Flash Drought Detection and Prediction

0:01 Speaker 1: Hello and welcome to the second and last session of this intermediate two-part RSAT training titled Agricultural Flash Drought Detection Using SIF and Soil Moisture Data.

0:12 Speaker 2: I'm Dr.

0:12 Speaker 1: Erika Potest. I'm a scientist at NASA's Jet Propulsion Laboratory and also an instructor with the RSAT program. Today's session will focus on flash drought detection and prediction. And we have as guest instructor, once again, Jackie Ryan from NASA's Jet Propulsion Laboratory. A reminder that there is a homework associated with this training. It opens today. You can access it through the training webpage, and the due date is October 20th. The certificate of completion will be awarded to participants who attend both live sessions and complete the homework assignment by the due date.

0:55 Speaker 1: Prerequisites. These are the three prerequisites for this webinar series. I'm the ARSA training coordinator for this webinar series. And today we have one invited guest instructor. She's, once again, Jackie Ryan. She's a data visualization developer at NASA's Jet Propulsion Laboratory, and she will provide a theoretical background and demonstration on flash drought detection and prediction. And these are the training objectives for today's session. By the end of the session, participants will be able to synthesize SIF data rapid change index time series data for a selected region using a provided Jupiter notebook.

1:40 Speaker 1: Retrieve and manipulate SMAP soil moisture and land process model-based soil water deficit index time series data and compare that with SIF RCI time series using a provided Jupiter notebook. And finally, compare the implications of these different methods for flagging and identifying flash droughts in real-world scenarios.

2:05 Speaker 3: How to ask questions.

2:06 Speaker 1: Please write your questions in the questions box, which is at the bottom right. There will be three points. Click on that, and there will be a pop-up menu. Select questions. And write your questions in that box, and we will address your questions at the end of this session. Feel free to enter your questions as we go along, and we'll try to get to all of your questions During the Q & A session, the remaining questions that we cannot answer, we will answer them in the Q & A document, which we will post to the training website in about a week.

2:45 Speaker 1: Let's get started with today's session, and I will hand it off to Jackie Ryan. Welcome back, Jackie, and great to have you supporting this RSAT training. Thank you, Erica, and welcome to the part two presentation on our RSIT course on agricultural flash drought detection using SIF and soil moisture. This part will cover flash drought detection and prediction. And I'll be your RSIT instructor, Jacqueline Ryan. Okay, so for part two, here are.

3:22 Speaker 3: Our objectives for today. And first, we're going to review why SIP and soil moisture are used together for detecting flash drought. And we've already covered the derivation of our drought indicators in the first part of the course. So what I really want to go over is the justification for why this technique makes sense compared to other drought detection methods that are more mature. And in the course, we'll be reviewing the creation of SIF-RCI and SWDI time series, but for new regions and time periods, so that you have a better understanding of how to quickly generate these for new regions.

4:07 Speaker 3: In particular, we'll be looking at a case study of the South African flash drought of 2023 to 2024. which you can see over on the right, a map of the root zone soil moisture deficit during the peak of this event in March, 2024. We'll be comparing different rules for detecting and identifying flash droughts, and then we'll be exploring other data sources for flash drought early warning. Okay, so reviewing what we covered in part one, In Nick's presentation, let's talk a little bit about agricultural drought and its definition.

4:47 Speaker 3: We define agricultural drought as a period of declining soil moisture with impacts to crop yield, regardless of the surface water resources. It doesn't mean that those don't matter. It just means that we care primarily about soil moisture for the definition. And this makes it distinct from meteorological drought, which is just a lack of precipitation over time. And so why do we find this distinction important? For one thing, severe precipitation deficit can often be absent during flash drought events.

5:20 Speaker 3: It might only be a mild precipitation deficit. For another thing, when we're interested in agriculture specifically, crops often don't behave like wildland ecosystems, such as forests. They're typically not native plants, and so they might have different water requirements from the environments that they grow in. An example of this is that a thirsty crop will fail during a relatively mild precipitation deficit or may fail due to high evaporative demand. If you're having trouble picturing what that means, if you look on the right at this image, This is a picture of the alfalfa fields in the Imperial Valley of California, which is a region that is extremely arid, and it's surrounded by shifting sand dune deserts even.

6:15 Speaker 3: Despite this, they're able to grow alfalfa, which is incredibly thirsty, due to the irrigation. And so if we're interested in what qualifies as a drought in that region, we would be more interested in soil moisture than precipitation. And when we talk about soil moisture, it's important to ask why do we need specifically root zone soil moisture? And this is because as opposed to surface soil moisture, which only describes the top layer of soil, we're better able to categorize the full supply of water that plants have access to.

6:54 Speaker 3: In the diagram that we're showing here, the way that roots on soil moisture is defined in models is the soil layer up to one meter deep, but it might not be fully one meter deep. So the depth of the root zone layer is defined by a parameter d in the land surface model that's used in this map data set that we're using in this course. Below that we have a layer of compacted soil or bedrock, and below that we have the water table. And so when we consider the water inflows and outflows from the root zone layer, one of the outflows is via evapotranspiration, which is from the vegetation into the atmosphere.

7:45 Speaker 3: And then for inflows, we have precipitation and infiltration. And infiltration is where water enters into the deeper layers of the soil. But in terms of another outflow, the water can infiltrate into deeper layers of the soil and percolate into the water table. And so a useful analogy I like to use when thinking about roots on soil moisture is it's almost like the water bank account over the cropland that we're looking at. And you can think of precipitation as our deposits into this bank account.

8:29 Speaker 3: And evapotranspiration, as well as infiltration, describe withdrawals. And the net effect of these is the root zone soil moisture that we're able to estimate, which is just describing the supply of water available to the crops. And when we talk about evapotranspiration, that's going to be correlated with temperature and vapor pressure deficit, which are commonly used meteorological variables when studying drought. The idea of why this is a more holistic metric, therefore, is that we're accounting for more traditional drought indices in the roots and soil moisture, but the net effect of them.

9:16 Speaker 3: If you look at the plot on the right, you can see an example of an SWDI time series from the original Martinez-Fernandez paper that developed the index. And one of the things that we can see up at the top, the bars show where there was precipitation events. And wherever we see precipitation, you can see a sudden uptick in SWDI. And that's relatively speaking, proportional to the size of the precipitation event. Now, this isn't a correlation plot, so we don't have an R squared. to tell us how tightly these two variables are correlated to each other, but the paper shows that they have a high degree of correlation.

10:10 Speaker 3: On the other hand, where we see decreasing SWDI, that can be associated with evaporative demand in the environment. And one of the last observations we can make from this plot is that near negative 10 SWDI, which if you recall the definition is based on the wilting point of the soil, the rate of decrease slows. And that's because the roots and soil moisture is approaching a equilibrium dry point with deeper layers of the soil. Or at least we can theorize that's the reason. And in comparison, the surface soil moisture is sort of like a recent transaction history.

11:01 Speaker 3: So it can be observed directly unlike root zone soil moisture, but it doesn't tell us about the total water content available to the vegetation. And so when we talk about what we did in the notebook, we want to compare why we use the soil water deficit index in place of RZSM. If you recall the formula that we introduced in the part one notebook, SWDI is defined as RZSM minus the field capacity or the wet point of the soil over the field capacity minus the wilting point or the dry point of the soil.

11:40 Speaker 3: And if you're confused about these variables, the key for what they mean is down in the bottom right of this slide. One other detail to keep in mind is that field capacity is not directly provided by the land surface model that is used to derive SMAP L4 data. So we have to obtain it ourselves using the column water capacity from the model divided by the depth of the soil profile, which is the D value from our previous slide showing the RZSM diagram. times the density of water, which is just 1,000 kilograms per cubic meter.

12:23 Speaker 3: And so what are we trying to accomplish with this formula, since it seems pretty directly correlated with RZSM itself? We're trying to divide out the bias in the RZSM estimate to allow for the creation of ensemble SWDI if multiple model sources are available. So another way of thinking about that is When you see an RDSM value, it's in these ratio metric units. And whether that value that we're seeing is wet or dry relative to the soil conditions, it's sort of hard to tell just based on the value alone.

13:05 Speaker 3: Putting it into terms of SWDI allows us, in theory, to say minus 10 is dry, 0 is normal soil conditions, and then 10 would be very wet soil conditions. And in this way, we can define a severe deficit of soil moisture by any value less than minus 5, regardless of the source of our RZSM estimates. The key criticism of this approach is that there's we have trouble having enough confidence to say that less than minus five is actually a severe deficit in soil moisture. And so these thresholds are only physically meaningful if you trust the estimates of field capacity and wilting point.

