Mapping Landslide Occurrence Using Earth Observations

Speaker 1: I am delighted to introduce the guest trainers for today's webinar, Dr. Robert Emberson and Dr. Pukkar Amatya. Robert serves as Associate Program Manager for NASA's Earth Applied Sciences Disasters Program Area. In this role, he helps administrate a range of projects funded by the Disasters Program, as well as supporting the disaster response activities for providing satellite data and insight to a range of stakeholders before, during, and after disasters strike. Alongside this programmatic role, Robert is also part of the NASA Landslides Research Team, where his current research focus is to better understand and model the exposure and risk faced by human settlement due to landsliding.

Speaker 1: The primary aim of his work is to better model the economic and social impacts of landslides globally and in near real time using satellite data. Pukar is an associate research scientist working with the landslides team in the Hydrological Sciences Laboratory at Goddard Space Flight Center. His main research is to develop open source systems that use machine learning and very high resolution imagery to develop automated open source landslide mapping algorithms to generate landslide inventories using optical and SAR imagery.

Speaker 1: He's also involved in the Committee on Earth Observing Satellites, Landslide Demonstrator, to highlight the use of commercial and space agency data for landslide applications. I'll now pass things over to Robert for today's presentation. Robert, over to you.

Speaker 2: So thank you, Sean, for the introduction. I'm excited to be here again to present part two of the training alongside my colleague, Bukhara Mathia, and talking much more today about how to map landslides using satellite data and specifically focusing on the sources of data information that NASA has made available and the tools that we've developed in-house. So let's start with first steps. So you'll remember from the training, the first part of this training, the type of landslides that we can think about with satellite data, the way in which we can think about the broad types of data involved in this.

Speaker 2: But let's say you have a landslide that has occurred in the area of your interest, really interested in determining details about that specific landslide. Maybe you want to know more about whether there are other landslides in the vicinity. So let's think about what might be necessary for mapping those landslides using satellite data. We typically talk about pre- and post-event imagery, basically as the type of data that's available before or after the triggering event that caused the landslide.

Speaker 2: So in terms of definitions, if you see that terminology used, typically we're thinking about imagery or data that brackets the timing of the triggering event for that landslide. So helpful to make sure that you have access to data from either side of this. It is still possible to map landslides with only post-event data, but again, pre-event is certainly helpful from a comparative perspective. And again, To really get into the detail for this, it's important to make sure that this kind of data is geo-referenced.

Speaker 2: So we know that aerial photography can be used widely, but again, if you want to make sure that it's able to map the location of these, you need to make sure that that information is geo-referenced. So you would also need a GIS software system to kind of capture this. And again, artillery data that we talked a little bit about last time, but things that can help you distinguish the location and potential source and run out zones of a landslide may be helpful as well. And we'll talk more about that in due course.

Speaker 2: So just quickly reiterate what we talked about in part one, what characterizes a landslide event? So these are often rapid and devastating downslope failures and starting of material and sediment that washes from a source point to a deposit or a run out point. And typically this leads to very highly visible surface changes. And so you can see in this picture here, the landslide which is running from the top left down to the bottom right into the river. There's a source near the road at the very top of that image where the red point is, and then the landslide has clearly run through another road and some population, some buildings and devastated those buildings where, and then it eventually runs out into the river at the bottom right here.

Speaker 2: So one thing that you might note is that these, the impacts to human systems are quite clear here. So the impact to roads and infrastructure. So that's one thing you would want to think about with respect to mapping. But the other thing here is that this is obviously a two-dimensional observation of what is ultimately a three-dimensional event. There's a gravitational movement downslope, and so we have to think a little bit about how we can convert our observational understanding from orbit, from two dimensions, into those three-dimensional kind of, how you can turn this kind of understanding from space into a 3D visualization you can consider in your head.

Speaker 2: How does that mapping operate? So in some cases, these kind of landslide failures are kind of obvious in some ways or you can kind of quite easily conceptualize the source and the run-up zones of these landslides. Here's an example from Big Sur in the west coast of California. You can see that the landslide has run from the upper part of this hill slope in this aerial photography and ended up as a large deposit in the ocean. So there's context clues here that you can use to inform your understanding of this system.

Speaker 2: The first, for example, is that the ocean here is obviously the lowest point in this landscape. And so the material source and the material deposit zone will certainly start from a higher point and then a lower point. And you can use that, obviously, the ocean as a very basic kind of context point here. And again, thinking about how you understand the landscape and where the kind of the lower areas are and higher areas can really help you inform the source and run out zones of a landslide. You can see here, again, highlighted those scar area and the deposit area.

Speaker 2: Another clue that you can see in this image here that gives an example of what's happening is there's certainly a large amount of material in the deposit area that looks a bit like the source, the kind of area around the landslide. In other words, that kind of darker brown area in the deposit mirrors the landscape a bit more than the kind of freshly scrubbed clean source area. So again, maybe we're looking for sort of areas in the deposit that look a little bit like the original kind of landscape.

Speaker 2: And so from, if we're thinking about that from an orbital perspective, you can see here, this is now from Google Earth imagery, the same landslide, you can see and begin to start to distinguish the kind of source and run out zones of this landslide. So these very large events are somewhat clear to understand from this kind of Earth observation imagery. You can see there's a very large change here between the color of the landscape originally, these green vegetated areas, and it's like, deposit and landslide area that's significantly changed.

