January and February. And you can see where there are negative anomalies. So it's colder than normal as in north western Australia as you can see here. So this allows you to quickly browse through what's going to happen in next nine months.

36:06 Speaker 1: You can get qualitative idea by looking at this map that it's going to be warmer, colder. Similarly, you can look at precipitation Let's focus more on, if you look at the global map, you will see that most anomalies are stronger over the oceanic region, especially this ITCZ region. So if you want to look at land, it's best to focus on one of these areas. And I'm going to pick South America and then submit. So here, this is again monthly anomalies from June, July, August, all the way to February.

36:47 Speaker 1: And as you can see, There is deficit of rainfall. These are negative anomalies in the northern part of South America. As you go down to southern hemisphere summer, so say in January, February, Brazil develops or shows deficit in rainfall. It starts developing these anomalies and then you can see these large areas covered in negative anomalies, a deficit of rainfall. So here's how you can actually look at different continents and get some idea of what might be happening. Here it's about normal in their summer months.

37:34 Speaker 1: So also you will find information about S2S system here and a lot of other information that you can browse through. So in the next exercise, we are actually going to calculate anomalies of surface temperatures and precipitation using QGIS. If you have not yet installed QGIS on your computer, you can go ahead and do that after this exercise is over. Your prerequisite did request you to install QGIS on your computer. So for this analysis of calculations of anomalies, I have downloaded monthly forecast data for July, August, and September of this year.

38:26 Speaker 1: These are ensemble mean forecasts. And also what I've downloaded is so-called drift files with respect to which anomalies are calculated. All these files are available to you to download from the training webpage. And these files are in NetCDF format that we will be opening in QGIS and calculating anomalies in Next. So I have, I'm going to open a new project. Just for your information, the version I'm using is QGIS 3.16.14. Once you open a new project, we're going to go down here to coordinate. In this window, we're just going to type world.

39:18 Speaker 1: And you will get continental outlines with all the country shape files also outlined here. And if you want to see which country you're looking at, you can go here, identify features, click here, and zoom into region of your interest and click. And it tells you in this window which country you're looking at. In this case, it is Zimbabwe or Brazil. You can click on the country and you can see the country outline. Now the data that available to you are global data. So we are going to look at global data, but what you can do is zoom into the region of your interest and then just focus on that.

40:07 Speaker 1: So let's start by adding data to it. So I'm going to add raster. So let's see, I'm going to demonstrate calculation of anomalies for July and then you can do it for August and September at your convenience. You will have some time after this exercise to work on that or you can do it at your convenience later this week as well. So let's start with precipitation July 24 at CDF file.

40:40 Speaker 2: Open that.

40:42 Speaker 1: And add. So this file is added. Now, I also want to point out something that if when you load this file and for some reason, if it does not appear where the map is, what you will have to do is click on control, click on the raster, go to layer CRS. Because this is a net CDF file, there is no CRS attached to it, so you will have to pick a geographic CRS and that is EPSG 4326. So once you click on that, if the map and data are not appearing at the same location, they should after you do that. So you can click on the file and go to properties and look at information.

41:35 Speaker 1: This is an extent of this file is given here in longitude and latitude. So it is global data as you can see. Width and height, which is eastward and northward, north-south grid points, 720 by 361. So this is half a degree data, global. Then, of course, it's the format and data statistics are available here. Minimum, maximum values, mean and standard deviation. Most importantly, you can see the units here. It's in kilograms per meter square per second. If you remember, just a few minutes ago, we looked at the atmospheric anomalies website.

42:16 Speaker 1: The precipitation data were in millimeters per day. So to be consistent with that, I've converted this data in millimeter per day. And just a quick conversion is that if you have precipitation in kilograms per meter squared per second, to get the column of water in millimeter, First of all, you will divide it by water density, which is about 1,000 kilograms per meter cube. And so if you divide by 1,000, you will get precipitation in meters per second. And to convert it to millimeters per second, you will have to multiply it by 1,000.

43:06 Speaker 1: you are dividing and multiplying by thousands. So basically what it says is that when you have precipitation in kilograms per meter square per second, actually it is millimeters per second. The same values you will get that. Then you multiply by number of seconds per hour and hours per day. So click here on the layer. and go to Raster Calculator. You can select the raster and multiply it by number of seconds per hour and number of hours per day. And I have already saved this file. You can create that and then say OK.