14:02 Speaker 3: And if we go back to the plot of SWDI that we got in the first part notebook, The question we would ask is, is SWDI less than or equal to minus five a severe soil moisture deficit in reality based on ground conditions? It can be difficult to say. So there's some considerations we have to make about where we get the values of these constants from. Over the weekend, it came to my attention that there's an issue regarding what I've said here about the calculation of the soil water deficit index. And specifically, the issue is that the GMAO catchment model constants are not meant to be used to derive field capacity and bolting point values in the manner that we've shown in the notebooks.

14:57 Speaker 3: So because of this, the SWGI values that we're getting are biased because our FC and WP are basically technically not physically valid for this application. This has given us reasonable results for the case study that we've looked at so far, as well as the South Africa case study that we'll be looking at today. However, the recommended approach is to derive these constants directly from the 10-year SMAP soil moisture record on a per-grid cell basis. And I'm going to update the Jupiter notebooks over the next week or so to use this method and I'll be updating some of the plots accordingly.

15:45 Speaker 3: So I apologize for any confusion that this may have caused. Nevertheless, the conceptual discussion of using soil moisture for flash drought monitoring remains valid and the overall interpretation of the results changes very little because SWDI is linearly related to roots on soil moisture. However, the absolute SWDI values shown in some examples, including in the homework exercise, are not going to be reliable compared to using the percentile method. And in the Louisiana example that's mentioned in the homework, the unusually large SWDI values can be explained by the inappropriate use of FC from the land model constants.

16:34 Speaker 3: Once the notebooks are updated, I do plan to update the homework. So yeah, that's all I wanted to mention about this. So when we want to compare SWDI to precipitation, evapotranspiration, or other metrics, we can look at other studies that have used similar or the same metrics. Again, looking at our first case study, Let's see if we can track notable precipitation events represented in the SWDI as well as from other independent sources. In the bottom left, this is a plot from the He et alia paper from 2019 showing precipitation as the blue bars and then vapor pressure deficit and temperature anomaly accumulation as the lines.

17:29 Speaker 3: So it's a little bit challenging to read this plot. to be honest, but two of the observations we can make are that there was a June minor precipitation event in the Arrow marked one, and then there was a more major September precipitation event in the Arrow marked two. And if we turn over to our own SWDI plot, you can see those same events marked in one and two. And Again, looking at the bars of precipitation from the He et alia plot, where we see precipitation deficit decrease is where we would expect there to be precipitation events.

18:17 Speaker 3: And similarly, we can see this represented in the SWDI line from the Mohammadi and Wang impact matters paper. And again, if we want to see that relationship map more directly, just in a qualitative sense rather than doing a specifically quantitative analysis of this correlation, you can see that the timing of the precipitation events

18:50 Speaker 3: lines up quite well with what we're seeing on our own SWDI plot. The relationship between VPD and temperature anomaly is more complicated, but we can see that periods of increasing values on these lines lines up to where we see decreasing SWDI, which is what we would expect. And to really drive the point home and look at our detection criteria more specifically, we can compare the results from the Osman et alia 2021 paper with our own. So what we're looking at specifically here is that similar to the Martinez-Fernandez paper, in the plot on the right, the bars at the top of the graph represent precipitation events.

19:46 Speaker 3: And we can see that these line up with increases in SWDI which is the black line on the plot.

19:54 Speaker 3: And in comparison, our SWDI plot shows a similar shape to what was obtained from the other study. The researchers in Osman and Alia used a slightly different technique for determining the start and end dates of the flash drought, where they looked at when the five-day moving average of roots and soil moisture falls below the 20-day moving average for at least four consecutive five-day time steps. And the way we can interpret that, if you look at the dashed lines, We have our 20-day in the blue dashed line, and then we have our 5-day in the other dashed line.

20:48 Speaker 3: And the vertical lines show when the flash route was first detected, which was around June 23rd for their technique. By comparison, in our technique, if we use a simple threshold of just when we fell below the minus 5 SWDI, we would obtain a start date for this event around July 3rd. We'll use a more sophisticated technique for detection in the notebook, but this is a good starting point for comparison. So just using our simple threshold, we detected the flash drought event starting a little later than the other techniques, but around the same time, within about two weeks.

21:39 Speaker 3: So when we're talking about SWDI, I want to review why we can use it as a metric on its own without needing to consider precipitation and evapotranspiration. And to go back to our definition of agricultural drought, the precipitation deficit alone lacks the relevance to say whether or not the event is impactful. Other papers have used evapotranspiration as a more sophisticated way of detecting drought than precipitation deficit. And two examples of metrics that those papers might use, and these were also considered in the Mohammadi and Wong paper, are the evaporative stress index and the evaporative demand drought index.

22:29 Speaker 3: So for ESI, That would be the ratio of actual evapotranspiration to potential evapotranspiration, where ET is derived from the land surface temperature, and then potential ET, or PET, is derived from meteorological variables like the VPD and temperature anomaly that we looked at earlier. EDDI, by comparison... is purely based on PET, which, as I mentioned, comes from the meteorological data. And although ESI and EDDI are definitely relevant to drought, we might not see the agricultural impacts from increases in evaporative demand until later.

23:17 Speaker 3: The reason for that is essentially because if you have a surplus of soil moisture going into a drought the vegetation might be able to use that surplus of moisture even in the face of increasing evaporative demand. And so the agricultural impact might not manifest until later. In the case of the 2017 case study, we can see that ESI and EDDI will rapidly fluctuate between periods of hot weather and periods of cooler weather. And while we might flag flash drought events based on periods when these metrics are particularly severe, they might trigger false alarms where we don't actually see those anomalies manifest in the SIF or the soil moisture.

24:17 Speaker 1: So at the end of.

24:19 Speaker 3: The day, The theory that we're working with about why we use roots-on-soil moisture instead of an evapotranspiration-based metric, which is more common in a lot of other studies, is that RDSM is more holistic. We're looking at a supply-side variable versus a demand-side one. A way of looking at that from the Hood-Alias study is that the evaporative demand was high earlier in the season and throughout the season. But the SIF didn't decrease until about late May or June. Now, they're working with monthly averages.

25:04 Speaker 3: So we don't quite have the time resolution to say when exactly this occurred, but that is the theory. So now let's turn over to SIF and why it's useful for agricultural drought monitoring. To determine what is an agriculturally impactful drought, we want to use a measurement that is highly correlated with photosynthetic function because that's what matters for crop growth. And it responds rapidly and linearly to environmental stressors. And we also want to use a measurement that doesn't require in situ, if possible, so that we can use remote sensing for a global context to our analysis.

25:50 Speaker 3: So the ideal measurement that we would want to use is gross primary productivity, which is a direct measure of photosynthetic function, but it can't be measured from space. It has to be modeled. On the other hand, SIF can be measured using the OCO-2 or OCO-3 instruments, as well as a number of other spacecrafts. While flux towers can measure DPP in situ, these have a relatively small coverage over the Earth. And in this next section, we're going to look at the Yoshida et alia study from 2015, which looked at a 2010 Russian flash drought by comparing GOM-2 sift to MODIS NDVI, as well as a model DPP data set and tower DPP.

26:45 Speaker 3: And one of the findings from this study that's also substantiated by a number of other research studies is that SIF and GPP are linearly correlated. And the study found that the correlation was stronger over croplands and grasslands, which is pretty convenient for our technique because those are the main two land cover classes that we want to look at. If you look at the plots on the right, I would say to not worry about the boxes and colors just yet. I know this is a bit of a confusing plot to read through.

27:18 Speaker 3: But what we're interested in is that the correlation coefficient between model CIF and model GPP is quite high at 0.8. And when we look at the observations, while it's definitely lower compared to model GPP and model CIF, we still get a 0.53 correlation coefficient.

27:42 Speaker 3: And so there's a number of different vegetation indices that we could use that have more spatial coverage and are more widely available. Why would we want to use SIF instead of these other vegetation indices? And the theory here is that the SIF has a closer relationship to GPP than other vegetation indices. And furthermore, it's based on an emitted signal as opposed to a reflectance ratio like NDVI or EVIR. One of the benefits of that difference is that the SIF, in theory, responds more rapidly to environmental stresses.

28:34 Speaker 3: When we look at the case study that Yoshidi Adelia looked at, in particular over this cropland region in box number five, they saw that the sift had this more jagged line represented by the black dots on the plot. In comparison, the NDVI dots in magenta are smoother. and they don't respond as quickly to changes in the soil moisture and vapor pressure deficit anomaly that's shown in the bottom subplot. These solid lines show the climatological mean for SIF and NDVI. So we can see that the fact that our dots are lower shows that there was an overall deficit in vegetation growth during this growing season, which is indicative of the drought event that occurred in 2010.