Speaker 2: So that's certainly something you could look out for, and this would strike you and jump out at you from the page in some ways. However, there are certainly other types of landslides where the obvious signals, like those very large ones, may not be so clear. So here's, for example, a very large rockfall that's occurred. This is in Oregon in the United States. This massive rock has landed on this road. And this, again, is a type of landslide. In this case, it's a rockfall. there may already be an existing scar here, like an existing landslide deposit that has just been reactivated.

Speaker 2: And so in this case, the signal that you're going to look for from satellite observations is this potential impact on the road, this change. We would know, for example, that the human system here, in this case the road, we would not anticipate that there would be a very large boulder in that road. And so this initial change, this landscape change that we see here, would be quite visible in a pre and post imagery. You can really pick out that change to human systems. So again, these are the kind of visual clues you can look out for.

Speaker 2: And this is something to think about as we talk more about the more specific sources of information going forward. So in general, we think about landslides as events from above, from two-dimensional imagery. And we want to think about how to outline the area in which there has been a change. And so in other words, we're often thinking about creating polygons that outline the specific area of landslides. and capture the potential overall shift in activity. So here you can see an example of a landslide in British Columbia, and this image was something we used in the Part 1 training.

Speaker 2: But here we're looking, the red outline is a polygon outline used in a GIS software of this landslide event. And again, you can see at the very base of this landslide, where it's run out towards the river, there is a road, and the impact again on that road might be more visible. So creating these polygons is something that we can do using various GIS tools, including Esri ArcGIS, Google Earth, and other open source systems to create these shapes. And then we can merge those, if there's many, many landslides, we can merge those shapes into what we refer to as a file or a geodatabase.

Speaker 2: And that is then what we would potentially refer to as a landslide inventory, a data set of multiple different sets of landslides with the outline of those areas highlighted and the kind of the zones so the entire impact zone kind of connected both source and deposit. Now we'll talk a bit more about delineating between source and deposit in upcoming slides, but I just wanted to note that generally when we're capturing landslide mapping, we would consider both of the source and deposit areas and map them as a kind of entire polygon.

Speaker 2: In some cases, if it is such that the landslide has fully run out into a river or the bottom of the valley, and it's not entirely clear where that landslide fully separates from the surrounding rivers or something like that, you would probably want to draw a polygon that outlines that final deposit area, the end point of that deposit as the low point of the terrain, the river, the valley bottom, or the debris flow channel that might be connected there. So we obviously just talked about polygons, but it's important to note that many landslide inventories are also simplified to single points for each landslide, particularly for emergency response, and here I'm showing an image from the Hurricane Helene Landslide Observations Dashboard, which was a multi-agency effort to produce landslide and landslide location estimates during the aftermath of Hurricane Helene in last year in 2024 in North Carolina and surrounding states.

Speaker 2: It is much quicker to map the location with a single point rather than specifically and carefully drawing a polygon around that landslide. It can be more efficient for emergency response to do point data sets. And in essence, this is a single point of latitude and longitude to highlight the specific location of a landslide. And then these can be potentially used to inform further analysis and polygon mapping at a later date. So again, this is more often for convenience and speed. Typically for multiple different uses, we would probably prefer a polygon data set.

Speaker 2: It's more comprehensive. But again, for emergency response, this can be highly efficient and something we typically want to lead with for a very short-term response. As I mentioned before, landslides have a source area and a deposit area. Here in this very kind of simplified schematic on the top right, you can imagine at the top of a hill slope where the landslide source is from and where it runs out to the deposit area. Those are connected. We would connect those. We typically talk about the angle of reach or the kind of slope of the kind of in European terminology, the farboschung angle, connecting the source and deposit area.

Speaker 2: The area between those two locations, between source and deposit, is often damaged or even stripped bare by the landslide, the motion of that material downslope. And so we're thinking a lot about mapping the entire extent from the top of the source area to the bottom of the deposit area. But we would also want to think about differentiating these zones. And this is particularly important for something we'll talk a lot more about in the part three of this training, susceptibility and hazard analysis.

Speaker 2: So we'd also want to think about how we can connect those source and deposit zones and separately consider them. So there are certainly Simplified methods we can talk about like here where we sort of talk about the most uphill zone and the most downhill zone. We can sort of think about this as this image down below from New Zealand where you can see again the very top of that scar area. There's a clear line that defines the top of the scar area, the vegetation where the vegetation ceases and then the deposit.

Speaker 2: You can see that there's broad amounts of material that's been dumped into the stream And then the deposit in a broader area kind of around that you can you can see where there's vegetation still on top of a lot of that deposited material So there's complex clues here that you can use to sort of distinguish between those source and deposit, but these can be quite hard to generalize This is particularly challenging when we're talking about longer run-up areas. So here I'm showing an image from Google Earth from southern Taiwan and We'll refer to southern Taiwan a few times in the course of this training because it has some really useful and illustrative examples.

Speaker 2: But here you can see these very large landslide scars at the top of this hill slope, and I'm just going to zoom in on those a little bit. You can see that there is these large blobs of source area at the top of the hill slopes, and then there's these long run-out channels where the landslide material has been channeled through the valley and ends up in this like large river channel. And it can be quite challenging to distinguish where the landslide itself actually ends. We typically would not map the runout itself into those river channels.

