Demonstration of Regional Drought Monitoring Tools

0:00 Speaker 1: This is a great example of where one portion of the nation might be within that exceptional drought category. And it's really clearly identified here. And so a drought declaration could be made using this information and some resources and mitigation activities could occur. For this project, we really cared so much about having and ensuring that it was partner driven. And so we had two different kind of training activities. And the first was about halfway through the project to do some beta testing and to get feedback on what we should include, what not to include within the tool, what's really useful for decision-making.

0:44 Speaker 1: We also have this online user guide that was made available for folks to take a look at and do some calculations and demonstrations on their own. So now I will jump into a really brief demonstration of Climate Engine and.

1:01 Speaker 2: The DESET tool.

1:07 Speaker 1: So I'd like to start this demo by first navigating here to the Climate Engine main page. And as I mentioned, I really want to point out that most of these data layers and analyses that I will highlight through DSET are also available globally through the Climate Engine website shown here. So all you need to do is, in order to launch this application, is to have an account. So it just requires a Google account. And you can come right here to climateengine.org and launch the application and if If you're logged in like I am, you'll start to see the interface here, and they have some really beautiful guided and video tours available.

1:53 Speaker 1: So I really encourage you to take a look at that if you're interested in some of these things that I'll be highlighting with DSET. So for this regional example, I just want to navigate to NASA's Western Water Applications Office. And again, this is our main website for the program that I spoke about earlier. And within our water portfolio, you can see we have a link to the Navajo Drought Tool User Guide. And this is the beautiful user guide that I mentioned where we have the place to visit the tool.

2:28 Speaker 1: But I also wanted to highlight we have the introductory video, a longer version than what you saw today. And we have videos and hands-on guided exercises for assessing drought, performing crop analyses, looking at the status of lakes and meadows, and gauging the state of snowpack. So you can come here and take a look at all of this information as well. I will just come back up here and go to our drought tool directly. And you can see here that it looks very similar to the main page for Climate Engine.

3:04 Speaker 1: So you have along the left-hand side here the ability to change some of the variables for the map layer and then also make a graph. So we'll just do a couple of really quick analyses here. You'll notice that we have the Navajo Nation and the chapter boundaries centered on the map. One thing that I'll just do here is change the base map so we can see the data a little more clearly to roadmap. So you can, again, change all of these similar features here or in Climate Engine as you'd like. Now we can see the outline of the chapter boundaries.

3:43 Speaker 1: I also want to mention that we have available the U.S. tribal nations boundaries as well. So you can turn this layer on and you can see all of the other federally recognized tribal boundaries, and do any kind of analyses and mapmaking over similar regions as well. I am just going to go back to our chapters layer that we have here and do some analyses. Now, as mentioned, the most important variable for the Navajo Nation drought emergency management is the six-month SPI, or Standardized Precipitation Index.

4:21 Speaker 1: And I want to start by adding that layer for a particularly interesting time period in water rights management for the Navajo Nation. On June 22, 2023, the Supreme Court ruled against the Navajo Nation over claims that the federal government had failed to assert the tribe's need for water access in the Arid West. There is a long and complex history of treaties, agreements, and court decisions that over decades have dictated how the waters of the Colorado River are allocated. And the situation is compounded by a changing climate and scarce water supply.

4:56 Speaker 1: So I want to start out by taking a look at the grid met drought four kilometer six month SPI as of the June 22nd, 2023 start date. If you come over here and take a look at the variables, you'll see that that is the default. And the only thing I need to do here is change the time period to June 22nd. So we have 2023-06-22, and I can click on Get Map Layer. You'll notice here that it set it to the closest valid date where they've computed the six months prior to this date. So it's essentially the average SPI over the six months prior to this June 22nd date, and in this case, June 24th.

5:51 Speaker 1: So what you'll notice here is that it was a particularly wet period on the Navajo Nation with the SPIs corresponding to a drought scale that's moderately wet in some portions and extremely or exceptionally wet in other portions. And you can look at that by clicking right here on the question mark. So this was a fairly wet time period as dictated by the monsoon. But now let's take a look at three months later in September of the same year. So all I want to do here is change this date to September and then click on get map layer.

6:39 Speaker 1: So now you can see that most of the eastern half of the Navajo Nation is in exceptional drought, showing how quickly these conditions can change. You can also take a look at other variables here along the data sets. You can take a look at different types of data and models. And within GridMet, you can take a look at other kinds of drought indices as well. You can also take a look at precipitation over this time period and do some calculation there as well. So for example, what we can do is we can take a look at the grid met daily and take a look at the precipitation, which comes up as the default here.

7:29 Speaker 1: And if you look over the same time period, you can take a look at the percent of average precipitation. So what we can do here in the processing over the total and the statistic is we can look at the percent of average conditions. And we can look over the same time period where we were examining the SPI from March to September of 2023. So you can see here that the custom date range has defaulted to that six month time period. And we can take a look at the Gitmap layer as well.

8:22 Speaker 1: So here you can also, now we're looking at the precipitation percent of average and you can see that in many portions of the Eastern Navajo Nation, it's much less than average. You can also get a specific value by clicking on get value and moving it to anywhere here on the nation and clicking show value. And so you can see what percent of average the precipitation was here In this case, it's almost 35% of average for the state range. Another difficult time period for the Navajo Nation and the world was during the COVID pandemic.

9:00 Speaker 1: And this year was particularly hard with lockdown and lack of rain to support crops and livestock. So if we take a look at May 2020, you can also see some variable conditions when considering the SPI. So let's just go back here to the grid met drought. And we're going to take a look at the 180-day SPI.

9:27 Speaker 3: And we are going to take a look at May of 2020. And click get map layer.

9:44 Speaker 1: Here you can see those variable conditions where it's slightly wetter in the southwest and slightly drier in the northeast. And again, if we look then towards September, for example, of the same year, we can see that this was a really exceptional drought in this region. And this was especially challenging due to the lack of potable water and lockdown constraints. And so these types of data alongside understanding the community challenges can help us understand the drought conditions. I also want to mention that there are many other variables available here.

10:30 Speaker 1: You can look at remote sensing data and you can look at Landsat, MODIS, Sentinel, and you can look at things like the vegetation index and many other data sets available. And in particular, on the Climate Engine website, there are even more data sets than we have available. So I just, the final thing I want to highlight is the ability for Climate Engine and DSET to make graphs. So if we click on make graph here, and let's say we're interested in examining the precipitation versus the SPI in the White Rock chapter, which is near Canyon de Chelly.