29:35 Speaker 3: And so the theory behind why this is the case for SIFT other than just that it's a measurement or an estimation of an emitted signal, is that SIF is sensitive to the fraction of photosynthetically absorbed radiation and the photosynthetic yield. And when we talk about drought, Drought specifically causes declines in these two metrics. So the idea is that we can observe them more directly with ZIP than with other vegetation indices. This relationship is the most acute over croplands and grasslands.

30:14 Speaker 3: And so when we look over at the plot on the right, the mixed forest, the diamonds and the Xs, have more of a horizontal spread compared with the croplands and grasslands, which are the pluses, the triangles, and the squares down in the bottom left of the plot. And the greater the horizontal spread on this plot, the less correlated the variable is with the linear trend. So the idea here is what we're trying to show is that croplands and grasslands are more directly tied to these vegetation stress indicators during a drought than mixed forest or broadleaf forest.

31:10 Speaker 3: And so why do we want to use SIF-RCI in place of SIF? To detect agriculturally impactful drought, we want to quantify a departure from the normal seasonal growing cycle. If we take a look at the time series of SIF in the top right, we can see that there's a flat top in our dotted line, which is indicative of the actual start of the drought event, but we don't actually see a negative SIF anomaly until later in the growing season, a few time steps later. When we use SIF-RCI, we're able to see that downtrend manifest itself sooner than the negative anomaly appears.

32:00 Speaker 3: And it's trying to characterize this flat top, our decreasing slope of SIF compared to the climatological mean line of SIF.

32:14 Speaker 3: So we don't want to just detect when we fall below the mean. but specifically when the cumulative anomaly becomes impactful.

32:25 Speaker 3: If we were to calculate CFRCI from the data in the Yoshieda et alia paper, like in the bottom left, we would see a similar effect where the drought manifests itself relatively earlier than we would see the negative anomaly become impactful. So now let's compare what we got for our results with the U.S. Drought Monitor to see if we would have had any statistically meaningful difference in when we detected the start of the event with when the U.S. Drought Monitor declared a drought in this region.

33:07 Speaker 3: If you look at the map on the left, this comes from the historical conditions page on drop.gov, which I've included a link to in the references slide at the end of this presentation. When we look at June 2nd, the most recent U.S. drop monitor map that we had available was from May 23rd, 2017. By comparison, The two time series that we have on the right come from our technique.

33:40 Speaker 3: And when we look at the region of interest specifically, represented by the box in the middle of Montana, there's actually no drought in this region in late May. And while we do see an abnormally dry region in the eastern portion of the state, it hasn't yet reached the level of a full-on drought. By comparison, SIF-RCI has already gone below zero for one time step. And SWDI is trending downwards, although it hasn't quite reached that minus five severe deficit level yet. So the question we need to ask is, is there enough information at this point using our technique to confidently declare a flash drought?

34:29 Speaker 3: The easy answer, based on our detection rules, is that No, not really. We need to see additional time steps below the threshold that we use and a certain level of decrease to be able to confidently declare a flash show. However, when we look at these time series, we can definitely see decreasing trends in both SIF RCI and SWDI, which would point us towards being on alert for the development of a drought if conditions got any worse. And one other thing to note is that arguably it's unfair of us to use May 23rd compared to our June 2nd data.

35:17 Speaker 3: But because the USDM is released about every two weeks, We don't always have the most recent USDM to work with, and that's just a limitation of it compared with our technique, which we can update every three days for SWDI, which is also fairly arbitrary. We could also update it daily and every eight days for CFRCI. So let's flash forward about two weeks to June 18th and look at where we were at with our metrics. In the USDM, we can see that a moderate drought had appeared in the eastern portion of the region and was spreading into the middle portion of the region of interest.

36:00 Speaker 3: In comparison, our CFRCI metric would have already flagged this as a flash drought event. SWDI had not quite reached the level of minus five, but if we were using a metric of SWDI less than or equal to minus three, for a moderate deficit, then we would have flagged this event as a flash drought by this point. While we would detect in the USDM that there was a drought developing, it's hard to say how severe this event would be. So maybe the farmers in the western portion of the region might not be considering this to be a drought event yet.

36:44 Speaker 2: Okay.

36:46 Speaker 3: Now, let's take a look at very late in the season, around the 8th of August. And by this point, we can see that this drought event was already well underway, as is indicated by the exceptional drought in the eastern portion of the region of interest. And for our own metrics, we can see that by this point, we would have flagged both SWDI drought and CFRCI drought. So if we were using the metrics here, we definitely would have flagged this event. So that's not the problem. The question is, would this have been an early enough warning?

37:34 Speaker 3: And the answer is essentially that we would have need to have flagged this event in early to mid June to have an advantage over the USDM. If we were flagging it as late as late July or August, we would have had no advantage over what was clearly evident on the ground. And that's why it's important to understand the context of the agricultural situation and what information is already available to most users when we want to look at the effectiveness of a technique. If it's telling us that there's a drought when other techniques are already showing us that there's a drought, then there's very little point in using this technique.

38:25 Speaker 3: By August, the fields were already drying down if the farmers were growing winter wheat or spring wheat or barley. So the harvest season was pretty close at hand. And farmers would have been well aware that there was a drought and that crop damage had already been done. Within another month, If you recall, when we looked at the map view, there were those burn scars in the reflectance data. There were wildfires that broke out in the region in September.

39:07 Speaker 3: We don't have enough time during this presentation section to fully go over the relationship in terms of the effectiveness of CFRCI and SWDI as flash drought detection metrics, but these slides will be available for you to look at after this session if you're interested in their relative effectiveness in terms of detection. The long and short of it is that the study that we're basing our technique off of showed that The two metrics had the greatest skill over croplands and grasslands and a relatively lower amount of skill in forest regions.

39:58 Speaker 3: Okay, so now let's start talking about our case study. And Melherb et alia, 2025, examined a summer flash drought in the 2023-2024 growing season in South Africa. although this event had ramifications across much of the southern portion of Africa. In particular, the region we're looking at has rain-fed cropland with maize as the primary grain. The event lasted about two and a half months, but it had a significant impact on yields, including crop failure in some portions due to its timing at the peak of the growing season.

40:42 Speaker 3: In the Jupiter Notebook, we're going to compare the effectiveness of our SIF-RCI-SWDI technique to the method used by the researchers, where the researchers essentially had their own advanced drought early warning system, which is a web dashboard that looked at a number of meteorological variables as well as a few reflectance-based vegetation indices. So, as I said, these researchers had their own system called ADUs, which looked at a number of different metrics, such as the standardized precipitation index from local weather stations, the vegetation condition index, which comes from MODIS-VERE's reflectance data, the percentage of annual seasonal greenness, and soil moisture data derived from GRACE and GRACE follow-on, where available.

41:40 Speaker 3: So what are the rules that we actually want to implement in our notebook? Well, there are no hard and fast rules for flagging a flash drought. In other words, all of the studies that look at this phenomenon use slightly different detection techniques and different indicators. There are a few guidelines that we can base our approach off of. So in our initial implementation, we're going to be using the detection rules that Mohammadi and Wong used. And this would be that the 30-day moving average of SWGI decreased by two points, and SWGI was less than or equal to minus five for at least 12 days, or four time steps.

42:28 Speaker 3: For CIF RCI, we'll use the rule that it had a value of less than negative 0.5 for at least 24 days. And this might be a little confusing because in the plots that we looked at earlier, we used less than zero as our criteria. But as you'll see, we're going to use these rules on a per grid cell basis, which provides them a greater degree of granularity compared with our spatially average time series. If we assess the case studies, and compare them to the USDM or the ADUs to see if they're good enough, at least qualitatively.

43:15 Speaker 3: We can't necessarily do a full-on correlation analysis. We can see the effectiveness of the rules. In general, these rules have been developed based on a maximization of their predictive skill, but they're still relatively subjective. And so they can be modified slightly if needed. And one of the additional things that I encourage considering, which doesn't necessarily come from the research, but just from understanding the overall context, is to look at the crop calendar. And we add the crop calendar to the plots that we looked at in this course as more than just a plot decoration.

44:10 Speaker 3: A flash drought is arguably more relevant if it occurs during the growing season than, for instance, during winter. We might just be looking at relative anomalies in dormant vegetation, for instance, if we see a SIF anomaly in the middle of winter.

44:31 Speaker 3: In general, a flash drought outside of the growing season may have little to no impact on yields. In fact, you may often see a negative SIF anomaly during senescence. If we had an early planting time for the particular year that we're looking at, you might expect to see early senescence, since the growing season is often defined as a number of days rather than harvest always occurs on X day, for example. In the U.S., one of the key sources we can use to get the crop calendar guidelines is from the link that I've included here.