Speaker 2: So the entire river channel you can see running from sort of the center of the image all the way to the right hand side. We would typically not consider that to be part of the landslide. But again, if I asked you to kind of distinguish where does this landslide begin and end, that can be quite challenging. And so oftentimes, we're going to think that we can, to generalize this process, we would perhaps want to think about some simplified geometric rules. And so this is a rough estimate that has been demonstrated by a number of studies that the perimeter and the area of a landslide can be tools with which we can estimate the kind of aspect ratio of a landslide and then determine the rough estimate for a scar.

Speaker 2: the relative area, I'm sorry, for the scar, and then using the scar area, we can kind of estimate what proportion of the overall map landslide is made up of a scar. So I'm going to go over that briefly here. So what we would do is we'd calculate using a mapped polygon. We'd calculate the breadth of the landslide. We would also calculate the area of that landslide. And then using the formula below, we can estimate the associated width of a landslide. And we would use the area and perimeter to calculate an equivalent ellipse and the aspect ratio, which is K here.

Speaker 2: And then again, using this formula, we'd roughly estimate the associated width of a landslide. And then we know from global data sets that the scar area proportional, that the scar area is approximately 1.5 times the width squared. And so once we know that scar area, we can take the overall area that we mapped of a landslide and take the kind of scar area estimates. And using that, we can kind of estimate an approximate relative proportion of the landslide. And in this case, we would probably assume that the higher part of the landscape, the most upslope area is the scar area.

Speaker 2: And that can be a way to approximate this over large landslide populations. So I'm happy to provide more information. More information is available in the resources for this presentation about that specific formulation. But again, this is an approximation that you can use for large landslide populations to estimate the source and relative source area for landslides. or scar area for a large population of landslides.

Speaker 2: Another important tool in our belt, I guess, for mapping landslides is false color imagery. So this is some imagery that you actually have seen from the intro material for this overall training. Again, this is a large catchment in southern Taiwan, which was devastated by landslides as a result of Hurricane Morakot in 2009. What you're seeing here is Vegetation in red and gray tends to be bedrock or channel fill. This is using NDVI band combination or normalized difference vegetation index. Details on how to create that kind of NDVI system, again, are available in the resources for this training.

Speaker 2: And this is a very helpful tool because changes in vegetation, particularly in tropical areas where there's extensive vegetation or forest, is often much more sensitive to surface change as a result of landsliding than any other kind of combination of bands. And this can be particularly important in areas where we have dense vegetation and we can pick out landslides very carefully. In this case, you can see that there's extremely large differences because of the landsliding. These gray areas get much more pronounced in this catchment.

Speaker 2: But you can also see, obviously, that the river itself has a large amount more sediment in it as a result of upstream landslides. So that can be something else you can pick out using this methodology. I would also note from this image that one of the important elements that you should pay attention to is cloud cover in the bottom-most image here. In fact, in both images, there's clouds around the top of these mountains. That's something you won't be able to see around. And again, we talked about that in part one of the training, that cloud cover is an important consideration for optical imagery use.

Speaker 2: Another characteristic of the imagery you can use to map landslides is the resolution of the imagery. We talked a bit about this in the first part of the training, but in more detail now. The number and size of landslides that you can observe is very dependent upon the resolution of the imagery that you're using. And that should make kind of common sense, but to illustrate this, using this figure here, we're looking at landslide inventories from a range of different locations around the world, mapped using a range of different imageries.

Speaker 2: And again, these different image types have a different set of resolution. So what you're seeing on the x-axis is the area of landslides and then on the y-axis is the probability density of the overall population. So essentially as you go higher up on that y-axis, these are more common in the overall inventory. That size of landslide is more common in the overall inventory. So the key thing to look at here is these populations at very small areas. And in this case you can see there's a handful of those inventories, these individual lines, that were mapped using very high resolution imagery.

Speaker 2: And that could be up to three meter resolution imagery. And in this case, you can see there are a large number of landslides in those inventories that are around about 10, I guess, 100 square meters or even smaller. Those are really small landslides. And so, or at least from an optical observation imagery perspective, those are small landslides. And so this high resolution imagery certainly means that we can see a large proportion of those small landslides. And then when you have access only to the kind of course resolution imagery, particularly you can see some of these points where what we refer to as the rollover in these distributions is a much higher area.

Speaker 2: You're capturing a lot fewer landslides at those small numbers. And so your population, your overall inventory is going to be made up of much larger landslides only. So this has been a big change in recent years as we have The community has had access to a wider range of high resolution commercial data, particularly that has enabled this kind of data set mapping. A lot of this through NASA has been available through the Commercial Satellite Data Acquisition Program, and it's enabled a much higher effective resolution mapping of these small landslides.

Speaker 2: Again, as I mentioned before, cloud cover is a really significant factor for optical imagery. Here's just an image, this is just a nice image of different types of clouds over Northern Australia using Landsat. You can see that there's small puffy clouds at low level and then higher level, you've got these larger cirrus forms. Each of these can kind of definitely, I guess, mask the areas where landslides have occurred or could occur. It's certainly the preference to seek cloud-free imagery for mapping where possible.