11:10 Speaker 1: So what we can do here is we can do a two variable analysis. We could pick a chapter here and we can take a look at, for example, if we're interested in the Canyon de Chelly region, we could take a look at the Chinle chapter and you'll see we are directed towards this region in our map when we select that. We can take a look at the climate and hydrology And we can take a look at the precipitation versus the SPI. So if we take a look at the grid met daily precipitation, we can, let's just use the 2020 water year as the example.

11:56 Speaker 1: So we can start off with October 2019 to September

12:14 Speaker 1: And then for variable two, we can look at the SPI over the same time period. So you can see here we have the grid met drought SPI, and we can just go ahead and copy and paste the same date range. And what we're going to do here is just make a simple graph looking at these two variables and how they might change over time. Now we click get time series. Great, so now what you can see is the precipitation in blue and the six-month SPI in green for the Chinle chapter, so that very specific chapter within the Navajo Nation.

13:01 Speaker 1: You can see how the SPI increases after many different rainfall events and then starts to slowly decrease after a pretty dry period, and there is clearly a lag effect of the six-month SPI versus the precipitation. I also want to mention you can download data as PNGs, so you can download these kinds.

13:28 Speaker 3: Of graphs as an image, or you can download the data here as well.

13:35 Speaker 1: And that just highlights some of the features of DSET. and the importance of these data during these really interesting time periods on the Navajo Nation. So I do encourage you all to take a look at DSET, take a look at Climate Engine, if you're interested in doing these kinds of analyses. So that concludes the demo for today. I want to thank you again for having me. And now I will pass the proceedings over to Erica. So over to you, Erica.

14:05 Speaker 3: Thank you.

14:09 Speaker 4: Hello, everyone. This is Erika Potest, and I am an instructor with the RSAT program and a scientist at NASA's Jet Propulsion Laboratory. The next demo will be focused on the use of a tool known as the Sustainable Forest Management Information System, or SFMIS, which generates land cover classifications. So in this case, we will generate land cover classifications during and after a drought to assess the impact on land cover. This tool was developed through a NASA ROSES ecological forecasting grant.

14:48 Speaker 4: This demonstration will be focused on how to run the SFMIS tool, which, as mentioned, is a land cover classification tool, in order to assess vegetation cover change due to a drought. The demonstration will be over the country or part of the country of Panama, specifically for 2015, which was a strong El Niño year, which brings drought to the country. And these drought conditions were also pervasive in the previous three years where there was unusually low precipitation. We will then focus on the year 2020, which was a slightly wetter than normal year.

15:32 Speaker 4: And we will generate land cover maps for these two years and compare them to assess the difference in land cover due to drought. The two figures that you see here show precipitation anomalies for 2015 and 2020, where blue are drier than normal conditions and red are wetter than normal conditions. Note that Panama has two distinctive seasons, the wet and the dry season. And in this example, we are working with satellite data from the wet season only, and we will also be focusing on the central to eastern portion of the country where the impacts of the drought were the strongest.

16:15 Speaker 4: The tool runs on Jupyter Notebook. It is open source and runs on open source data. And so during the demo, we will walk you on how to set up the tool and how to run it. The demo will only use optical data from Landsat 8. However, the tool can also run with other optical data such as Sentinel-2 or a combination of Landsat 8 and Sentinel-2, as well as with radar data such as Sentinel-1 or PALSAR data, data sets that are available on the Google Earth Engine database. The tool also uses ancillary data sets to help inform the classification In this case, a digital elevation model.

17:01 Speaker 4: And the classification algorithm used is a supervised approach based on random forest. The training is from a land cover reference map that was generated by the Panamanian Ministry.

17:15 Speaker 5: Of the Environment.

17:19 Speaker 4: So the tool has different classification approaches. The first classification approach, which is the one that we will be showing in today's demo, is based on the medium composite of all Landsat 8 images over our study area during the wet season, which spans from around May through early December. And the reason for focusing the classification on the wet season is because land cover can vary between the wet and the dry season. The second approach classifies each image within the Landsat 8 collection individually.

17:56 Speaker 4: And then the class assigned to each pixel, whatever each pixel was assigned, the final class is the majority of the classes for that pixel. So that's the final class that is assigned to each individual pixel. And then the third classification is focused on combining data from multiple sensors. And in this approach, say we're using optical data from Sentinel-2 and Landsat-8, then we can then classify Sentinel-2 and Landsat-8 image collections separately and combine their classified information. So basically, the class assigned to each pixel was decided based on the majority vote for a class.

18:46 Speaker 4: And this approach yielded the highest classification accuracy from all other approaches. And also by construction, this approach is guaranteed to have the highest coverage. Let's get started with the demo, and you can access the notebook through the NASA RSET GitHub page, and the link is on the slide here. I will now turn it over to Dr. Rita Majumder, who will show you how to set up and run the tool.

19:17 Speaker 2: Dr.

19:17 Speaker 4: Majumder supported the development of this tool while he was a PhD student at the University of Maryland. He is currently a postdoctoral fellow at North Carolina State University, working on several problems pertaining to the statistical modeling of extreme weather and climate change. He was a member of the 2021-2022 Future of Fire cohort of the USGS Climate Adaptation Postdoctoral Fellows. He has a PhD in statistics from the University of Maryland Baltimore County and specialization in spatial statistics, Bayesian inference, and probabilistic machine learning.

19:56 Speaker 4: Welcome, Dr. Majumder. It's a pleasure to have you here.

20:04 Speaker 2: Thanks for the introduction. I will be walking all of you through the demo of the SFMS tool for land cover classification. We're gonna be using a few tools in this process. This assumes you are logged into Google Colab, into your Google Drive account, as well as you have an account at Google Earth Engine. There is some additional setup that we need to go through, but I will be walking you all through that. Okay, the first thing you need to do is go into your Google Drive. This could be done through the web browser.

20:40 Speaker 2: like I have here, or it could be done through your offline Google Drive software, and create a folder inside there called GEE underscore exports. This is where the output from the code you run will get exported. I'll briefly discuss later that it can be changed and the name, you can modify that, but for the time being, create this folder inside there and make it case sensitive. So the same case that you see in this case. So that's kind of step one. The second thing is you will need to download the Python notebook, which is run as a Jupyter notebook from the RSAT page.

21:24 Speaker 2: And you should have the exact link there. Once you download it, you can download it on your local drives, but I will recommend downloading it within your Google Drive. Doesn't need to be specifically within this folder, but since we will be using a lot of Google resources throughout, it's easier for the pipeline if the file itself is downloaded within your Google Drive. Okay, so now that both of those are kind of set up, we will be mostly working through two different websites. The first website is going to be Google Colab.