45:13 Speaker 3: Unfortunately, if you're looking outside of the U.S., there's no single universal source for the crop calendar, so you'll need to do your own research to find what the typical planting and harvest dates are for a given crop. And if you look at the figure on the left, You can see an example of a field of den corn being harvested by a combine. The farmers that harvest this grain want the crop to be dried down so that there's less water weight during the harvest. And so let's look at some other years for our two case study regions.

45:54 Speaker 3: If you look at the top left, you can see the plot of SIF RCI is showing that there was a relatively higher SIF anomaly early in the season, and this might correspond to an earlier planting date. On the other hand, later in the season, close to where we've highlighted the harvest occurring, we definitely have a moderate late season flash drought because we see a SIF RCI deficit as well as a SWDI deficit decline reaching into severe levels during the harvest period. Whether or not this was impactful to yields really depends, but we might not necessarily consider this event to be as impactful due to occurring during the crop senescence period rather than during the peak of the growing season.

46:54 Speaker 3: In our South Africa case study, we can see in a year other than the year that we're going to look at, there was a soil moisture deficit before the start of the planting season. And whether or not this actually impacted crop yields is not necessarily clear, but we might place less relative importance on this event since it might have primarily occurred in wildland regions rather than over the cropland regions. We'll look at the spatial distribution of our flash route detections to determine a little more granularity about our detection approach.

47:41 Speaker 3: And so if we want to extend the technique to the present day, we can look at more recent data in worldview. And so on the right here, I've included a comparison between the July rootstone soil moisture levels in eastern India on the left side and a previous year at the same time of year on the right side. One of the things that we can see is that you can use the color map to notice that this eastern region of India has much lower levels of roots and soil moisture in July compared with previous years.

48:28 Speaker 3: And this basically indicates that there might be something going on here that we could look at in more detail. What we would need to do though is because our virtual data set that we used for retrieving soil moisture only goes up to October 2025, we would have to create a new virtual dataset using the Appendix Notebook, which I've included instructions for.

48:57 Speaker 3: Some other things that we could do is to assess the predictive skill of different detection rules more broadly by

49:08 Speaker 3: using the data on a map at one degree spatial aggregation and comparing it to a catalog of known flash trends. that have occurred in the past. This is easiest to do in the United States when we have the USDM to work from. But if you have a catalog of flash droughts in other countries, you could try that as well.

49:38 Speaker 3: If you just want to look at one discrete event for a region of interest, though, you should be able to use the notebooks as is. And I've included instructions for that. What you would want to do is pick your region of interest, look at the crop types and the crop calendar, and then input the values for your region into the Part 2 notebook. And you'll probably also need to run the appendix notebook as well to get a climatology So I'm not going to go into this particular study in detail, but if we wanted to get more fancy with our detection technique, we could use a classifier machine learning model instead of simple rules.

50:24 Speaker 3: And one study from Tong at Alia did exactly that, which you'll be able to read more after the presentation if you're interested. And so... One of the key problems with GoSIF is that it's not forward processed. So we don't have data going up to the present if you wanted to study 2026 and beyond flash drought events. While the researchers who developed it do update the data set on a yearly basis, you're going to need to use a different source of SIF data if you want to study present day events. Some of the options that you have for this would be tropomy SIF, which has a 5 kilometer by 7.5 kilometer resolution.

51:10 Speaker 3: And that's what I'm showing over on the right. Alternatively, it is perfectly fine to use NDVI or NDVI if you'd like.

51:20 Speaker 2: But as we.

51:22 Speaker 3: Discussed, those have some lag compared with SIF, at least in theory. You could, however, convert them to an RCI time series to potentially improve your early warning lead time compared with other techniques. And soon, the ESA Flex mission will measure SIF at a 300-meter resolution, and that will provide a new data source for SIF. Okay, so that takes us to the end of the presentation. Here are the references that I used, including some useful links if you wanna look at the maps that I showed and the slides and worldview.

52:07 Speaker 3: And now we're going to turn over to the demo portion. Right, so now we're going to look at the Jupyter Notebooks. So for session two, we're going to talk about the 2023 to 2024 summer flash drought in South Africa, as I mentioned. We're going to actually implement our detection rules, and then we're going to look at the spatial distribution of our detections on a map to see what we can figure out.

52:35 Speaker 2: By the end of.

52:36 Speaker 3: Session two, you should be able to apply the techniques for deriving CIF-RCI and SWDI time series across a broader range of contexts and time periods. You should be able to compare different flash drought detection rules in the context of the case study event. You should also be able to identify the spatial component of flash drought development. And lastly, you should be able to compare the differences between SIF and soil moisture as early warning signs of drought. Okay, so I've pulled up the Jupiter Notebook for part two.

53:15 Speaker 3: And if you wanna follow along with me, I cover the instructions in the README of the code for the course. you'll just be able to run the setup script again, and it will open up the JupyterLab environment. It should open the drought indicators notebook, the one from part one automatically. But if you click to detection down here, that will open up the part two notebook, which is what we're looking at right now. And so, as I mentioned, we're going to cover flash drought detection. And We're going to implement the detection rules for SIF RCI and SWDI that I talked about.

53:59 Speaker 3: There are two key criteria that we need to think about. Number one is the drought severity, and this is determined by a threshold value below which we declare a drought. The second criteria is the speed of the drought development. In other words, how quickly it's dropping. And we don't use the speed of drought development for CFRCI since it already takes it into account with its value to some degree. So we'll just be looking at the speed of drought development in terms of the SWDI metric.

54:35 Speaker 3: In research papers, these values are chosen based on a rigorous correlation analysis. But we're just going to take our values from an existing paper and then modify them a little bit in different contexts to see if it makes a difference. And so let's talk a little more detail about our case study. As was mentioned before, we're looking at South Africa in the 2023 to 2024 growing season. And remember, since this is the Southern Hemisphere, the summer season is from December to March. The region of interest covers the Free State, of South Africa in the Maize Belt, as well as a small portion of the country of Lesotho.

55:23 Speaker 3: And in the paper, the regions that are talked about in our particular region of interest that's used in this notebook are the Windberg, Senecal, and Bethlehem Magisterial Districts. The paper also discusses other districts, but we're going to be using a smaller, about one degree, by one degree lat-lon area for our discussion. And when we talk about maize in this region, we want to think about the crop calendar, like I mentioned in the presentation. It takes about 140 to 160 days for maize to mature.

56:04 Speaker 3: And the planting window is usually from late October to early December. This corresponds, therefore, to a harvest of around April to mid-May. And these values come from a paper about the maize growing period in Free State, which you can read if you're interested. And the first thing we're going to do is essentially a repeat of the first notebook, but for our new region of interest and our new dates. And I've cleaned up some of the code that we went over in the first section to be more streamlined to do this analysis quickly.

56:44 Speaker 3: What we're going to do is, number one, retrieve the necessary Go SIF data for this time period, compute a time series of SIF RCI and SWDI, saving the results as a CSV, and then plot the time series in a single figure with two subplots. Another modification we're going to make from part one is that we're going to look at the Z score of SIF instead of what we called the SIF standardized anomaly in part one. So that z-score that we looked at was the z-score of the change in sift between time steps, whereas this time we're looking at the z-score of the anomaly for a given time step.

57:27 Speaker 3: And that's effectively summarizing the departure from the climatological mean as opposed to the departure from the trend in the growing cycle. And the reason why I made that change is because it provides more diagnostic information about what's going on. And so when we look at this cell down below, I've also included instructions in the comments about which procedure you're going to need to follow to change this for your own region of interest if you want to do that analysis on your own after this demonstration.

58:12 Speaker 3: And so this first cell contains all of the inputs that we're going to use for doing this flash drought analysis from beginning to end. I've identified a region name, which I'm calling South Africa here. And I also have just a string specifying the name of this region that I want to appear in my plot titles so that it's easier to read. The first important value that we need to input is the coordinates of our bounding box. And unfortunately for right now, you're only going to be able to use rectangular bounding boxes, but that's going to be good enough for our analysis in this case.

58:58 Speaker 3: And the way we input these coordinates is in the west, south, east, and north order and basically

59:14 Speaker 3: in the eastern hemisphere we have positive values so our convention for the coordinates is minus 180 to 180 where minus 180 will be in the western hemisphere and 180 will be in the eastern hemisphere and And then 90 to minus 90 going from the North Pole at 90 and the South Pole at minus 90. And then we need to input the date range of our analysis. So in this case, we're studying the 2023 to 2024 growing season. So I put in a date time object that goes from September 1st to August 31st. And this basically covers from just before the spring planting through to a period after the harvest.