44:04 Speaker 1: And now you will see numbers. Compare this in kilograms per meter square per second. And this is in millimeter per day. So you can remove this if you like and keep the layer which is millimeters per day. So let's set symbology so we can see values. Go to properties, go to symbology, here you will pick single band pseudo color and I would say 0.5 to say about 60 we can pick that here in the labels it should be maybe significant to two decimal points and you can have more intervals here so equal intervals and then how many classes you want you can see that And then you can invert this so that you can have lower values for blue and higher for red.

45:15 Speaker 1: And say OK. And you will see the data here. Now, what I'm going to do is click on the world map, go to properties, and select just simple outline rather than shaded continents. So you can just see the outline and data.

45:40 Speaker 1: So now you can see the data and you can zoom into the region of your interest and see what the precipitation structure looks like for month of July. You can see that numbers here say over the US, maximum rainfall is more like 2025. So now you will have to change colors to be able to see the structure. So you go to property, go to symbology and change the range here. Let's see. I'm going to say 15 and say okay. And now you can see that you can see more structure. So precipitation. You may have to change and reduce the range even further if you are interested in any of these regions.

46:31 Speaker 1: Now, I have already pre-calculated precipitation and temperature anomalies for July 24, so we are going to walk through those steps now. So again, for precipitation July prediction ensemble mean that we just saw, I have loaded in this project. The highest value here, the range is up to 50 millimeters per day. Next file that we have is this drift file for the same month. You can consider this as a long-term mean. So anomalies are found with respect to this file. So when we want to find anomalies, we take actual prediction and subtract this drift file precipitation from that.

47:21 Speaker 1: And so that's what we've done. So for that, go to Raster Calculator. select July precipitation prediction and subtract the precipitation from the long-term mean or drift file. So select that. And then you can select name of the output layer and save it on your computer, which I already have. So once you put that file name in, you can say OK, and then you will get a new raster, which is precipitation anomalies. And that's what is shown here. So first of all, this is actual July prediction. And this is prediction.

48:06 Speaker 1: This is drift file precipitation. And these are anomalies. And if you look at the scale here, it's minus 2 to 2 millimeters per day. So anything that is very dark blue, it's deficit either minus two or less and red is plus two or more more precipitation it's better you can zoom into the region of your interest here you can see that parts of Canada in the central US in Colombia and northern Brazil and southern Brazil there is rainfall deficit for July 24 if you move on to Southeast Asia, China, there are parts of these areas where deficit of rainfall is there.

48:56 Speaker 1: If you zoom in a little more, Bhutan, parts of Nepal and eastern India also have less rain than expected or than normal. So you can look at that. I have calculated temperature anomalies as well. So let's start by looking at July prediction of surface temperatures and scale shown here is 205 to 315 Kelvin warmer tropics and colder as you go away from the tropics. The next file is the drift data for surface temperature. So you can see that this is actual prediction and this is drift, actual prediction and drift for July.

49:42 Speaker 1: And when you subtract that to find anomalies, you get temperature anomalies and the values here are minus five to five Kelvin or degrees Celsius, it's the same. Now you can see that parts of the US are warmer, Canada, Again, Brazil, Greenland, many regions have warmer than normal temperatures. And wherever you see green and blue colors, that's where it's colder than normal. So now you can analyze both together precipitation and temperature. The regions which are drier and warmer are shown here, either in the US, in the South America, in Brazil.

50:35 Speaker 1: This is for month of July. The files available to you are for August and September also. So all the units are converted into millimeters per day. You don't have to do that. And anomalies are also calculated. So all these files are available to you. All you have to do is just load these rasters in QGIS, zoom into the region of your interest, change symbology to look at the patterns and see how you progress from July, August, September, how the temperatures are changing and precipitation are changing in the forecast.

51:18 Speaker 1: And that will give you some idea of what's unfolding in next two to three months. So that concludes our exercise, our demonstration for today. So with these demonstrations, we conclude today's session. So just to summarize, We had an overview of S2S prediction system and we also looked at how to analyze data. GMAO, seasonal prediction group, they use coupled earth system models and analyze in conjunction with satellite and in-situ observations to study and predict phenomena that evolve on sub-seasonal to decadal time scales.

52:06 Speaker 1: S2S prediction is not same as weather prediction as Dr. Mollard explained. It's a statistical approach where ensembles of forecasts are needed to predict probabilities. Multiple uses of S2S data are already there. They include drought forecasting and water resources forecasting, studying dust sources in the US, soil moisture, precipitation and temperature, prediction for landslide prediction, ecological forecasting, and data to drive a pest model. Data for users are tailored and distributed from an FTP site.