22:02 Speaker 2: I'm going to open that up really quick. The links to that is collab.research.google.com. This should already be set up like a Google Drive. You can see on the right hand side if you're logged into your account. I would recommend doing that because you'll need access to run a bunch of the things. So once you open that, you have an option to just open up the Jupyter Notebook file that you've downloaded for this training. It's technically possible to open it up directly from the GitHub. but for a couple of security reasons, I will not recommend that.

22:40 Speaker 2: Ideally, you want to go to the Google Drive tab and open up the file from there. Usually, Google Drive is good at showing you Jupyter Notebook files that are already there on your Google Drive. If it doesn't, you can use the search function and search for it. In our case, the file we will be using is called LandCover Classification Landsat 8. Click on that. That will open up. Now, a couple of brief things to go through. The second thing we will need in terms of setup is for Google Earth Engine. Now, the website for Google Earth Engine, again, is code.earthengine.google.com.

23:21 Speaker 2: You should be able to see that you are signed in. You need to at least be signed in beforehand. The second piece of the setup that needs to happen within Google Earth Engine is you need to set up a project that all the code you will be running is associated with. This lets Google assign resources to that project, and without this, you will not be able to run the code. You will see that I am already signed on to a project. It's this ssms-347819. This number will also appear inside this Jupyter Notebook file, but that project name is unique to me.

24:03 Speaker 2: The others will run into permission issues if they try using that. So you should create your own project for doing that. I'm going to walk you through the steps of that right now. For some of you who might already have a project created, and this is not new information for you, you can just change cloud project and switch to an existing project. But I'm going to assume you don't have a project set up and just show you the process from scratch. And so, yeah, if you click on your user avatar here, you'll see the option for register a new cloud project.

24:38 Speaker 2: So it's going to be non-commercial, unpaid usage. The project type, in my case, is academia and research. This will hopefully apply to most of you. And then, again, if you have an existing project that you want to link to, you can do that here. Otherwise, create a Google project. Organization can be left blank. The project ID is going to be a unique identifier that you just saw on the Google Colab page as well as the Google Earth Engine page. So for example, let me call it something like ssms-001.

25:13 Speaker 2: It needs numbers and letters. This is its text field, which is kind of like a description.

25:22 Speaker 2: Call it ssms-test and then continue to summary. This is happening because I had something else created like this. Let's try this once more. So it has to be a unique name. You can't have an existing project with that name. So I'm gonna create a new one. Takes a couple of seconds to assign resources and set it up. So the important piece is this. You should copy this or make a note of this. You can see this from your account later as well, but this is the unique identifier that you will need going forward.

26:03 Speaker 2: So do all of that, click confirm. Again, it's gonna take a few seconds. It's gonna say it's registered. It's redirecting you to the code editor. Now the code editor, again, you saw it automatically redirects you back here. The new project should be opened up already. So you can see it's already assigned me to the new project I just created. So again, make a note of this. I am going to switch back to my existing project because that is what it will draw resources from. This was for demo purposes.

26:40 Speaker 2: So you can always change cloud project and then from the dropdown list, it'll show you the ones that exist. So this is the new one that I just created. This is the one that I've been using so far. I'm gonna select that, okay. So now that it is set up, we are ready to kind of start and start running the LandCover classification code. So I'm gonna close this for now, but we will open up this page again later on once we have started running a job. Okay, now this obviously is gonna be empty for you for the time being.

27:20 Speaker 2: And therefore let's go back to the Colab page. So the way Colab works is you need to give it access or you will need to connect it to your Google Earth Engine account and project. You will also need to connect it to your Google Drive. The reason you connect it to your Google Drive is because some of these codes can generate large files and you don't want to have to manually babysit the whole process. You want this to be automated especially if you're running multiple pieces of code and you want them to be downloaded to your Google Drive once it's done automatically.

28:00 Speaker 2: So this kind of facilitates that. It goes without saying that your Google Drive needs to have some empty space. None of the files we will be working with today are going to be massive. Probably a couple of hundred MBs is sufficient, but just have some space in your Google Drive so that the files can be exported there. A couple of packages will be installed when you once you start. The moment you start running this code, you will notice that Google has assigned you or Colab has assigned you RAM and disk.

28:35 Speaker 2: This is why we prefer running all of this on the cloud because it doesn't strain your system. Google's resources are usually good enough for running this sort of code. And so whenever you're connected, it'll show you that all of this is running. Okay, so every time each of these are called cells, every time a cell finishes running, you will see a tick mark beside it. That's usually the sign. I'm gonna hide the output from here. Now it's gonna import a bunch of packages where one of them is the one we just installed, one of them already exists on Colab, and the third one is for some tabulation stuff.

29:14 Speaker 2: Okay, all of this is imported. The next two steps are the important ones. The first one is authenticate. This is, first of all, connecting Google Colab to your Google Auth Engine account. So allow, then it's gonna bring up this sort of Google login screen, which several of you might have seen before. If your email ID that you're not logged in, that you're logged into Google Earth Engine doesn't show up here, you need to log in and select that. But usually if you're logged into stuff, this will show up there.

29:50 Speaker 2: Click on that, click on continue, click on continue. Okay, check mark, this has worked perfectly. So now your runtime is connected to Google Earth Engine. Next, you'll allow it to have resources, which will connect you to the project you just created on Google Colab, sorry, on Google Earth Engine. So if you run the next line, that happens. Now, I'm gonna zoom in here for a second. As you can see, this is the project that I have on Google Earth Engine, SSMS-347819. So if I go back to the Earth Engine website, that's the project I'm using.

30:41 Speaker 2: You should just replace this with, you know, whatever project you just created. So for example, if your project is called SSMS-003, you just replace it with that. If you don't, if you, for example, say link to something that doesn't exist, or leave it blank.

31:02 Speaker 2: It will usually give you an error. In my case, probably because I'm already signed in, it's not showing that. Okay, I'm gonna switch this back to what I had. We're gonna stick with that. Last bit of setup that is required is allowing Google Drive access to your account, which is what this does. you'll have a very similar setup like you just had with this eEauthenticate function. You say connect to Google Drive. Again, select the account. I think it's not necessary, but it is convenient if the Google Drive account, the Google Colab account, as well as the Google Earth Engine account are all connected to the same Gmail ID.

31:47 Speaker 2: Not necessary. It can be across different IDs, but I think it's just simpler to keep track of. So you select that. give it permissions, continue. Okay, and that has successfully run. Later on when we are at the export stage of this file, I'll show you where this comes into play. Okay, now we're actually getting into running the actual analysis. There is going to be some importing data, some importing functions, initial setup. Part of it is preloaded, part of it the user can modify. Let's work through that.