Speaker 2: When producing an inventory of landslides, particularly for emergency response purposes, it's important to provide polygon outlines of the cloud locations in the associated inventory. In other words, you want to provide a separate set of shapefile that captures where those clouds are in the imagery used to map this, so that anyone using that data set can then go back and say, ah, I understand this area was masked by clouds, so there could still be landslides here. And it makes sure that anyone using this inventory for further analysis does not falsely assume that those areas may be landslide-free.

Speaker 2: It's only that they were masked by cloud. So again, capturing where cloud cover is is an essential part of building a. And again, this is one reason why we would strongly suggest to seek cloud-free imagery where possible, because then it avoids having to have a sort of separate shapefile of clouds within your overall landslide inventory. So when producing... a landslide inventory, we'd want to think a lot about the type of data to go alongside this. So for each individual landslide, you would want to think about the types of information available.

Speaker 2: And then for a broader inventory, similarly, a number of different data sets, a number of different characteristics are relevant to capture. And this would be what we refer to as metadata. So we would certainly want to think about the location. That can be either obviously the individual landslide locations, either as polygons or points, but also more generally providing that kind of contextual information about the overall landscape, if you have that. That can be provided as comments in the data package that's part of the inventory.

Speaker 2: We'd also want to think about the potential triggering event. That can be very easy in some cases to point to a specific earthquake or seismic event. Or in some cases, you would think about, you would know that this was triggered, these landslides were triggered by a significant rainfall event. So providing that information would be helpful. We want to provide the imagery date. So certainly to ensure that we have informed the user of this dataset what the date of the imagery used for mapping was, so when that image was collected.

Speaker 2: Particularly, this is important if we're using pre- and post-imagery, so that we can determine whether the landslide inventory includes just the landslides triggered by a specific event bracketed by those imageries, or whether that includes historic events prior to that as well. We'd want to include information about the source, the type of sensor, and resolution. So you could say Landsat, operational LAD imager, resolution 30 meters, for example, or for a variety of different sensors, depending on what was involved.

Speaker 2: You'd also want to include information about the type of mapping that was used, whether that's points or polygons, whether it uses a manual method or an automatic method, which is something we'll talk more about in a few slides time. If it's produced using a manual method, we'd also want to give information about who that person mapping is. Again, all of that information is so that this data set is reproducible or as best as possible reproducible. Typically, we want to talk about any associated publications or a DOI that attaches to this information.

Speaker 2: Again, the extent of the mapping area, as we mentioned, talking about cloud coverage and including that information in the metadata is essential. And if possible, the type of landslide for the individual events. should be recorded. And this is often tricky if we don't know exactly what kind of information this is, and field validation is certainly relevant. But if we're typically mapping individual small landslide events or shallow landslides, we might want to flag that that's what was mapped here.

Speaker 2: If we're thinking more about deep-seated landslides, as we mentioned in part one, we'd also want to flag that those are the type of landslides that are considered. So again, all that kind of information should be considered and included in a metadata package. I mean, that can be part of a geodatabase, or additional notes or documentation alongside the shapefile that's produced. As mentioned, we can think a bit about new or historic landslides. So after an initial landslide, the deposited area and the scar area can certainly be revegetated.

Speaker 2: Plants can grow back and fill that area so that it's much harder to spot where that was. So again, we're returning to that location in southern Taiwan. You saw the image on the right-hand side here before, but when that catchment was fresh from this incredible rainfall event in 2009, there was a lot more obvious landslide scars in that area. And so these are the same location, but one image is taken in 2011 and one image is taken in 2024. And you can see that a significant part of that area, especially with the smaller landslides, has been revegetated significantly over time.

Speaker 2: These older landslides may be invisible as scars are through vegetation. And so ultimately, we want to, this is why it's important to flag the timing of the imagery and critically to get imagery, if possible, as close to the timing of the landslide event as possible to ensure that we capture the full extent of those areas. One thing that's also obvious here is that those larger landslides here are still vegetation-free, and that may be because they've been reactivated or kind of the remobilization of material or loose scar material may be continuing as rainfall continues over the last 15 years.

Speaker 2: And so that's something to also pay attention to, again, that like larger landslides may continue to be obvious in the landscape for longer than smaller landscape landslides. One thing that we can think about more about, we'll talk more about this with automatic mapping, is that multi-temporal landslide inventories can certainly improve the kind of longer-term analysis of these kind of areas. So in other words, we can build a landslide inventory not just once, but over many years, like once a year, maybe even once a month, potentially, if you have access to that kind of information, to assess the changes over time and whether there's been reactivation, new landslides, changes to existing landslides, or essentially if only small landslides have occurred.

Speaker 2: So all of these kind of contextual clues and the timing of these events are critical to consider when building an inventory. And again, you've seen a number of examples now of manually, of Lantivites and how you can distinguish them using manual mapping. We're gonna talk now about some of the tools used to map Lantivites automatically. So the next section, we're gonna kind of drill into what the critical elements to consider are and how we can accomplish those using algorithms and automated methods to build these inventories.