1:00:03 Speaker 3: And it gives us some buffer around those planting and harvest dates that appear in the crop calendar. The next thing that I find is the crop calendar dates itself. And these don't have to be precise. It's really just a guide for us to understand the relevance of anomalies in the data that we're going to see in the time series. And in our case for this part of South Africa, the planting we're going to say starts at October 20th and ends at December 10th. Likewise, the harvest starts at April 1st and ends at May 15th.

1:00:47 Speaker 3: So I'll just run this cell. And this is not going to do anything at first. You can see that It ran over here, but we have nothing happening yet. The next thing I'm going to do is download the gosif data. Now, I've already done that, but if you haven't, this is going to take a few minutes to download those dotif files, and they'll be in that same data directory, data slash gosif, that we had before.

1:01:16 Speaker 1: Okay.

1:01:20 Speaker 3: In the next cell, we're going to actually compute the time series themselves. And this is another thing that will take between 5 to 10 minutes, mostly due to the length of time that it takes to compute the SWDI time series. And these will appear in the data directory at the top level. And so since I've already done that, this cell will run basically instantly. And you can see that it's already run. And we can look at the outputs here. So here is the SWDI time series for 2023 to 2024. And we can see the dates as well as the values of SWDI that we get for our spatially averaged time series.

1:02:16 Speaker 3: One key difference from what we did before is that we also save rasters of our computed drought indicators for each time step. And this is going to be located in the South Africa time step directory. If we go in here, we don't have to look at these in detail for right now because we'll look at them later in the notebook. But you can just see that we've created a number of different swdi and then a date.tif files. And these are the swdi values for each grid cell at each time step. And the same thing has been computed for CIF.

1:03:05 Speaker 3: So, going to the next cell, this is where we're actually going to plot the time series. All right. And so, This is going to be a little big to show on the screen, but I'll open it up in a new tab. All right, so I've opened up this plot in a new cell so that we can look at it a little more closely. As I mentioned at the start, we changed the z-score that we're looking at to the true anomaly in CIF versus the anomaly in the CIF increment. And the SIF RCI is the dotted line that we're looking at here.

1:03:50 Speaker 3: In this case, we see a really strong flash drought signal in the SIF RCI. And looking down here, we can see that that first reaches below that zero point around early February or late January and reaches its lowest point in March. and then it recovers later in the season. The reason we're using SIF-Z score in this case is to point out an important point about interpreting positive SIF-RCI. And so SIF-RCI is a statistical tool that we use for detecting flash drought, but when you see this return to positive values you might naturally assume that the vegetation recovered during this period of the year because we're seeing positive SIF RCI.

1:04:50 Speaker 3: But because of the way that it's formulated, that might not actually be the case. Early in the season, we do in fact have a positive SIF anomaly associated with strong vegetation growth. And the SIF anomaly, the Z-score, is a lagging indicator So it doesn't actually reach negative values until several time steps after the SIF RCI does. And then later, we have increasing SIF anomaly, but it's still negative. So the vegetation is still being impacted by this drought event. And then it reaches about the climatological mean a little after what we consider the harvest period to be.

1:05:33 Speaker 3: So actually, we would not necessarily consider this vegetation to have recovered by late season. And in fact, if we're looking at just the maize crop, it definitely wouldn't have recovered in this particular season. Now turning over to the SWDI, one of the things that we can see is that unlike our previous case study, the soil moisture is persistently pretty low throughout the entire year. And While there is some recovery early in the summer from what it was during the planting season, the soil moisture then decreases throughout the summer and reaches a low point in March.

1:06:22 Speaker 3: These spikes upwards, as was mentioned in the presentation, are associated with precipitation events. And especially over here, early in the season, the reason why it's a bit higher than later in the summer is due to above average precipitation in that December to January period. And this observation is substantiated by what's shown from standardized precipitation index data in the study.

1:06:59 Speaker 1: Okay.

1:07:01 Speaker 3: So I wanna go over one last point before we move on, which is about interpreting SWDI. Because the values for SWDI that we see throughout the entire year are very low, there's a couple of different interpretations that we can have. If we trust SWDI, we might say that this is an example of a transition to a longer term drought. But if we wanna drill a little deeper into the values of the constants that we've chosen to define SWDI and the accuracy of the roots and soil moisture itself, we can look a little deeper.

1:07:41 Speaker 3: To talk about the accuracy of the RZSM itself in a little more detail, we can look at worldview. So if I follow this link to worldview, it's going to open up a map of this case study. And one of the things that we can see when we look at this, here we're plotting the roots and soil moisture. So this is the data that SWDI is derived from. We can see that there's quite a big discontinuity between the South Africa side of the border and the Lesotho side. And we might reasonably question how real is what we're seeing in this data.

1:08:24 Speaker 3: And the answer is that it's partly due to uncertainties in the model and assumptions in the way that the model considers these land processes that creates discontinuities like this. It's partly due to the retrieval of soil moisture itself. And it's partly due to real conditions in the environment that are being modeled. So If I remove this layer and we look at the underlying land surface reflectance, and I remove international boundaries to see, can we see something in the land surface reflectance without the political boundaries being visible?

1:09:11 Speaker 3: The answer is actually yes to a Here we see this clearly defined change in reflectance color between the Lesotho side of the border and the South African side. I'm not able to zoom in very far on this 500 meter resolution data, but if we look at this more closely, we can see that the vegetation density is genuinely lower on the Lesotho side compared with the South Africa side. One of the consequences of that to soil moisture that we might reasonably expect is the lower amount of vegetation drives lower levels of evapotranspiration.

1:09:56 Speaker 3: Another difference between Lesotho and South Africa is that Lesotho is significantly more mountainous, so the orographic lift drives greater amounts of precipitation. And if we were to look at the Level III soil moisture product, the observations themselves, we would actually see that there is somewhat of a difference between the soil moisture in Lesotho versus on this side of the border in South Africa. Now, on the other hand, one of the challenges that we have is that the soil moisture is very difficult to estimate over urban areas.

1:10:39 Speaker 3: And so we also see a higher level of soil moisture compared to the surroundings over here, which is associated with more urban areas. So those are some of the challenges of using roots and soil moisture that we can see in practice. And nevertheless, in this primarily agricultural region, we'll see that we're still able to detect the drought signal anyways. So now let's talk about applying the detection rules.

1:11:21 Speaker 3: As was mentioned in the presentation, for CIFAR-CI, we have a very simple rule, which is just that if a given grid cell is below negative 0.5 for at least 24 consecutive days, or three time steps, then we flag it as flash drought. And that will naturally stop flagging the flash drought once we see a time step above that threshold. For SWDI, we look to see when our values have decreased by at least two points over the preceding 30 days or 10 time steps. And by default, we're going to use this severe flash route definition, which is SWDI less than minus five for at least 12 days in addition to meeting the qualifications for rule one.

1:12:16 Speaker 3: In the homework for the course, you're going to look at a case in southern Louisiana where we use this moderate flash drought definition due to the different climatological conditions where the soil is much wetter in that region compared to the regions that we've been looking at so far.

1:12:36 Speaker 3: In terms of an end date, we would look to see that SWGI is greater than minus five for 12 days or more. And then in terms of the combination of these two metrics, we're not going to implement these particular combinations, but we'll talk about them briefly. If we see that SIF RCI flags a start date for a flash drought without a corresponding SWDI start date, we might consider this to be a significant SIF anomaly, and it might not be due to flash drought. It could, for example, be due to floods or fires that could also cause detections in the SIF RCI.

1:13:20 Speaker 3: Another key thing that we want to look at is the slow drought transition. So this would be when we detect no SWDI end date within the time range. Another definition that we could use for the slow drought transition would be to consider when SWDI is below a given threshold, such as minus five for 60 plus days. Okay, so In the cell that we're looking at here, we have two functions at the beginning, detect flash drought in SIF and detect flash drought in SWDI. The implementation of these functions is actually in one of the course code files because they're a little involved.

1:14:13 Speaker 3: But if we open them up over here, you can look at their specific implementation if you're interested. So, for instance, here we have a docstring describing the implementation. And we're essentially going through the rasters in the raster directory that we created earlier when we computed the time series. And we're going to add a column to the CSV that we wrote earlier with the percent of grid cells that have a positive detection for a flash drought. And we go through and do this iteratively per raster.

1:15:09 Speaker 3: And we do the same thing for CIF-RCI. Once we've done that, we're going to then create plots of those two metrics. And what we'll say to start is that when more than 50% of the grid cells in this region, which is about one degree, as I mentioned before, when more than 50% of those grid cells are flagging a flash drought, we'll consider the whole region to be in a flash drought. And that's what we'll use as our detected start date. And similarly, we'll flag the end date. So I'm going to run this plotting code now.