52:48 Speaker 1: They have to be requested. Maps of two-meter temperature and precipitation monthly anomaly predictions for nine months are available from the GMO site that we just saw. Also, examples of ensemble mean predictions of surface temperature and precipitation anomalies were analyzed in QGIS. We looked at the month of July. You can look at August and September to see where dry and warm conditions might persist over the next two to three months. And specific data can be requested by contacting Dr. Andrea Mollard, and her email address is available here.

53:30 Speaker 1: Now, next part three will focus on overview of NASA Earth Exchange Global Daily Downscale Projections, or NEXT-GDDP, for the Coupled Model Intercomparison Project Phase 6, CMIP-6. So we will look at how to access and analyze NEXT-GDDP CMIP-6 climate projections. We'll look at precipitation and temperature and see how long-term drought conditions will evolve under climate change. A note about homework and certificates. There will be one homework assignment and that will open on 1st of August and that can be accessed from the training webpage.

54:12 Speaker 1: Answers must be submitted via Google Forms and the homework due date is 15th of August. Those who attend all four live webinars and complete the homework assignment by the deadline will receive a certificate of completion in approximately two months after the training is over. With that, we want to thank Dr. Andrea Mollard. Andrea, thank you so much for your time and for the excellent presentation about S2S prediction system. Also want to thank you for making the data available to us for analysis.

54:50 Speaker 1: Here is the RSAT website link. and we are on the social media and these are our sister programs you can look into. And that brings us to the end of today's session. So I hope you were able to work a little bit on the exercise. And now we can go to question and answers. I'm going to go through all the questions one by one and see how many we can address.

55:28 Speaker 2: We have Dr.

55:28 Speaker 1: Mollard and she will be answering these questions about S2S system. So question one, I'm going to start Andrea and you can answer these questions. Will the GEOS S2S 3's hindcast climatology spanning 1991 to 2020 also be released in August?

55:57 Speaker 1: So you can unmute yourself and just verbally answer the question, Andrea.

56:00 Speaker 3: As I wrote here, the answer is yes. Frankly, we here at NASA, as in many, many places now in the government, we're fully committed to this open science. Any data that we produce should be available to the public. Our only obstacle is having enough public-facing file space to get everything out there. So the answer is yes. This is not that extensive a data set compared to others. We have access to something that we call a data portal. We're working out how to do that, and we will make that URL public.

56:35 Speaker 3: And we'll have a little readme on our website about how to access the data through the data portal.

56:45 Speaker 1: Thank you. Question two is also for you. How long will Jios S2S version 2 be available after the release of version 3?

56:52 Speaker 3: Yeah, we're going to keep it running probably until end of calendar year. We have a user, one user in particular, that needs this to be running for a while because they need to download every single one of our ensembles through the whole retrospective forecast period, and that's going to take quite some time.

57:15 Speaker 2: So we're going to keep S2S2 running for them.

57:19 Speaker 3: The plots on our website and the data that are going to a lot of our individual users are going to flip over to be from S2S3 either next month or the month after.

57:33 Speaker 2: Nice.

57:33 Speaker 1: Question three is, what are the possible ways to get the ensemble to look good if it is not?

57:40 Speaker 2: Right.

57:40 Speaker 3: So this is the calibration that I talked about. So we generate the ensemble and we do some sort of calibration. Frankly, we remove the drift. That's the simplest calibration that.

57:50 Speaker 2: You can think of.

57:52 Speaker 3: But, you know, this gives us essentially an unbiased anomaly. There are many other methods out there. The National Multimodal Ensemble, for instance, applies a rescaling of the variance as well. And there are many, many, many different machine learning-based methods out there and still under development. So this is an open question for how to adjust the ensemble. But again, we're doing and advising the simplest one, but certainly there are many options out there.

58:33 Speaker 1: Great.

58:33 Speaker 2: Thank you.

58:34 Speaker 1: Question four, is GEOS S2S an open access model? Can I use it for Kazakhstan?

58:42 Speaker 3: So a couple of things. One is the output is global. So if somebody is looking to use the output from our forecasts over Kazakhstan, then they should be able to. It seems that there may be an issue of resolution, horizontal resolution related to that, but certainly it's available. In terms of somebody else running the model, We're an open access model only in the sense that anybody can download the code.

59:08 Speaker 2: We have it up on Zenodo.