Speaker 2: So why do we care about automatic mapping? So far I've shown you a number of images over the first part of this training as well as the first section of today's training that consider essentially one or two or maybe even a handful of landslides. But in some cases where there is a large earthquake or a major rainfall event, many thousands of landslides could be triggered by a single trigger. so we'd end up on a landscape that looks a bit like this this is again the same catchment in Taiwan and it would be impractical to ask an individual to map these using hand mapping or kind of clicking to create the polygons this would be an enormous task in some places this has been done and I myself have been part of you know several efforts to map thousands of landslides but this is not time efficient and particularly if we're doing something that requires rapid response and that could include emergency response for disaster analysis.

Speaker 2: One example of that would be NASA's disasters program which supports landslide mapping as part of our response for disasters, particularly in the US but also abroad. We want to use automatic methods to speed up this process as well as provide a more kind of reproducible and robust and I guess objective analysis of these areas to build a more systematic landslide catalog. So typically, automatic methods will exploit the changes in texture, color, and spectral properties of the recently disturbed areas.

Speaker 2: So what that means is the change between an image before and after can be looked, you can look at the change in the kind of pixels or the areas to assess those characteristics and how the landslides have introduced that. Typically, there are two methods, either pixel-based or object-based, and we'll talk about both of those. One reason why this, however, is One critical element we want to address today is amalgamation as well. I'll talk more about that in a few slides time. But again, automatic methods can be highly productive to do this kind of work.

Speaker 2: And a real advance has been made in the Landsat community in recent years. There are certainly caveats, but we'll talk more about that in the next few slides. So to drill into this in more detail, let's talk about how this kind of process works. So let's give a very broad overview before I hand it over to my colleague, Bukhara Mathia. But typically, we want to think about a computer is going to, an algorithm will look at a way to characterize locations that look like landslides or don't look like landslides based on information that has been made available.

Speaker 2: And typically, that's the training information. So in other words, we would often have a small subset of landslides that are manually mapped as an input to a model. Those would then be, an algorithm would assess the characteristics of those landslide locations and the areas without landslide locations. and then compare that with the larger, broader analysis area that's to be mapped, and pull out the areas of that kind of broader parity map that will produce an estimate of where landslides are likely or not.

Speaker 2: So one example of this is using random forest models where an initial data set is subsampled, decision trees are used to assess the likelihood, either a yes or no, of landsat location. That's done over a large number of subsamples. And then this kind of multiple set of trees is produced in the majority voting estimate to produce a final prediction. So this is just an example of one of many different types of machine learning or statistical methods used to do this. And typically we were, again, we're going to try to characterize between landslide and non-landslide zones based on this training information.

Speaker 2: And we have a wide variety and I think we will not have time today to talk about all these different sorts of forms of analysis. But again, statistical regression, machine learning or AI, they have widely been used recently. And so the outputs of all of these, however, is ultimately a prediction of likely landslide locations. One of the most important things to remember here is that the outputs of any statistical model are only as good as the inputs. And so ultimately the training data that's used is going to be one of the most important factors that determines the success or failure of an automatic model.

Speaker 2: And so still, this is why we have to stress the importance of manual mapping and manual inputs, because ultimately a human eye will still be one of the most important elements to build those training data sets in most cases. So I'm now going to pause and pass over to my colleague, Pukar Amatya, who will take over to talk a little bit about some of the automatic methods that have been developed by NASA scientists.

Speaker 3: Thanks, Robert, for the great overview of landslide mapping. Hi, everyone. My name is Pukar Amatya.

Speaker 4: I mostly work towards development of automated landslide mapping system. So right now, I'm going to talk about automatic mapping techniques to map landslides using optical imagery.

Speaker 1: So remote sensing of.

Speaker 4: Landslides can be categorized into two groups, pixel-based and object-based image analysis, or OBIA. The pixel-based method used the spectral signature of single pixels and have been successfully used to map landslides. However, this method suffers from the salt and pepper effect due to single pixels being demarcated as landslides. On the other hand, in OBIA, image segmentation is used to convert homogeneous pixels into objects. So an object in literal sense is a polygon made up of similar pixels.

Speaker 4: Spectral, textural, morphological, geometrical, and contextual characteristics of these objects are then utilized to detect landslides. So comparative studies have found OBIA to detect landslides accurately with fewer false positives. One caveat is it is computationally expensive in comparison to the pixel-based methods. Due to the reason mentioned in previous slide at NASA, we have been utilizing object-based image analysis-based methods. Landsat mapping using OBIA consists of two important steps, segmentation and classification.

Speaker 4: In segmentation step, segmentation algorithms such as multi-resolution segmentation, mean shift segmentation, water set segmenting, etc. are used to group homogeneous pixels into objects. In Figure B, you can see objects being created using segmentation. The main challenges in implementing a segmentation algorithm is determining the segmentation parameters. These segmentation parameters are very important as it defines the size of the objects overall, influencing the accuracy of the entries produced.

Speaker 4: So these parameters are often set using iterative trial and error methods or automatic parameter calculation methods. So the second step is classification.

Speaker 1: Most of the studies.

Speaker 4: Utilizing OBIA for landslide detection involves creation of a knowledge-based classification rule set for classifying segmented objects to landslides. This machine executable rule set is created based on site-specific characteristics and manual thresholding of landslide diagnostic features such as NDVI and slope. Alternatively, machine learning methods such as random forest and support vector machine spectra have been successfully utilized for this step as well. Although OBIA has been used to map landslides all around the world, the accessibility of the available algorithms is an issue because most of the studies have relied on commercial softwares.