1:15:53 Speaker 3: And you can see that for our case study over South Africa, we flagged the CIF RCI flash drought as beginning on February 18th and then ending on March 21st. And likewise for SWDI, we're flagging the beginning of this event on February 10th. And these two dates are very close to each other. So that's a very promising sign that we're looking at the same overall trend in the environment between our two different remote sensing metrics. Another thing that's interesting to note about this data is that we're rejecting detection of flash drought in other parts of the year.

1:16:46 Speaker 3: So for instance, in CIF-RCI, and this is the same CIF-RCI time series that we looked at in the previous plot. We're not seeing any grid cells going into flash drought when we go into a negative value for a brief period of time.

1:17:08 Speaker 3: And then over here in SWDI, we do actually see that briefly during the planting season, there are a few grid cells that showed positive detection of flash drought, but we don't flag the entire region as being in flash drought. On the other hand, when we get later in the season and we see this steadily declining SWDI, we do actually flag this event. And another interesting thing to note here is we don't actually exit the drought. This would imply that we're transitioning into a long-term drought.

1:17:48 Speaker 3: It's much more difficult to detect a long-term drought transition in SIF data and especially SIF-RCI. So really we're going to be using our SWDI data as the main criteria for deciding if a flash drought has transitioned into a longer-term drought event.

1:18:10 Speaker 3: So now let's compare our results to the ones obtained from the study. And one of the things to keep in mind is that the paper itself does not flag specific dates for the start of the flash drought event like we did. But they do specify that January to March is the peak period of interest for the onset of this event. And that's kind of where they compare the climatological data from January to against what was available for February and March. So our data sources corroborate the observation that there was above average rainfall in January, which is why we see this relatively higher level of SWDI at that time.

1:19:02 Speaker 3: However, the overall supply of soil moisture was still pretty low because we're hovering around this severe deficit level, which means that when there was a precipitation deficit later on in the summer, it rather quickly moved the whole system into a drought. As I mentioned at the beginning, the region of interest we're looking at in our data roughly corresponds to the Windburg region. Senegal and Bethlehem Magisterial Districts. In their study, they have a table, Table 2, which I've copied here below, and they indicated that those regions saw 59%, 59%, and 74% of normal maize yields in 2024, respectively.

1:19:53 Speaker 3: While our own technique does not have a rigorous way of estimating the severity of flash drought in terms of the end of season impacts on yield, the positive detection from the plot in both CFRCI and SWDI lines up well with the timing and the yield figures show that this was in fact an agriculturally impactful drought. Another thing that we can see when we look at the table is that these districts had significant precipitation anomaly. over a longer period of time. And that indicates to us that the longer term precipitation deficit would lead to a longer term deficit of soil moisture, since the theory is that those two variables are correlated with each other.

1:20:52 Speaker 3: So let's try to apply the same technique to the 2017 Northern Great Plains case study from part one. And to save time, I've already done this for you. And also because one of the open-ended questions from the homework is to apply this analysis to your own region of interest and to reproduce a plot from the homework. I'm not going to show you exactly how to follow the instructions in the comments at the beginning of the notebook. So this is a reproduction of the SIF-RCI plot for the code from above here.

1:21:40 Speaker 3: So it's the equivalent of this plot. And one of the things that we can see in the northern Great Plains region is that we actually don't flag a flash drought using our technique of detecting per grid cell, and then flagging a start date based on when we reach over 50% of cells in flash drought. And that's kind of an interesting result because our time series definitely shows that there was a flash drought during this period, a long excursion into negative CIF RCI values. So what's the problem here?

1:22:20 Speaker 3: And one of the ways that we can look at that is to look at the spatial distribution of these flash drought detections in CFRCI. If there's a bias where one side of our region is detecting flash drought and the other side is not, or one land cover is detecting it and the other is not, then we would be able to see that when we look at a map.

1:22:45 Speaker 3: And this is really important as we have larger regions of interest. In this case, our first case study, We looked at a five degree by five degree area, and that's generally pretty large to spatially aggregate. So there could be a lot of differences between one side of this region and the other. And if you recall from the presentation, this drought developed from the east to the west. So we might be leaving a lot of detection skill on the table when we average together such a large area and consider it all the same.

1:23:24 Speaker 3: Now let's turn over to the SWDI and what we were able to detect using the rules that were outlined here. And you can see that we reach our 50% area threshold around July 24th, which based on just the time series of SWGI alone, if we were applying it to the whole spatially aggregated area, we would flag this event much earlier in the season, in late June. And if we were applying it based on when individual grid cells were indicating a flash drought, we would detect it even earlier. The first grid cells to show signs of flash drought began appearing in mid to late May, whereas we didn't reach a 50% area until July.

1:24:22 Speaker 3: So again, this makes a strong case for us to look more closely at the spatial distribution of these points. There's also a number of different ways that we can modify the rules themselves if we find that it's not adequate for the specific context that we're looking at. So as I said, we're going to look at the map, but before we get into that, we can also change the rules. One of the things you could consider would be to lower this area threshold that I talked about. When we go up here, back to our cell, I have this FDThresh value over here, and this is the 50% area threshold that I mentioned.

1:25:11 Speaker 3: If you lowered the values for this, in this particular case here, you would be able to detect the event sooner than you would have been able to otherwise. But there's a problem with that, which is that you might have more false alarms So for instance, if any value over 0% detection counted, then we might consider this event in December to be a flash drought as well, which we don't necessarily want to flag those events. And similarly over here, we could definitely improve our ability to detect this event if we lower the threshold, but we'll get more false positives.

1:25:58 Speaker 3: So that's not necessarily the best approach to just lower the area threshold. We want to consider the size of the area that we're using and maybe decrease it. So the recommendation is that if you're making these time series for a region of interest, try to stick to one degree by one degree lat-lon regions. And this is aligned with the way that studies that look at the skill of rules overall work is that they grid a larger area into one degree sections. And then another thing that we can consider is to just change the absolute threshold for drought.

1:26:47 Speaker 3: And that goes back to the South Louisiana example that I talked about before. We're not going to be doing that in this case because the threshold of SWDI less than five is justified in this climate. But in general, you can change that by modifying the value of this ABS threshold or absolute threshold in the rules above. And to go back to the cell, we can see that we have this detect flash drought SWDI function that I mentioned. and it has a number of different keywords, and one of them is absolute threshold.

1:27:26 Speaker 3: If you wanted to use our moderate FD definition over here of less than minus three, then we would use abs threshold equal to minus three.

1:27:42 Speaker 3: And you can see that this is all documented as well in the code in analysis.py in the doc string. We can also change some of the other parameters, although these are less recommended, such as the length of time that SWDI is allowed to decrease over the drop threshold. But if we change the number of time steps in the look back, that changes the definition of the speed of drought development. So if you use a lower value, or sorry, a higher value, such as 20, we would say that the drought is allowed to develop over 60 days in that case, which might not meet what we would strictly consider a flash drought.

1:28:35 Speaker 3: If you use a lower value, you make the definition of flash route stricter and say it has to develop over a shorter period of time.

1:28:45 Speaker 1: Okay.

1:28:47 Speaker 3: So now let's look at a map of the spatial distribution of our flash route detections. And this map is

1:29:00 Speaker 3: I'll admit, a little hard to interpret. It was based on what was developed for part one. I sort of modified it. And one of the things that we can do when we look at this, right now we're looking at CIF RCI, and the opacity is 25% where we have not detected a flash drought, and then it will go up to 100% when we do detect a flash drought.

1:29:28 Speaker 3: It makes it a little difficult to see the value that we're looking at, but if I remove the NADR BRDF, the NBAR layer, then it's a bit easier to see the values that we're looking at. If I switch over to the SWDI layer, I can also see the values of SWDI at each timestamp. As we flip through the date, it's going to jump in three-day increments, and we can see the development of the drought. So let's start by looking at CFRCI. And remember, in our plot for this South Africa region, we looked at the 50% threshold being reached around February 18th.

1:30:16 Speaker 3: So that's what we'll look for over here. And I am going to jump. And we can see that all of a sudden, we had only about 0.3% of the area on February 1st. Then three days later, we had 4.3% of the area. And then a few days after that, we had 23% of the area. And we're seeing that actually the flash route detections are appearing more strongly over here in the northwest and less strongly over here close to the border. And then here's where we reach our 50%. And the SIF RCI is the lowest over here in the Northwest.

1:31:10 Speaker 3: And this corresponds with the Wittenberg district. Here's Senecal and then here's Bethlehem.

1:31:21 Speaker 3: And then we peak here in late February, early March.

1:31:28 Speaker 3: We're reaching 83% of the region in flash drought. And it's much easier to pick up this detection when we only consider one degree at a time, as opposed to using a larger area. You can, of course, consider larger areas, but it's useful to spatially aggregate at this scale. And then our CIFAR-CI signal begins to go away a little later on in late March. And eventually by April, we're no longer detecting negative CIF RCI. And to go back to the first plot, this wasn't because the vegetation necessarily recovered, but rather its senescence was decreasing at a rate similar to normal for the time of year.