59:09 Speaker 3: We have it up on a couple of different, you know, version control, SVN, and Git, and a few different ways to download it. But in a practical sense, it needs parallel computing to run at any reasonable resolution, so some kind of a cluster. And there's a lot of auxiliary data that reside on the NASA computing machines.

59:31 Speaker 2: That we need to run it.

59:33 Speaker 3: And so in a practical sense, it's very difficult for somebody to run it somewhere else. We have options to run it at the two NASA computing centers. We're also not a community model in the sense that some others are where there's a whole office full of people to provide help.

59:51 Speaker 2: On how to download and run things. Yes and no is the answer to that question.

59:59 Speaker 1: Yes, I think so. Question five is, can the GEOS S2S model be applied to local scale, small scale studies? What is the special resolution of its output?

1:00:12 Speaker 3: Yeah, so our model runs It doesn't run on a lat-lon grid, but we interpolate our output to a 50-kilometer lat-lon grid for atmospheric fields, temperature, precipitation, things like that, 25-kilometer ocean resolution, sea surface temperature, sea surface salinity, and any fields at depth in the ocean.

1:00:34 Speaker 2: We provide it 25 kilometers.

1:00:36 Speaker 3: In order to get something at smaller scales, you'd need to go through some kind of downscaling exercise. There are a few users around that take our output and do downscaling, but what they're doing is very specific to their specific downscaling application. I don't have any kind of sort of general recommendation. This is how you would downscale in general for all uses. So that's something that would have to get worked out. There are dynamical and statistical downscaling methods out there.

1:01:12 Speaker 2: That's not maybe as.

1:01:13 Speaker 3: Useful as you would like, but I don't have a better answer for that.

1:01:18 Speaker 2: Yes.

1:01:19 Speaker 1: In next session, we will have one statistical method that you can learn about from NextGDDP. That's the downscaling effort. But as Andrea said, it's really you have to be sure that for your specific application and region, it's applicable. I think downscaling is a whole area of research and application. Question six, are S2S 2.1 maps available for download as GeoTeX?

1:01:49 Speaker 2: So, no.

1:01:52 Speaker 3: What we've got is we upload a bunch of. png images on our website that somebody could grab, but that's a very limited number of fields, and there's.

1:02:01 Speaker 2: Nothing in GeoTeX. Okay.

1:02:07 Speaker 3: I would imagine if we found some kind of very broad use for.

1:02:10 Speaker 2: This, we could find some way to get this done. But as of now, the answer is no.

1:02:16 Speaker 1: Question seven is how do you check for Africa?

1:02:20 Speaker 2: So again, the model is global. The data are global.

1:02:24 Speaker 3: My understanding is that some of the demonstrations may have been for U.S. or something, but certainly Africa.

1:02:32 Speaker 2: Is part of.

1:02:37 Speaker 2: Any of the global fields. So that should just be included when you download data.

1:02:47 Speaker 1: So the data you have, the sample data to work with, they have, it's global. So you should be able to look at Africa, zoom in. And the question eight is, can they be viewed? So this is about data file. Can they be viewed using Panoply since it's NetCTF? And answer is yes, you can open these files in Panoply.

1:03:09 Speaker 2: Okay, thank you.

1:03:11 Speaker 1: Question nine, is there a Python notebook you could recommend for this exercise for people not expert in QGIS?

1:03:21 Speaker 3: I can't help with this. I'm wondering if one of you guys can.

1:03:25 Speaker 1: So, I mean, that is a good question. We probably will think of that instead of QGIS, slowly switch to Python and notebook-based calculations. So we'll definitely take this suggestion for our team to work on. Yes, we'll do that. Right now, we don't have it, but this is something that we can think about. Question 10. It is very interesting to me that the forecast predictions are listed as ensembles. This leads me to ask this specific question. What specific motif or patterns are within this forecast prediction and exist across North America for landslides?

1:04:12 Speaker 3: This answer is a very long-windowed way to say I don't know. But there are people that either are still or were using our output specifically to look at landslides. And I can't remember exactly who, but I can find out quickly. And I could answer, you know, in whatever format you guys would like. But I think this question gets posed to them.

1:04:35 Speaker 1: Yeah. So, yeah, if we can, you know. find out who is using, maybe we can ask this question to them.

1:04:45 Speaker 2: There's somebody in the hydrology branch, and I just can't remember who. Okay.

1:04:49 Speaker 1: Okay. Question 11, what are some examples of custom data requests that get uploaded to FTP?

1:04:58 Speaker 2: Yeah.