32:33 Speaker 2: The first bit is a data import. A lot of these cells are going to be minimized by default, but you can go in and look if you want. The couple of things I will bring your attention to are these three files. These are the land cover reference maps. So the ground truth for 2012 and 2019, which is used to train our model. The 3C, you will see a 2C later on as well, stands for three classes, in this case, forest, non-forest, and pasture. This is our region of interest by a polygon within the country of Panama.

33:07 Speaker 2: And there are a bunch of other files that are imported, as you can see from the path they're imported from. They're all imported from a Google Earth Engine account, in this case, mine. But they're all openly available, and those are the ones that are used to train the model. I'm going to minimize this in a second. and I will run this. Okay, all data has been imported, so you can see the tick mark here. The next piece that needs to be done is re-projecting some of the maps. So the Google Earth Engine, which hosts the Landsat 8 dataset and the Sentinel dataset, they are all in this 4326 projection.

33:52 Speaker 2: The land cover maps that we have in the beginning are not. So I'm just gonna reproject them. So they're all in the same protection and there are no misalignment issues. There are a bunch of functions. Most of them are standard. One of them is supplied by Landsat 8. I'm not gonna go into details, but when we use them, I'm going to tell you what each of them does. So these functions need to be defined. Okay, now comes the interesting bit, which sets up the parameters of the models. Now, the two things that you will need to take into consideration when you run all of this is what year you want the land cover map for, and what season you want the land cover map for.

34:35 Speaker 2: The years, I think the Landsat goes back earliest to roughly 2014-ish, and it should work easily until 2021. You'll see in a bit that it'll give a warning about a deprecated data set. I'm gonna discuss that. But up till 2021, All of this that is set up should be fine. As for the seasons, the seasons have been defined specifically for Panama, and I will not recommend changing these dates. There is an option to change them later on, but you should stick with these because some of the underlying setup has been defined based on these dates.

35:13 Speaker 2: So annual is just the entire year. Dry season is January to March. Wet season is May to December. Okay, so... for example, if we want this to be 2015, and W is wet, and then so it's gonna be D, W, or A for dry, wet, or annual. In this case, it's gonna be the wet season, so I'm gonna say W. No need to change any of these. These are fairly standard and have been tested, so just stick with all of that, run this. Okay, so now we're gonna do classification 2015. This cell, Again, no need to modify anything here, but basically it sets the export file name.

35:56 Speaker 2: And the reason I'm showing you here is I want to go through what the file names look like and kind of what they represent. So the first two letters represent the data set. So lands at eight. The second two letters represent the what is classified. You've seen probably in the PowerPoint that there are two types of methods we'll be using. We'll either be classifying a median composite or we will be classifying the entire image collection, which is this IC. So it's kind of defining that, the year, followed by the analysis scale is the resolution.

36:35 Speaker 2: It is set to 20 meters by default, which is what a lot of the bands and information are. I'll not recommend changing this because at least making it less than 20 meters because it requires a lot more processing power, a lot more time. And if you try to do it for large spots of the country, you will run out of resources on Google Earth Engine. And finally, it adds the season at the end. So, you know, W or A or otherwise, I think it's just gonna make it D. Okay, so this just generates the file name beforehand, so it's automated later on.

37:08 Speaker 2: imports the Landsat data. So you can see it's importing the image collection for Landsat. It's filtering it by the start date and end date. That depends on what season you chose. It's chopping up by the shape of the entire country because that's what the model is trained on. This is a cloud mask that is applied to each image. This is a water mask that is applied to each image. This is a band for the elevation info. that is applied to each image. And finally, an NDVI calculation that is done based on the Landsat

37:46 Speaker 2: bands that is also applied to each image. This also loads up the training data sets that we will be using, which is just a composite of the Landsat data for 2019, which is our reference here for the three seasons. And when you run this, you will get a warning that this dataset is a deprecated asset. So once Landsat 9 came out, they deprecated a bunch of these things. If you go to this link when you run it, you can see the discussion on that. But the current dataset that we are using is still completely okay, anything up until 2021.

38:26 Speaker 2: Beyond that, they have a discussion on this webpage which will tell you to just change the name of the data set. So all you need to do is basically change this data set to whatever they recommend here, depending on the year. But for the purposes of this training, we are good. I will not rerun this cell, but this essentially just tells you how many images are in that image collection. So for like, I don't think this is 126, it was for 2015. It's probably for 2020. The reason I'm not running it, it takes a bit of time.

38:59 Speaker 2: It's gonna be a very similar number when you run this. And it's kind of like a sanity check to make sure the data has been imported properly. Some more setup before we run the random forest. Gonna prepare the data for classification, which will create a median composite. So takes the collection, calculates its median. It creates the three classes that we're gonna be using, forest, non-forest, and pasture. It selects the bands of interest. These are the optical bands. Elevation and slope come from the DEM function, which is the SRTM dataset.

39:35 Speaker 2: And NDVI is calculated based on the Landsat bands itself. And finally, the data import up top, if you look at it later on, you'll see it's got like some 70 or 80 training polygons across the three different classes. This just imports those. I'm gonna minimize and run that. Okay, this says do not run in all caps because this is essentially going to try and download this entire composite with all of these bands. The reason it says do not run is I think the resulting file is a few gigabytes in size.

40:13 Speaker 2: You can do it if you want to, but it's not gonna serve any important purpose. Okay, now we are actually getting to the meat and potatoes, which is the actual classification. This is the random forest classification on the median composite that we just created in the previous cell. So we first define a classifier, which is a random forest. The random forest has a bunch of parameters. All of these parameters were set up top in the same cell where we set the year and the season. So no need to change any of this.

40:49 Speaker 2: You tell the random forest that it's gonna train on the training features, which is the forest, non-forest and pasture that we had. And it will train using the bands that we just defined previously. Once the classifier is defined, you're gonna take the annual composite or the seasonal composite in this case, select the bands of interest, classify using the classifier, and we rename the layer for ease of use later on. Okay, one thing you'll notice, like this is supposed to have taken longer, but it didn't, it just said it's done immediately.

41:28 Speaker 2: That's because a lot of these software have, I think the technical term for that is a lazy implementation, which kind of queues up these things, but don't actually run the classification until you try to export the output. That's why these cells are running fast, but the moment I try to export this, you will see that the actual underlying work has begun. Okay, before I try to export this, a couple of housekeeping things. We eventually care about forest and non-forest, but as you saw up top, I mentioned there was forest, non-forest, and pasture.