Speaker 4: We at NASA developed a Python-based open source semi-automatic landslide detection system called Sallet, which is easily deployable in both Linux and Windows operating system. Sallet combines various Python packages to create a system capable of mapping landslides using OBIA and machine learning. The system when provided optical imagery and DEM uses means of segmentation to generate objects So object means of spectral, topographical, and textural landslide diagnostic features, such as NDVI, slope, brightness, GLCM texture, etc., are then calculated and used for landslide mapping.

Speaker 4: As this is a machine learning-based method, a manual landslide inventory is used to train a random forest model, and objects are classified into landslide or non-landslide classes. Finally, objects belonging to LandslideCast are dissolved to create a first-pass landslide inventory. We then run a quick visual check to remove some obvious false positives, such as river sand, barren areas, roads, et cetera, to obtain the final landslide inventory. The code for SALAD is open source, and if anybody's interested, you can get it off of NASA's GitHub.

Speaker 4: The link is also provided in this slide. So to put into visual perspective of what I just described, I will use a test case that was used in the original SALAD paper. So during development, SALAD was tested in a 125-kilometer square area along the Pasang Lamu Highway in Nepal, where a large number of landslides were induced by the 2015 Gorkha earthquake using a rapid eye imagery. So we divided the study area into training and validation area. Within the rapid eye imagery itself, 623 landslides were manually mapped.

Speaker 4: Out of that, 148 landslides in the yellow polygons were used to train the random forest model, and the remaining 475 landslides were used to validate the results of landslide inventories generated automatically using silence.

Speaker 4: So the figure on the top shows the segmented objects created from the rapid eye imagery using mean shift segmentation. We can see the landslide objects were clearly separated from other objects. The segmentation, however, did result in some over segmentation. Although the purpose of segmentation is to create meaningful objects, over segmentation is not necessarily a bad result. Generally over segmentation is preferred to under segmentation because merging is possible in final steps. So the bottom figure shows a zoomed in view of SALAD detected landslides.

Speaker 4: So comparison with manual inventory for validation highlighted SALAD was successful in detecting 70% of the area of manually mapped landslides and 20% were false positives present in the detected landslides. A thing to note here is some small amount of manual corrections will be needed to generate a complete inventory if failure generated landslides or any automated method outputs are to be used for detailed landslide hazard modeling and risk analysis studies. So once you have a trained model, then you can start doing large scale mapping.

Speaker 4: Video on the left shows the yearly distribution of landslides along two major highways in Nepal. The thing to note here is the sheer number of landslides that were produced by the.

Speaker 3: 2015 Gotha earthquake.

Speaker 4: One can note, as we are not using a change detection-based approach, we map every landslide present in a particular year's tile. This type of mapping, however, enables us to visualize evolution of landslides through time. The video on the right shows the evolution of a landslide that led to a catastrophic failure widely known as the Jure landslide in Orniko Highway of Nepal. You can see that landslide started developing in 2012 and failed in 2014, creating a flood hazard downstream by blocking the river.

Speaker 4: And since then, it has started to dwindle in size. Here's another example of a multi-temporal inventory that was developed using SALAD to support risk assessment study for upcoming hydropowers in the Orange River Basin in Nepal. So, although SALAD had an accuracy of about 70%, as mentioned before, the speed with which it mapped landslides over large areas of considerable advantage. For example, we spent about six hours manually mapping landslides induced by 2015 Golkha earthquake for a single rapid eye tile.

Speaker 4: With about one hour of mapping landslides for training purposes, we were able to map same landslides using SALAD in about 45 minutes, which is a significant amount of time gained and this capability might be very valuable for supporting rapid response efforts during disaster events. So as mentioned earlier, inputs to SALAD are a post-event image and a DEM. As it only uses the post-event image, it does not have that ability to separate all landslides from new event-induced ones. Sometimes when we need to do mapping to support rapid response, it is very important to remap landslides induced by the event only.

Speaker 4: A change detection-based approach utilizing pre- and post-imageries can help us achieve that. Hence, we introduced a scene detection-based module to the SALAD framework. We call this new system SALAD CD. The new framework incorporates image normalization, image co-registration, and instead of using single image landslide features, we use scene detection-based matrices, such as NDVI difference, principal component analysis, and independent component analysis, which is basically a dimensionality reduction method capable of capturing change-related information in smaller number of bands.

Speaker 4: from a stack of pre- and post-event satellite imagery.

Speaker 3: So in the figure, you.

Speaker 4: Can see individually mapped landslides that were induced by Hurricane Eta and Iota in 2020 in South America using SALUS-3D.

Speaker 4: So recently, deep learning-based methods have been found to map landslides more accurately and faster than OBIA or pixel-based methods. So deep learning takes in an image and applies several mathematic operations in particular filters to identify objects in an image separating one from the other. However, the prerequisite to run a deep learning based model is high quality large training data set, which might not be readily available to support disaster response. Fortunately, in recent times, there are multiple community reference data sets open source.