1:32:19 Speaker 3: Now, if we look at SWDI, I'm going to go back to the beginning.

1:32:27 Speaker 3: And actually, I'll hide it so that I don't spoil what's going to happen. And when we look at January, we can see that we have no SWDI detections for now. And this corresponds with where we were looking at in the SWDI data earlier on. where we see the SWDI being relatively higher than before or after the flash drought event, corresponding with rains in January. On the other hand, we do have pretty low SWDI. Our definition for the drought detection rules are a little different from what was used in the Mahamadi and Wong paper in the sense that we look for decreases of two SWDI points from any starting value.

1:33:27 Speaker 3: And that's essentially to account for the constants maybe not being as accurate as we might hope or also for variations in the accuracy of the RZSM estimates.

1:33:50 Speaker 3: And as we go on, the flash drought begins to propagate from this border region where we have the strongest

1:34:03 Speaker 3: precipitation deficit from the paper. So that's a little different than the SIF RCI, which is an interesting observation. Another thing that we can note is that because the SWDI is higher on the Lesotho side of the border, we're not necessarily detecting flash drought on that side of the border compared with on the South Africa side. Nevertheless, as the season goes on, we see that the flash drought detections spread out from the border region northwards as we get into February. And eventually, we even do see the flash drought detection signal propagate to the other side on the Lesotho side of the border.

1:34:59 Speaker 3: And eventually, we're seeing low SWDI throughout the entire region. When we compare the spatial distribution of our flash route detections back in February, one of the things that we can note is that there are areas where the detection aligns with each other and areas where they disagree. There's a pretty complex relationship between CFRCI and SWGI, which is something that you'll look at more closely when you look at the South Louisiana example from the homework. And so we wouldn't necessarily expect the spatial distribution of the detections to be exactly the same between the two metrics.

1:35:47 Speaker 3: But when we aggregate their values to a one degree by one degree area for the purposes of declaring a flash scout or not, that should align approximately in terms of date in many cases.

1:36:11 Speaker 3: So I want to give some final thoughts before we close out this part of the demo. Individual grid cells often enter and leave positive detection states between time steps. So this reinforces the idea that just lowering the area threshold, that FD threshold that I talked about earlier, might result in false positives. A more time-stable approach could take into account the length of time that individual grid cells have been in flash drought when counting the overall status of a region, although we're not going over that in this part of the course.

1:36:51 Speaker 3: Another thing that we can observe is that SIF-RCI is well-suited to detecting the onset of the drought, but it tends to rebound to positive values over a period of weeks. A better detection criteria for flash drought might consider the Z-score of SIF, that other indicator that we looked at in the plot at the beginning of the notebook, in addition to the SIF RCI. In other words, the detection rule could use the consecutive time steps with negative SIF RCI to declare the onset of a flash drought detection, but then consider the Z-score of SIF to figure out when the end of that drought detection occurs.

1:37:34 Speaker 3: And then another observation is that our SWDI detection rule hasn't specifically identified when long-term drought occurs, although we do note that there's no end state, so that's a clue that this is a developing longer-term drought. Another thing, I brought up a few times that our land model constants might not be ideal to use for deciding what constitutes a severe soil moisture deficit. Another approach that we could use, which comes from studies that use SWDI, is to look at the climatology instead.

1:38:16 Speaker 3: And in those cases, we can use the fifth and 95th percentile of soil moisture over the 10 plus year SMAP record to derive the field capacity and wilting point instead. of just taking values that come from the land surface models. The downside of doing this in the notebook is that we would have to retrieve 10 years worth of data, which would take probably around an hour. So we're not going to go over that in this case. And you may have noticed that we didn't spend very much time talking about the detection rules themselves.

1:38:58 Speaker 3: And this is because I want to encourage you to understand and interpret what SWDI and SIF RCI are telling you, rather than to just apply detection rules and hope for the best. Ultimately, These detection rules are fairly subjective in nature in terms of the values that they choose. And the CNN-LSTM approach, like the paper mentioned in the Part 2 presentation, would be probably a more rigorous approach that can consider the overall trend in what these metrics are doing. So that would be using a classifier approach.

1:39:45 Speaker 3: And then the last point I want to make is if you want to run your own region, you also need to run the appendix notebook to generate the climatology inputs that you need. This will require downloading all 20 years from 2001 to 2020 of GOSIF data, which will take quite a while. So that's why we haven't gone over it in this demo. So that covers the second notebook and we'll now turn it back over to Erica.

1:40:25 Speaker 1: Thank you, Jackie, for that great presentation and demonstration. Next, I will do a summary of the most important points covered in today's session. In summary, flash drought is characterized by a rapid decline in plant available water that can significantly reduce crop productivity, even when precipitation deficits are relatively small. Now, root zone soil moisture, that's at the one meter depth, is a more meaningful indicator of crop water availability than surface soil moisture because it represents the water that is accessible to the roots, to the plant roots.

1:41:08 Speaker 1: So the Soil Water Deficit Index, or SWDI, standardizes the root soil and soil moisture, enabling then comparisons across regions and identifying periods of significant soil moisture stress. SIF provides a direct measure of photosynthetic activity, and it responds much more rapidly to vegetation stress than traditional greenness or vegetation indices such as NDVI. Now, combining SWDI and SIF-RCI provides complementary information on soil water availability and crop physiological response, which improves confidence in flash drought detection.

1:41:54 Speaker 1: A joint analysis of SWDI and SIF-RCI can provide earlier detection of agriculturally impactful flash droughts than the traditional drought monitoring approaches for many regions. The detection criteria should consider crop calendars so that drought is evaluated during periods when crops are actively growing and vulnerable to stress. The workflow that was demonstrated in the Jupiter notebook can be adapted to other regions, crop types, and growing seasons using local crop calendars and regional data.

1:42:31 Speaker 1: And finally, keep in mind that future improvements may incorporate things like machine learning, additional environmental variables, and other satellite observations to further improve flash drop prediction and early warning.

1:42:50 Speaker 1: Contact information, here you can find the email addresses for Nick Parrazu and Jackie Ryan should you have any additional questions about the material that was presented today. We've also shared our RSAT website page as well as our RSAT YouTube page. page or link. And if you have any questions or comments or you'd like to share how you've used our trainings in your work or for your studies, please send an email to the email address that you see here. And finally, to join our mailing list or stay up to date on our latest training, visit our contact page to subscribe.

1:43:33 Speaker 1: We've reached the end of today's session and this webinar series. I'd like to give thanks to Jackie Ryan for serving as guest instructor for the excellent presentation today, as well as for session one. I'd also like to thank Dr. Nick Parizou, as well as Karen Yuen, all from NASA's Jet Propulsion Laboratory OCO-2 team. We have collected your questions and we will begin the question and answer session shortly. Great, so you should be able to see now the Google Doc that has the questions that we've been collecting and our guest instructors have started answering them.

1:44:26 Speaker 1: So let's just start from the top down. with the first question. But however, before we begin, Jackie, would you like to remind participants about the issue that you addressed earlier? Yeah, so.

1:44:47 Speaker 3: Again, the thing about SWDI, as I showed it in the notebooks and the presentation, is that field capacity and wilt point came from the land model constants. But I'm working on updating the notebooks to derive these from the

1:45:08 Speaker 3: long-term percentiles of soil moisture.

1:45:15 Speaker 1: Okay, so let's start with the first question, which is a long one. So let's see. When SIF and soil moisture observations disagree, for example, soil moisture decreases but SIF does not show an immediate decline, how do researchers determine whether this represents an actual early stage flash drought, a delayed plant response, or uncertainty in the satellite observations?

1:45:42 Speaker 3: Yes, this is a great question. So, essentially. One of the common scenarios that we can see across a number of climates where you have higher precipitation and higher soil moisture is that soil moisture can decline without seeing a corresponding SIF RCI flash drought detection. So what can happen is that If you remember, I said that the roots on soil moisture is a supply side indicator. Vegetation is able to draw on a greater supply of moisture. And so it might not indicate stress even during high evaporative demand.

1:46:34 Speaker 3: So while you may see ESI or EDDI flash drought detections through those metrics, you might not see them in RZSM or in ZipRCI. And so the common case when SFAR-CI does not show flash drought, even when soil moisture is decreasing, is that the soil moisture available to the vegetation is not causing detectable vegetation stress.

1:47:14 Speaker 1: Okay, great. Thanks, Jackie. Let's move on to the second question.