42:08 Speaker 2: We classified 2C, which is two classes, takes the classified, which had three classes, and runs it through this function aptly named pasture to non-forest. So it essentially collapses the pasture class back into non-forest. So you're left with two classes. So it's gonna do that. And this is the export statement. Couple of things to note here. It's gonna export it for that ROI that we had. Second, my mouse. Yeah, mouse is being finicky. ROI, which was at polygon. This is the file name. As I showed up top, the file names are automatically generated.

42:48 Speaker 2: And as you will see, that's why you don't have to manually input them here. This is the folder name. This is the folder that I asked all of you to create right in the beginning of the training. Now, if you want to mess around with it, you can change the folder name here, in which case you will need to change the folder name here. But the Colab will look for this folder name inside your Google Drive because it's given it access to your Google Drive when you try to export this. And just a couple of other housekeeping things is I think this piece.

43:26 Speaker 2: It's gonna have this unmask statement in there. And when we open up the output file, you'll see what it does. a lot of, because you saw in the beginning, we applied a water mask, we applied a mask for the clouds as well. We kind of want to unmask them later on so you have access to the entire data. And so this is required before the export is run. So I'm going to run this export. And so this one shows the status that, you know, it's starting, it's running. And you can actually track this in more detail on the Google Earth Engine page.

44:08 Speaker 2: Before I go there, you can always also track this from here. So whenever you run this, it'll show that it's been running for these many seconds. Obviously, this is machine time, so it's not really readable. But like how many attempts, priority, the name of the export file, and the state, whether it's running or completed. How do you actively track this? So, yeah, once you run the export, if you go back to the CodeEarthEngineGoogle.com page, it might take a few seconds to populate, but you will see under Tasks, the current tasks that are running.

44:45 Speaker 2: I accidentally have a duplicate running. I'm going to cancel one of those. But some multiple tasks will run, and you can track them in more detail if you go to the Task Manager. The task manager will actually show a bunch of your previous tasks as well. So like tasks I've run in the past. I'm going to cancel these two for the time being because they take, I would say in the ballpark of 20 minutes to run if you look at all of these previous tasks. So you can also individually cancel them or you can do bulk cancel mode, which will cancel everything that's running.

45:18 Speaker 2: So if you accidentally start a task more than once. Okay, so these have now been canceled. But that's fine. Otherwise, it's going to run for 20 minutes or so, and then it's going to show you this tick mark that it has finished running successfully. Before I show you the output, I'm going to also show you how the classification works on the entire image collection. It's very similar. It's two additional lines of code. But as you saw, these were the classifications on the median composite for 2015.

45:48 Speaker 2: If I go back to the Google call app page, The next section is classifying on the image collection and interpreting it.

45:58 Speaker 2: It's all wrappers of functions that you just saw previously. So all the prep work is already done. Running it now is just two extra lines, where it takes this entire collection of images, that hundred and something images, it runs it through, sorry, this is the actual collection of images, the Lancet collection, It runs it through this function called RF collection, which does the whole series of things, which is selecting the right bands, selecting the forest, non-forest classes, running the classifier.

46:31 Speaker 2: And then also does this last step, which is taking the three classes version into the two classes version, which is just forest and non-forest. So all of that is kind of done automatically through this. If you really care, this RF collection function is in that which had all the functions, if you really want to parse through it. So you want to run that. And again, as you saw, this run automatically, but nothing has started yet. Let me refresh this. It's gonna sometimes takes a few seconds to populate.

47:08 Speaker 2: So nothing has started yet, even though I have run it. Because as I mentioned, none of the code starts running until you ask to either display the results or export the results. There's some more housekeeping that happens here because you saw in the PowerPoint that it talks about how to interpret if you have all the images that are classified. It kind of combines all that information together. And the export statement looks basically the same. The classification file name, again, is automatically set and it takes the file and it just exports that.

47:44 Speaker 2: This will, by the way, take longer to export than the median composite. That's I think the main difference because the median composite is only exporting one file. This has to do a lot more pre-processing in terms of classifying every single image before it exports it. So this takes longer. And then when you run that, we give this a few seconds, there we go. This is already populated here. If you again go to the task manager, which is this one, you will see that this is running. You can run multiple tasks at the same time, but kind of Google prioritizes, like there is a priority thing.

48:20 Speaker 2: If you are running multiple, the first couple will be priority 100, then it'll be lower priority. Since the resources it allots you are finite, if you run too many, small chance the later ones, you know, it shows it's running, but it'll probably be pending until the earlier ones are completed. So I will not recommend running more than maybe two to four tasks at the same time. Okay, I will cancel this one as well because it's gonna take a while. But luckily for you guys, I've already kind of generated all of these data beforehand and I can show you what each of them look like and how you can interpret it.

48:57 Speaker 2: So once they finish running, I'm gonna pull up an example of something for me that has finished running. You can go here and then you can click on Open in Drive. which will basically open up the same folder you saw at the beginning of this demo, the GE exports folder where everything gets saved, right? It's the same identical folder. Now, whichever file then you want to see the details of, you will need to download them, which is very standard for Google Drive. You just click on download. You see the file sizes are not very large.

49:31 Speaker 2: The image collection ones are slightly larger. The median ones are slightly smaller, but each of them is going to be fairly small. Download them locally. You can also directly open them from your Google Drive. That is also fine if your Google Drive is synced to your local hard drive. But otherwise, just download them locally, and then you can open them up in any compatible software which can read these diff files. The one that I usually use is the software called QGIS, which makes looking at all of this data fairly straightforward.

50:07 Speaker 2: So that's what I will... show you for the last bit. So yeah, once you have a software like this opened up, these are the files that have been generated. Let me show them slightly large. Can we make this even larger? Yes, we can. As you can see, the names are fairly self-explanatory. They're all going to be Landsat 8. They're all going to be classifications. Some are image collection. Some are medians. The years are going to be 2015 or 2020 or anything in between if you change that. And then you will see the resolution, which is always going to be 20 meters and the season.

50:45 Speaker 2: In particular, let me show you this file and I will also show you the corresponding file from 2020. Right, we're going to take one of these files and let me make sure I'm grabbing the right one. Yeah, classification median. And we're going to drag it onto QGIS. which will display it, but as you can see, the band information is weird. We need to make some more changes to the properties. So I'm gonna double click this, which takes me to the layer properties of this file. And then you see the minimum value is minus 9999, which I pointed out was what we unmask missing values to.

51:25 Speaker 2: and the maximum is two. So there are actually four bands in here that we will now assign to different colors. So the render type is going to be single band pseudo color. The interpolation is going to be exact because there is no interpolation. There are exactly four values in here. You want to set this to be quantile because that's the easiest to set up and there are gonna be four classes. Okay, so this stays as is, this stays as is. I'm going to rename these two. The value is the actual value of the class.