1:04:59 Speaker 3: So we have some of the FTP fields that you'll find there are Courser resolution, because there are some multi-model ensemble groups that want everything at the same resolution, so we coarsen our fields. Others just want subsets because they're working at a university or at a small place that can't handle huge amounts of data. We do a little calculation for them on the fly. One example of this is there's an atmospheric river intercomparison project that we contribute to that the California water resources folks use.

1:05:41 Speaker 3: And we do a calculation from our outputted moisture and momentum fields, the winds, to perform what they call an integrated moisture transport and then upload that to them. And there are a few examples of little calculations that we make just on the fly, horizontal interpolations. Most of it comes under either subsetting horizontal interpolation for little calculations.

1:06:11 Speaker 2: Great.

1:06:12 Speaker 1: Thank you. Next question is one question related to the seasonal forecast. Is there any long-term or 30-year or longer daily data set of historical variables like air temperature and precipitation?

1:06:27 Speaker 3: I'm interpreting this question not as is there a long-term observational data set, but is there a long-term record from our forecasts?

1:06:36 Speaker 2: And the answer to that is we produce three hourly fields of two.

1:06:40 Speaker 3: Meter temperature and precipitation or all the model levels of precipitation and then average them later for seasonal forecasts to monthly averages or so. So it's a lot of data, but certainly at the request of user, we could give these three hourly fields for the ensemble members. Now, I would be very wary to give daily fields out or three hourly fields out for anything beyond the first 45 days at most of the forecast, because beyond that, the daily fields are just not predictable. And so I would have a hard time standing behind them.

1:07:24 Speaker 3: Given this, if somebody wants them anyway, of course it's available. Data are public and they can just have them. But we have these three hourly fields for the new forecasts, but also we have them for our entire record of retrospectives.

1:07:38 Speaker 2: It's a lot of data.

1:07:39 Speaker 3: But yes, there's a record and a user can ask for it.

1:07:46 Speaker 1: Great. Question 13 is the anomaly layer. shown for July for precipitation, there are some areas, countries in South America and Middle Eastern countries that were white, which represent zero anomaly. What is the reason for that? I do not think not having a deficit would be the reason. So the white color was between 0.4 to 4 millimeters per day, if you go back and look at it. So anything that was in that range around zero was white. So what it says is that the anomalies were smaller compared to meat.

1:08:24 Speaker 1: So what we saw was more like above minus one and plus one, they were colored, a lighter color between 0.4 and those numbers. So white color basically shows that the anomalies are non-zero. They are around zero, hovering around zero, between minus 0.4 to plus 0.4. And you can change that range. That's what I was talking about, that if you change symbology, you can do that. That range you can change. What you want to keep white or you don't want to keep anything white. Question 14.

1:09:07 Speaker 3: I would just add to that a note that I wrote there. This only means that the expected rainfall is the usual amount. The usual amount is the number that appears in that drift file. So if you want to know exactly how much rainfall there is, you can go to that drift file and see, because what we're telling you is that that's the number that we're expecting next month or the month after.

1:09:28 Speaker 1: Yeah, so actually, you know, that's something that's a great point that I actually wanted to make when we were going through the exercise is that you know, wherever you see anomalies, those numbers should be compared to like what the actual rain is like in drift, right? So that you can see what the percentage deficit there is compared to long-term mean or, you know, in this case, this drift. So yeah, that's a good point. So it's more or less normal with respect to that drift data that you have, the white colors.

1:10:06 Speaker 1: The next question is, it was a little bit too fast to follow the calculation. So yes, it's just the kilogram per meter square per second precipitation. It's equivalent to millimeter per second. And so you multiply a number of seconds in day and you get millimeter per day. But the files that you have, they are already converted. So you don't have to go through that step. Question 15 is, do you only produce precipitation and temperature data under GEOS S2S?

1:10:41 Speaker 2: Yeah, we have quite an extensive list of output fields.

1:10:45 Speaker 3: We are predicting the winds, the temperature, the humidity, as well as all kinds of what we would call diagnostic fields, surface radiation, all kinds.

1:10:54 Speaker 2: Of things like that.

1:10:55 Speaker 3: That document, we call that our file specification document. provides the whole list for GS S2S2. The S2S3 output is going to be totally reorganized into different groupings and things like that. We're preparing that document now, but you can imagine that it's at least these fields plus quite a number more.

1:11:19 Speaker 2: Quite a few more.