Speaker 4: which can be used to train the deep learning based model and establish a pre-trained file. So a pre-trained deep learning model once established can easily and effectively be transferred or used on similar landslide mapping tasks. So our test here at NASA has verified deep learning to map landslides faster than OBIA method and is something we are pursuing further for the near future. So just to summarize which method to use really depends on individual preference and mapping conditions. If one has a pre-trained deep learning model or have a large enough training data, I would suggest using deep learning methods.

Speaker 4: If one has small data set and training needs to be done from scratch and relatively quickly, we can use OBIA. And if one just needs to have an idea, like basic idea of like how large of a landslide event it is, one can relatively quickly and easily run a pixel-based method. So, thank you. With this, I would like to hand it over to Robert again to continue on with the presentation.

Speaker 2: Thank you so much, Prakash. Very interesting presentation. Some of the exciting work has been done from the landslides research at NASA. So, just to note a couple of additional considerations with respect to automatic mapping. So again, as we mentioned, the training data strongly influences the outputs. The method that Prakhar described, SALAD, has been used also to map flood damage, relying on change detection. And ultimately, this is looking at specific information about where rivers have changed or the damage as a result of high floodwaters have occurred.

Speaker 2: In some places, that can look similar to landflights, but ultimately not exactly identical. And so again, whatever you feed into the model is going to strongly influence the output therapy. So in this case, for example, using DEM data, the topography or the slope, for example, could be used as an underlying layer to like distinguish between areas that are, for example, floodwater impacts or landslides on hill slopes, ultimately because we understand that there's a difference between high slopes and landslides occurring on slopes and impacts occurring in river channels as well.

Speaker 2: So again, adding in that kind of ancillary data to the training analysis tools is certainly beneficial in some of these analyses.

Speaker 2: One other note is that areas where there has not been a significant change in the textural color or spectrographic analyses or observations, detection of landslides may be difficult. And this is particularly important in desert regions or areas where landslides have been reactivated. In those places, the bedrock color or the subsurface color or the soil color may look very similar to the sort of surface color that you would see from over there. And so ultimately, there may be very limited amounts of change that could be detected.

Speaker 2: And in that case, we may need to use non-optical methods to do this, and we'll talk more about that in a couple of slides. One other critical element to consider is amalgamation. So I'm showing two images here of a small area of part of China that was impacted by the Wenchuan earthquake. These are the same zones, and on the left-hand side is using 15-meter aster imagery. The landslides were mapped automatically using a variety of different analysis tools, you can see that these polygons are produced are relatively coarse and they kind of align around these zones of change, including landslides.

Speaker 2: On the right hand side, this is 2.5 meter spot imagery of the same area, but now you can see that there's a lot of detail within the landslide areas that's highly complex, and in some cases, specific landslide bodies that we've mapped as individual large bodies of landslides, single landslides, may in fact represent multiple small landslides sort of merged together by the kind of course analysis. And this is a real challenge, particularly if we want to do something with that landslide inventory for or sustainability modeling, or particularly run-out modeling.

Speaker 2: The main reason why this is the case is because the amount of material produced by a landslide typically scales with the size of the scar. So the larger the scar, the more material is moved. But this is a nonlinear scaling. The volume of material produced is certainly scales in a, as I say, nonlinear fashion with the scar area. And what that means is if we misidentify the overall scar area, the errors in our volume estimates are even significantly and proportionally larger. So this amalgamation is really important.

Speaker 2: And it's one of the reasons why high resolution imagery is certainly better. And in any automatic method, it's also important to look at again at these automatic polygons or the data that's produced and identify, for example, if any of them may be amalgamated landslides. A human identification of some of these events would focus on potentially, let's say, these larger blob of landslides that have been highlighted. breaking them apart into smaller individual areas and using context clues to determine whether these maybe are multiple different events all merged together.

Speaker 2: And this ultimately produced a more robust and reliable landslide inventory for further analysis.

Speaker 2: So now, as we mentioned, we're going to talk a little bit more about non-optical detection of landslides. So there are certainly strong cases to be made for the use of optical data, and this has formed the bulk of landslide science and mapping for many, many years. But there are multiple different sources of information available from orbit, and some of them have been used effectively for landslide mapping in recent years. And we want to touch on some of those today to give a brief overview of how they operate.

Speaker 2: So one key element is the use of SAR data. SAR stands for synthetic aperture radar, and this is using radar data to image the surface of the Earth. That sounds somewhat simple in principle, but there are many factors that influence the way in which landslides are detected. This schematic here shows roughly what's going on. A satellite flies along a specific flight path and puts radar data, points radar data out at the surface of the Earth and then detects the reflection of that data, of those radio waves back at the satellite itself.

Speaker 2: The satellite records both the amplitude, the frequency and phase of that information coming back. So there's lots of critical information captured within that that can be used to assess the surface characteristics of the Earth as it flies below the satellite. However, there are several factors that influence the way in which a satellite can observe. Certainly mountains areas, which are really critical for landslides. Three of those key elements include shadow. So in other words, areas behind a mountain as seen by a satellite orbiting at a shallow angle are invisible.

Speaker 2: In other words, anything that's behind the mountain itself is not something we can see. So landslides on the far side of a mountain like that would be invisible. There is also an effect called foreshortening. And this is because of the relative angle of slopes facing a satellite versus slopes facing away from a satellite. Sensor facing slopes tend to look smaller than the ones facing away from a slope. So that's something else to consider. The relative can definitely be corrected for. And then finally, layover is an effect whereby steep slopes may appear as being flipped over.