51:58 Speaker 2: The label you can set to whatever, but zero. I'm going to set this to one. For completeness, let me also change the labels. It's missing. Zero is my non-forest. One is my forest. And two is actually water. I'll also obviously change the colors to match. So the missing data I'm gonna put in gray, the water I am gonna put in blue,

52:27 Speaker 2: the non-forest I'm gonna keep in yellow, and the forest I'm gonna turn green. Okay, all of that is done. Now, if I apply that, you have a usable map. So all of this is water. Whatever is gray is almost all missing data. There are a few exceptions. shapefile that we have for Panama has the lakes. And for example, if I zoom in, like the Panama Canal area is also as kind of missing, which is fine. That's correct for our intents and purposes. But yeah, it's either cloudy data or it's like an internal water body, an inland water body.

53:09 Speaker 2: But otherwise, everything that you see here, green is forest, yellow is non-forest. Let's do that really quick for the corresponding 2022 data. So 2022 median wet season drag and drop. I'm going to do this really quickly. So pseudo color exact on tile for classes. Let me make this zero, one and two. I'm not gonna bother changing the names again. This is zero. and then two is water, one is forest, and apply. So if you wanna see the differences between the two, there are obviously more involved tools in the software, but visually you can kind of see how 2020 has a lot more forest classification than 2015, even if you take into account all the missing data.

54:08 Speaker 2: And yeah, that's basically the entire process. You can always track what you have exported here. You can always track job status from the Earth Engine file. Okay, so the one last thing I wanted to point out that this is mostly just methodology for generating land cover. And in practice, you would want to validate a lot of these results. Now, as you saw in the beginning of the Colab file, that all of this was done, the baseline year for us is 2019. So it's very easy for us, well, relatively easy for us to validate this for 2019 because we have this file here and you can also write your own export statement, export this LC 2019 file and do analysis for 2019 and validate.

55:06 Speaker 2: But if you're doing it for a different region or a different time period, validation would be done by getting access to the ground truth and then comparing against the ground truth and seeing are there any kind of biases? Are those biases particular to a season or do they appear in annual data as well? All of that is kind of critical. This kind of just, the demo today just demonstrates how you would do the random forest classification if you had ground truth for a particular year as well as access to Landsat or Sentinel data and all of that.

55:43 Speaker 2: And also, as I mentioned, the dataset, this particular dataset has only been tested up till like roughly 2021, which is beyond that, as you saw the thing, the dataset is deprecated, so it might not perform as well for years beyond that, but anywhere between say 2015 and 2021, you should be good. Then that was most of the steps of the demo. And I will now pass it on to Erika, who will give her closing comments.

56:17 Speaker 4: Thank you very much. That was a great demo. Just to remind participants that the demo is focused on the methodology to generate a land cover classification. You can play around with this demo and modify it to generate a land cover class over your area of interest. You would need to change the study area, the period of time of interest, and also you need to update the labels. So that would be the training and validation labels. And as long as you have the right data to validate for the same year that you're generating your results, you can validate that.

57:02 Speaker 4: And at the end of this code, there is a step to validate your results. And one thing to mention too is that you can increase the number of classes. Here we had three and then two classes. But if you have multiple classes that you can train on, you can certainly increase that capability in the code as well. So with that, I will now hand it over to Amita.

57:36 Speaker 6: Thank you very much, Amber, Erica, and Ritam for the excellent presentation and demonstration. With that, we conclude this four-part training on droughts. And just to summarize, this four-part training focused on drought monitoring, prediction, and protection based on Earth system observations and models. In part one, we talked about Droughts occurring on different timescales, ranging from weekly to decades. We talked about drought classifications and various types of droughts, including meteorological droughts, which are indicated by precipitation deficit.

58:20 Speaker 6: Agricultural droughts, indicated by soil moisture deficit, low vegetation, and low crop yields. And hydrological droughts, indicated by low stream flow and reduced groundwater.

58:35 Speaker 6: NASA Earth observations provide near real-time to more than 10 years of data, and in some cases, more than 20 years of data for drought monitoring, including precipitation, surface temperatures, vegetation cover, soil moisture, and groundwater. In part two, we talked about NASA's sub-seasonal to seasonal prediction system that provides an ensemble of forecasts of temperature, precipitation, and soil moisture on monthly to seasonal timescales useful for water resources and drought management. Then, in Part 3, we talked about NASA Earth Exchange Global Daily Downscaled Projections for Coupled Model Intercomparison Project Phase 6, or NEXT-GDDP-CMIP-6, climate projections for the 21st century, which are available from multiple global climate models downscaled at daily quarter-degree latitude-longitude resolution.

59:41 Speaker 6: We reviewed satellites and sensors for Earth observations useful for deriving drought indicators for different types of droughts that include precipitation temperature, normalized difference vegetation index, evapotranspiration, soil moisture, and groundwater. And the satellites and sensors, they are listed here in this table. In part one, we introduced drought indices, including standardized precipitation index, SPI, Palmer Drought Severity Index, BDSI, normalized difference vegetation index, NDVI, and vegetation condition index, VCI.

1:00:21 Speaker 6: We also saw NDVI and solar-induced fluorescence applications for monitoring agricultural drought. In part two, we had an overview of NASA global modeling and assimilation of its S2S prediction system. And in part three, we had an overview of next GDDP CMIP6 climate projections. Today, Amber presented NASA WAO drought projects, including Western Land Data Assimilation System, LDAS, and grace-based drought indicators. She also talked about Navajo Nation Drought Severity Evaluation Tool, or DSET. And then we heard from Erica and Ritam about Sustainable Forest Management Information System, SFMIS tool.

1:01:11 Speaker 6: We had several demonstrations. of Drought Monitoring Tools and we also conducted calculations and analysis for drought assessment. In Part 1, we had demonstrations of Drought.gov and US Drought Monitor and today we had demonstrations of DSET and SFMIS tools. In Part 1, we calculated SPI and VCI using Google Earth Engine. In Part 2, as to as predictions of temperature and precipitation anomalies were analyzed using QGIS, and in Part 3, climate change projections of temperature and precipitation were examined till 2100 from NextGDDP using Google Earth Engine.

1:01:59 Speaker 6: There is one homework for the training, and it is posted today on the training webpage. The answers for the homework must be submitted via Google Forms, and the due date is 15th of August. Certificate of completion will be awarded to those who attended all four live webinars and who completes the homework assignment by the deadline. You will receive the certificate via email approximately two months after completion of the training. Here's the contact information for today's speakers. This is RSAT website link and RSAT social media links.