1:47:22 Speaker 1: So let's skip the second question. We'll answer that in the Google Doc. The third question is, ESI and EDDI give somewhat false alarms. Why do we observe such data rather than getting more appropriate prediction-giving data parameters?

1:47:40 Speaker 2: So....

1:47:43 Speaker 3: The use of those two metrics is well established in the literature, and they are valuable. It's just that what we're arguing in this course is that soil moisture can

1:48:00 Speaker 3: tell us more about the agricultural impact. In practice, a drought monitoring program like the one operated by NOAA that I've linked below uses a mix of soil moisture vegetation, and ESI and ADDI.

1:48:21 Speaker 1: Great. And Dr. Nick Papazou is here as well. And I believe he'll add to this answer as well. Go ahead. Thanks, Erica. Yeah.

1:48:33 Speaker 2: I mean, that's already a great answer. I'll just add, there's a, if you recall from the part one, presentation I summarized a range of different types of droughts and so there's a number of applications and there's a number of drought indices and it's important to have as many indices as possible in general to get as much information as possible to kind of detect and monitor drought flash drought agricultural drought ecological drought hydrological drought the more information the better is the general approach to take.

1:49:12 Speaker 1: Four, what was the reason for the large error bars on some of the SIF points?

1:49:19 Speaker 3: Yeah, so I think this was in reference to the slides I had on that Yoshida et alia study looking at Russia. And one of the reasons is that the regions that they aggregated were fairly large. And then the other reason would be due to the retrieval uncertainty from the GOMS SIF that they used in that study. And we use OCO2 SIF in our Jupiter notebooks.

1:49:56 Speaker 2: Yeah, and my response is just reiterating that the SIF data is super noisy because it is a small signal. It's only 1% of the total radiance. And so it's a hard retrieval to make. And then the signal is also very dynamic. So it's changing constantly as light changes, as temperature changes, as the kind of physiological response to environmental conditions change. So that just creates a noisy, dynamic signal.

1:50:26 Speaker 1: Great. Thank you. The next question, number five, is the SIFTPP relationship linear at the leaf level?

1:50:36 Speaker 2: So I'll take this. It's actually highly nonlinear at the leaf level. And this is because of the light reactions of photosynthesis are constantly evolving with light conditions. And so you have something called non-photochemical quenching that kicks in when you have saturating light and high temperature. And so this regulates the amount of absorbed light that's used for photosynthesis. And so I provide this document here, the great reference for the scale dependence of the CIF-GPP relationship. But generally, as you go higher, or I guess up in scale, the more linear it becomes.

1:51:24 Speaker 1: Super. Thank you. Next question, number six. Why do we see high values of CIF and JJA instead of OG, MAM, or SOM? Is it because plants should directly use SIF in summer to lose excess energy and photosynthesis should bloom post-monsoon as we often see from NDVI?

1:51:48 Speaker 3: Yes, I tried to interpret this. And by the way, I think the OD there is just a typo.

1:51:52 Speaker 1: It's meant to be of.

1:51:56 Speaker 3: So, yeah, the relationship between the specific months when you see peak SIF and is going to depend on the area that you're looking at. For the Northern Great Plains, you know, we're looking at North America. We're going to see that there's maximum APAR or available photosynthetically active radiation in June, July, and August. And yeah, I'll let Nick handle the rest of this answer since he provided a more detailed answer.

1:52:37 Speaker 2: Yeah, my response is that SIF is a function of absorbed light. And so in general, if you have a healthy plant and more chlorophyll and leaf area, you're going to have higher SIF. And so that changes when you have physiological stress, as we have seen throughout the course. And so for the monsoon, I'm not sure if there's a seasonal dependence there, but in general, if you Go from a monsoon to a post-monsoon, I think you would see a reduction in cloud cover and an increase in soil moisture. And so both of those things would lead to a higher SIF signal.

1:53:15 Speaker 2: Of course, that depends on plant health and whether the soil is saturated or not and other factors.

1:53:24 Speaker 1: Great. Thank you. The next question, let's go to question number eight. While SIF correlates strongly with GPP, SIF signal changes are also influenced by canopy architecture and leaf area index during crop growth cycles. How reliably does SIF RCI distinguish between physical structural canopy changes, such as late season natural senescence, and true physiological drought stress?

1:53:54 Speaker 2: This is another really great question. You guys are killing it today. Um, so for, for natural vegetation where you have more or less predictable vegetation and phenology over time, um, SIF RCI is, is great. Um, you're, it's accounting for historical conditions as well as recent change. And so if you're able to account for that, um, with the SIF RCI is fairly robust to physiological change and structural change over the season. In cropping systems, there is more uncertainty because you may or may not have variability in planting date or crop type or management practice.

1:54:38 Speaker 2: And so to get the most possible robust information out of SIP RCI, you'll want to train it on historical data that has similar conditions to the year of interest. So if you're looking at a A cornfield in 2026, you want to make sure that you're looking at 20 years of corn with similar planting dates for that pickle.

1:55:06 Speaker 1: Super. Thank you. Let's go to question number 12. What are the major limitations of using SIF and satellite-derived soil moisture for early detection of agricultural flash drought, especially in heterogeneous agricultural landscapes?

1:55:25 Speaker 2: I think you actually know the answer to this question. So good job. So the satellite data is coarse relative to the size of the agricultural field. And so you're going to be mixing a lot of information into a single value. And so this washes out the field scale information that you're after. And so there's a lot of work to try to address this challenge. And I've listed a couple of examples here. One is to just avoid it and focus on uniform landscapes. Another is to apply fractional weighting. If you know the area of different crops within the satellite pixel, you can do your best to account for that.

1:56:09 Speaker 2: And that's actually quite effective as some more recent work has shown. And then another option is to use higher spatial resolution information from some, you know, either reflectance-based optical imagery that you can get from like Sentinel or commercial satellites to do downscaling. So you're taking the SIFT information at core scale and you're distributing that over space based on the reflectance information. And that has its own uncertainties, but it's another approach that people are taking.

1:56:46 Speaker 1: All right, great. Thank you. So let's go to question number 14. Is the SIFT data set available in Google Earth Engine or a way one can access it through GEE?

1:57:02 Speaker 2: Yeah.

1:57:02 Speaker 3: So the SIFT data sets that we've been using are not available on Google Earth Engine. Nick can correct me if I'm wrong on that. I.

1:57:15 Speaker 3: I don't use that tool, so I can't really speak to what the features of Google Earth Engine are, although I do know that the Tropomy SIF dataset is available as a stack catalog. So if GEE is compatible with that, that could work. Did we skip the livestock grazing question? Because I actually wanted to provide some additional insight on that. I know that Nick answered it as well. Sure.

1:57:47 Speaker 1: Yeah, we had skipped it, but yeah, let's definitely take that one. That would be number 13. Is there any study involving influencing factors related to excessive livestock grazing?

1:58:04 Speaker 2: Yeah, I've just noted one paper in here. I think there's a body of evidence outside the SIF literature showing this, but for SIF it's been very limited. But there is one study that looked at some rangelands in the southwest U.S.

1:58:22 Speaker 3: So what I wanted to add was when we looked at South Africa and Lesotho, one of the reasons that we saw that distinction that was even visible in reflectance across the political boundary was due specifically to livestock grazing practices. So in Lesotho, we could consider it to be an area that's more overgrazed compared with South Africa. And the soil moisture was higher on the Lucidia side of the border.

1:59:03 Speaker 1: Great. So we've reached the top of the hour. As you can see, there are many more questions that have not been answered right now or that we didn't get to, but we will be responding all questions in the doc and we will be posting this document on the training webpage. So that marks the end of this training, I'd like to thank our guest instructors, Dr. Nick Parisu, Jackie Ryan, and also very special thanks to Karen Yuen from JPL, as well as the RSAT team, Brock Blevins, Selwyn Hudson-Odoi, Jonathan O'Brien, Sherry Morris, and of course, all of the participants for their great interest in this topic and all of the great questions.

1:59:56 Speaker 1: We hope to bring another SIF training in the future, so please make sure you fill out the survey that will be sent out within the next day or two and provide your feedback about what you would like to see for future trainings related to SIF. Before I close, I'd like to give the floor to our guest instructors. Jati, would you like to say any words before closing?

2:00:28 Speaker 3: Yeah, I just wanted to say thank you all so much for attending. And I hope you found the course interesting and useful. And yeah, I'll turn it over to Nick.

2:00:42 Speaker 2: Yeah, I want to thank Erica and Brock and everyone for organizing this and Jackie for putting in a ton of time and effort to prepare the materials and then to all the participants for their really great questions today. We really appreciate it.

2:01:00 Speaker 1: Thank you. Thank you very much. And with that, then we have reached the end of this training. Thank you, everyone. And until next time, wishing you all a great day. Bye-bye.