1:11:22 Speaker 1: Thank you. Question 16 is, while performing exercise part one, I encountered some error in GEE. Can someone guide who I can contact to correct those errors? I think you can email it to us. Please, yeah, email to RSET email address and then we'll forward it to Sean McCartney or anybody else who can answer your question. Question 17 is, as the forecast provides data on the anomalies, how do you use that data to calculate drought indices like SPI? Yeah,

1:12:08 Speaker 1: go ahead, Andrea.

1:12:10 Speaker 3: So that's an interesting question because the reason that we do all of these anomalies and things like that is that provides the reliability of the forecast, the correspondence between the error and the spread. when you're computing a particular statistic or a particular index, there are a few different ways to do it. Certainly, you could take the raw uncalibrated output and compute the SPI and do the same from the Heimkast. And so you would have an SPI anomaly more than, you know, higher than usual, et cetera.

1:12:46 Speaker 3: But there are a few other options that one could think of. And so that would be something to reach out to us, please, and we'll try to figure something out. You know, the idea would be to try a few different options and see which one provides you the most skillful forecast of SPI.

1:13:08 Speaker 1: Great. That's great to know. Question 18 is, I was wondering if there is any package in R that allows us to download GMO S2S data? So, Andrea, I think that probably there isn't a package that would allow you to download GMO S2S data. You have to request them to be staged on an FTP site.

1:13:36 Speaker 2: Yeah.

1:13:37 Speaker 3: Either grab what we have or request that they be staged, or there is some stuff that we're going to put on that data portal. And, again, as we finalize all of this, we'll let people know.

1:13:53 Speaker 1: Question 19. do we need to perform the raster calculation to convert units? And no, for the files you have for this exercise, you don't have to convert any units. They are already converted. If you are interested, if you download data from the FTP site and you want to convert the units, you will have to do the raster calculation.

1:14:18 Speaker 1: Question 20, I cannot download the anomalies. I'm not sure this is the GMAO S2S2 current. Andrea, do you recognize that question, question 20?

1:14:36 Speaker 3: I see the question, but I'm not quite sure that I understand what somebody is trying to do.

1:14:43 Speaker 1: Yeah, this is the atmospheric anomalies page, I believe, and those are PNG images, so I'm not sure.

1:14:53 Speaker 3: Right.

1:14:53 Speaker 1: Yeah, yeah. So they're not digital data. So that way, you know, one of the previous questions answered that, that we cannot download as digital data. Question 21, what is the best scale to work with this data set in the forecast? Not sure what you mean by scale, if you can be more specific. Where can I get VPD data for 2000 to 2020? It would have to be calculated from temperature and humidity. What's the chronological arrangement of the layers?

1:15:54 Speaker 3: So we interpolate our data from model levels to pressure levels to provide output, and they're sequenced bottom to top, but the metadata in the NetCDF file should tell you which level is which.

1:16:16 Speaker 1: Question 20, yeah, 25. Are there any further recommendations to consider when modeling seasonal droughts such as parameters, scales, and data?

1:16:31 Speaker 3: I am not an expert on drought prediction, so this is not a question for me. I'm a reader of the papers on the subject, so I know that there are lots of different ways to measure the different kinds of drought, the agricultural drought, the meteorological drought, etc.

1:16:54 Speaker 2: So I don't know how to answer this question.

1:16:57 Speaker 1: So Andrea, I had a question about that. Does S2S version 2 or 3 will have surface soil moisture?

1:17:05 Speaker 2: We have soil moisture.

1:17:06 Speaker 3: So certainly, that's one of the parameters that will be downloaded. We actually have three soil moisture levels. We have a surface soil moisture, a root zone, and a deep.

1:17:20 Speaker 2: So all of those are available.

1:17:23 Speaker 1: So I think... not about the scale, I'm not sure, but if you look at precipitation, surface temperature, soil moisture, then you will have a good idea about meteorological and also agricultural drought based on these data sets. So at least that you can look at seasonal anomalies you can look at if you want more seasonal droughts.

1:17:50 Speaker 3: We are particularly pleased with our soil moisture initialization methodology. Frankly, it was developed here quite some time ago, and now other centers are starting.

1:18:03 Speaker 2: To adapt this method. But our soil moisture initial conditions are better than many.

1:18:12 Speaker 1: Great. So I interpret scales means temporal scale, that if you want to look at seasonal drought, do you use daily data to begin with? But you can use monthly data, right? to look at seasonal evolution. Yeah.

1:18:31 Speaker 2: Yeah. Yeah.

1:18:33 Speaker 3: One caveat for this that I just want to have clear, which is we do not.