1:02:41 Speaker 6: Here are our sister programs for capacity buildings, development survey, and links are given here. And we want to thank all our speakers today, from RSAT and our guest speakers, from NASA, Compton Tucker and Andrea Mollard, from NOAA, Kelsey Sattolino, Brad Pugh, and Steve Ansari, Amber McCullum for NASA Ames, and Hritham Majumdar from NC State University. A special thanks to our set coordinators and editors, Natasha Johnson-Griffin, Sarah Karchel, Susan Monti, Brock Levins, Sabin Hudson-Odoi, and Jonathan O'Brien.

1:03:21 Speaker 6: And thank you all very much for attending today's session. This concludes our webinar series. In four parts, we have covered a lot of information about many aspects of droughts. We'll go to the question and answer session. So we'll start with question one. Is the GRACE data open or free access? And Amber, you have answered the question, so you want to unmute and address that.

1:04:07 Speaker 5: Yeah, fantastic. Thank you, Amita, and thank you, everyone, for joining our session.

1:04:13 Speaker 7: The short answer is yes. All NASA data are freely available. There are a few different places where you can access the GRACE data and visualize the GRACE.

1:04:24 Speaker 5: Data, and I've linked a couple of those here in my answer.

1:04:29 Speaker 7: There's this really nice great follow-on data analysis tool that allows you to visualize the data as well. And then in the project that we highlighted with the Western LDAS project, you can access the maps and analysis that we presented as part of that project at this link here that I provided as well.

1:04:52 Speaker 5: So those are for the U.S.

1:04:55 Speaker 7: And also many of that is available globally as well.

1:05:01 Speaker 6: Thank you. Second question I'll say for you, is the drug too limited to the USA only?

1:05:09 Speaker 4: No.

1:05:10 Speaker 7: So I presented essentially our spinoff of Climate Engine.

1:05:17 Speaker 5: So for those of.

1:05:18 Speaker 7: You interested in doing these types of analyses, and using the data on a global basis, I would go for Climate Engine. So the drought severity evaluation tool was really tailored to.

1:05:34 Speaker 5: The specific needs of the Navajo Nation.

1:05:39 Speaker 7: And so we had very specific things like their administrative boundaries, their rain gauge data, and the ability to calculate some area average statistics over those specific regions. However, because this was built as a spin-off of Climate Engine.

1:05:59 Speaker 5: Many of the same analyses, data, are available via Climate Engine.

1:06:05 Speaker 7: So you might not be able to do a very specific thing like the area averaged six-month SPI over this specific administrative region.

1:06:15 Speaker 5: But you can do.

1:06:17 Speaker 7: nearly all and many of the same analyses that we did with DSET at a global scale. And in fact, Climate Engine has actually more data available on it than our DSET tool. And we did that purposefully as well, where we sort of scaled back all of the data sets available. So Climate Engine now also has some forecasted data and a lot of really interesting model data to use. So I would go ahead and access the tool via the Climate Engine website that I mentioned here. And the only thing that you need is essentially a Google account.

1:06:55 Speaker 7: And Climate Engine is built off of the Google Earth Engine.

1:07:02 Speaker 5: Skeletons, basically.

1:07:04 Speaker 7: So you heard all of these really fascinating things you can do with the Earth Engine code.

1:07:09 Speaker 5: And that is essentially building.

1:07:12 Speaker 7: Off of the power of Earth Engine and pulling in data available globally.

1:07:18 Speaker 4: Thank you.

1:07:19 Speaker 6: Great. Question three is, are SPI values standardized across the globe or normalized locally in some way?

1:07:27 Speaker 5: Yeah, that's a great question.

1:07:28 Speaker 7: So the SPI essentially quantifies observed precipitation as a standardized departure from a probability distribution function that models the raw precipitation data.

1:07:42 Speaker 8: So what this is doing is it essentially is looking at the region of interest, in this example, sort of the pixel of interest, and looking back at, say, the climatological average of precipitation in that specific place over time and comparing whatever timeframe you're interested in looking at to that average.

1:08:07 Speaker 7: And so if, for example, we were looking at the six-month SPI starting today, the model would look at the previous six months of this year and compare those six months of this year to the long-term average of precipitation.

1:08:26 Speaker 8: In this same place.

1:08:28 Speaker 7: And so that's actually a real benefit to the use of the SPI because it's specific to the region of interest, but it is also an easily transferable index across many different climatological regimes. You can look at SPI on a variety of different time scales too. So you can look at this month compared to the climatological average of every other August in history in this region. Or you can look at it at a larger time window. There are, however, limitations to the SPI, because the only thing it's including is precipitation.

1:09:05 Speaker 7: And we all know, and from this fantastic training, there are a lot of factors at play when thinking about drought. So just wanted to make note that other indices include things like.

1:09:15 Speaker 5: Evapotranspiration, which is an important component of drought as well.

1:09:19 Speaker 7: And I've linked a nice website that describes the SPI in a little more detail there as well.

1:09:26 Speaker 6: Thank you, Amber. I also want to remind everyone that we went through a description of SPI in part one, and our prerequisite job webinar also has great detail about SPI. We used Sean McCartney demonstrated the Google Art Engine code, how to calculate SPI, and that's one of your homework questions based on that. exercise. So hopefully you learn more about SPI when you work with it. Thank you. Question four, is the DESET tool available for any country in the world, any other country?

1:10:03 Speaker 7: Yeah, so that's the same answer to question number two. DESET was tailored specifically to the Navajo Nation. However, all of the data available, with the exception of the Navajo Rain Gauge data, And even additional data are available globally via Climate Engine.

1:10:23 Speaker 5: So if you're interested in that, take a look at the Climate Engine website. Thank you.

1:10:29 Speaker 6: Thanks. Question five, does GEE exports have to be loaded in Google Colab? I have done the GEE authentication, connected the project CFMIS 347819 and authenticated Google Drive. But for the data import, I have an error. EEE exception, Earth Engine client library not initialized. Run EEE initialize. So is there a step that I missed?

1:11:00 Speaker 2: I can take that.

1:11:01 Speaker 1: Yeah.

1:11:02 Speaker 2: So yeah, I mean, I'm assuming this problem eventually got resolved because I did see follow-up questions from the same poster. But yeah, it's easy to miss that EEE.initialize step because it's a single line. And that's the easiest way to check whether you're doing things in sequence, because if you don't run this, everything afterwards will basically throw an error.

1:11:22 Speaker 7: Okay.

1:11:24 Speaker 6: Question six, is it possible to import the GEE map and view the results directly in Google Colab?

1:11:31 Speaker 2: Yeah, so thanks for pointing that out. There are previous versions of this where we did have a GEE map in there to kind of display results within the... interface, the Jupyter Notebook itself, but I've recently taken that out because depending on the size of the raster and the number of layers you want to display on it, it can be a little finicky depending on what browser you're using. But it's a fantastic package and it's got a lot of resources. So for those of you who are not familiar, I'd recommend looking up the package.

1:12:07 Speaker 2: If I remember right, it's just geemaps. single word, all small letters, which will allow you to display a lot of these rosters directly inside the Jupyter Notebook.

1:12:21 Speaker 1: Thank you.

1:12:21 Speaker 6: And question seven is for you also. Can you briefly discuss connection to Colab?

1:12:26 Speaker 2: Yeah. If anybody has other specific issues, feel free to put it in a question so I can resolve them in any individual issues you might have. But generally, it's fairly straightforward. It's similar to, you know, logging into any Google service. You just go to the webpage that's listed there. If you're already signed into your Google account, it should show up on the top right that you're signed in, like you have with Google Docs or Google Drive. If you're not, you'll see a sign-in button, and you sign in with that, and you should immediately be logged into Google Colab, and you'll see just like the pop-up page I had in the beginning, which was asking me to load up the notebooks that should immediately be there next time you're at the webpage.

1:13:10 Speaker 6: Great, question eight. When I select tasks in GEE and run LHCIS ME 2019-20A3C, it shows image collection load. This T1SR is deprecated. See the website. for how to migrate to collection tool and set.

1:13:38 Speaker 2: Yeah, again, thank you for bringing this up. So between the time the demo was recorded and essentially last week, the data set that was just deprecated before and through a warning, like you saw in the video, has since been deleted from Google Earth Engine altogether and it's forced the migration to the new version of the data set, which is CO2 slash T1 underscore L2. And so the Jupyter Notebook that you will end up downloading from the RSET webpage is actually updated. So it isn't what was in the demo.

1:14:13 Speaker 2: It's updated to the new data set now. So A, it shouldn't show that deprecated warning. It should run without any issues. But you will also notice a couple of small changes that I should point out. The largest one being that, if you remember, there was a section which showed There was three pre-trained training rosters that were already there for the two seasons and the annual one. Those lines you will see are now commented out with a note that it's because the data set is deprecated and the current version of the code directly generates and trains on the training data while you're running the notebook.

1:14:58 Speaker 6: Question nine, how many samples are needed for each class? maybe use Sentinel images. Are there options we could include in our data?

1:15:07 Speaker 2: Yeah, this is a great question, and this is actually two different questions, so let me address them one by one. Yes, you can include Sentinel-2. If you look up Google Earth Engine dataset, you will see Sentinel-2 is available. I will recommend using the surface reflectance one, which is, I think, underscore SR, which generally has provided better classification performance in our tests. And it's also at a slightly higher resolution. But the downside of Sentinel-2, at least the surface reflectance dataset, it doesn't go back as far as Landsat 8 does.

1:15:46 Speaker 2: And so because we wanted to show droughts and we wanted to go back as far as 2015, using that was not an option in this demo and that's why we resorted to Landsat 8. As far as the number of samples are in each class are concerned. There are a couple of ways to go about this. The short answer to your question is that it's probably much fewer than we are using. The way we selected the samples was by manually drawing polygons in the Google Earth Engine's JavaScript interface at code.earthengine platform.

1:16:21 Speaker 2: But you can also, there are ways to randomly sample locations. And so in our case, because we were drawing polygons, we have, I think, something on the order of 10 raised to six training points across the three classes. But if I were to hazard a guess, I think you can probably get away with much less, something around, you know, 10,000, so 10 raised to five, that order might actually be sufficient spread out across the three classes.

1:16:48 Speaker 6: Okay, question 10 is, On the Navajo Nation case study precipitation graph superimposed on SPI graph, how can you explain the lag on the SPI with respect to the precipitation? This lag was mentioned during the lecture.

1:17:10 Speaker 5: Yeah, thanks for that question.

1:17:12 Speaker 8: So what was shown here was the six-month SPI.

1:17:16 Speaker 7: I'm sure, Amita, you all covered differences in time.

1:17:21 Speaker 5: Scale when assessing droughts and what that means for the SPI calculation.

1:17:27 Speaker 7: Because the calculation is including essentially the last six months of data to create this one value at that new data point of the standardized precipitation index.

1:17:40 Speaker 5: There will in turn be a lag.

1:17:42 Speaker 7: In these high precipitation events until you're seeing the SPI value change as a whole when looking at it over that sort of mid-range time scale, like.

1:17:53 Speaker 5: The six months.

1:17:55 Speaker 8: If you were, for example, to take a look at the one month SPI, you might see that SPI.

1:18:00 Speaker 7: Responding a little more quickly following one of these large sort of rainfall events. That also sort of highlights the benefit and maybe, you know, pitfalls in some way of using the SPI. So you can look at the different time scales of this index to evaluate changes to.

1:18:23 Speaker 8: Different parts of the water cycle and.

1:18:26 Speaker 5: The water system.

1:18:27 Speaker 9: So, you know, the soil moisture and the precipitation is going to change on a shorter scale, but things like groundwater, stream flow, reservoir storage, those are going to be changing on a longer-term scale and may necessitate that longer-term analysis.

1:18:42 Speaker 7: And Amita and team, feel free to add to that response as you are the experts there as well.

1:18:48 Speaker 2: Thanks.

1:18:49 Speaker 1: No problem.

1:18:50 Speaker 6: You've got your answer right on. So, yeah, because we're just looking at precipitation here with respect to long-term mean. And so that lag, depending on what time window you choose, will be there because of these factors that you mentioned.

1:19:10 Speaker 6: Question 11. If in my study area I have specific and historical precipitation data from meteorological stations, how do you recommend integrating them into a drought analysis? So Amber, you can respond to this, but you can calculate based on meteorological station SPI can be calculated based on precipitation data. And that would, of course, be limited to to station data, if you can interpret, if you have high density meteorological station, then you can maybe make a special interpolation, make a map from that of precipitation.

1:19:54 Speaker 6: You perhaps can have a map of SPI, but otherwise at each station, you can calculate SPI. Just like, yeah, go ahead.

1:20:08 Speaker 8: Yeah, I was just going to say that.

1:20:09 Speaker 2: That's spot on.

1:20:10 Speaker 7: I mean, you could calculate it at a specific station. You could use a network of stations, like a mesonet network or a variety of.

1:20:19 Speaker 5: Simultaneously gauged stations to

1:20:23 Speaker 7: Create this interpolated map, exactly what you're saying.