Speaker 2: In essence, the point at the bottom of a slope versus the top of a slope may be inverted in the observation that returns to a satellite. And so again, correcting for that is something that's critical when considering the landslides in these mountainous areas. The other thing that matters is the type of surface for SAR data. So the reflection of the surface Radar data is going to depend significantly on how smooth or how rough the surface is, particularly if the wavelength of the roughness is close to the wavelength of the sensor or within those kind of brackets.

Speaker 2: So very smooth surfaces like water, for example, almost give no reflections. We see that almost as dark in satellite and star imagery. Rougher surfaces produce more diffuse scattering, which potentially leads to wide variety of changes to the reflection. So again, landslides may not just affect the overall sort of shape of the landscape, but also the roughness of the surface. You can imagine a bedrock surface may be significantly either scoured or kind of these plummetly textures of landslides to have very kind of rough areas.

Speaker 2: And again, this is going to strongly influence the way in which satellites and radar later is able to observe those systems. There are two key approaches to assessing landslides using SAR data. Those are coherence-based approaches and amplitude-based approaches. Both types of analysis use pre- and post-imagery to compare. And so, in other words, you're thinking about the difference between an initial image and an aftermath image of a landslide triggering event. Typically, an amplitude-based assessment records the change in the intensity of the signal return.

Speaker 2: These typically have lower sensitivity, but more recently, and recent studies have been used to assess changes, large changes in volume of LASA material, so dense volumetric changes of LASA. I'll go over that in a second. And then coherence-based approaches are changes that affect the signal of the radar return, the signal frequency, and how much of that is incoherence or out of coherence, if you like, with the initial radar signal. This is highly dependent upon various factors, including atmospheric influence.

Speaker 2: And so an associated change, a significant effect of surface textural change is what we're looking for here and the scattering properties that result. This has been used in some cases quite effectively for landslide detection in some settings. So to briefly introduce non-optical threat, the amplitude-based assessment, several teams at NASA, including some supported by the Disasters Program, have used amplitude-based changes, including using Google Earth Engine to assess dense areas of landslides.

Speaker 2: So on the right-hand side, we're looking at what we refer to as a rapid response heat map. The black lines are showing the polygon outlines of landslides that have been mapped using optical imagery on the aftermath. But the low to high density color scheme here indicates areas where the star data has indicated a large amount of change potentially in the amplitude, and so we're potentially seeing a lot of landslide effects. And in this case, those areas align very closely with the areas of where lenses were eventually mapped afterwards.

Speaker 2: And this has been used effectively to see through clouds and to look at areas where there's been significant cloud cover afterwards. A critical example was in the 2021 Haiti earthquake and shortly thereafter there was a major rainfall event. This method was used rapidly afterwards to assess changes in landslide density shortly after that triggering event sequence. You can see here On the left-hand side, an image of SALAD-mapped landslides. I'm using the same method that my colleague Pukkar spoke about just a second ago, looking at the kind of yellow polygons of the landslides mapped in the center of the image.

Speaker 2: You can also see the number of clouds that were evident here. And this is an area that's highly cloud-covered. And often, although we could see some areas with optical imagery for mapping, not all were immediately visible. And so SAR data was used effectively to map in the very short aftermath of the triggering events to look at locations where there were significant numbers of landslides. And this was an important tool to, number one, to highlight the locations of potential risk, but also to flag areas for further mapping as cloud-free imagery became available.

Speaker 2: So that was a bit about amplitude-based assessment, but we also have coherence-based approaches. This typically looks at the change in coherence between two site images. This has been used, as I mentioned, for landslide mapping. This example from Myanmar and a recent study by Modini and some co-authors shows this big change in coherence, this bright spot in the cell image in the middle of the data set that aligns with the location of this landslide, as you can see on the left. This can be a very useful tool for the assessment of landslides, but it is very dependent upon a number of factors.

Speaker 2: and areas prone to de-correlation, so where the imagery, the return dataset does not align coherence with the output image, with the source imagery. Particularly forests, coherence change may have a very low detection ability. So in this case, this works, but it's important to note that this dataset and analysis approach is not always robust in every setting. So lastly, a word about InSAR. So this is interferometric SAR. This uses two or more images to reveal change in surface motion. So in this kind of schematic here we're looking at two images of two star images collected at different times and the difference between them is used to assess what's referred to as an interferogram and this builds sort of very small you can look at very small gradual changes in motion or change in surface.

Speaker 2: This is highly effective in some settings, but it does require surfaces that are typically coherent. In other words, ones that are consistent between each image. And so we'd want to find specific points that are unchanged on the surface of a landslide. And so this typically works best for slow moving landslides. Ones where we might be able to find a specific reflection point that is consistent from one point to another, and you can assess the small amount of movement between those, the before and after images.

Speaker 2: So here's an example from California. Sorry, this is in Nepal, but the team is based in California, my apologies. There are a number of landslides. You can observe the velocity per year, for example, of the individual landslide bodies. And over the course of a large landscape, you can assess areas of motion, and those would be used to highlight potential slow-moving landslides that may not be visible from any other source of data, even from the ground surface where these small changes maybe non-evident.

Speaker 2: And so again, this can be an important way to highlight potential risks in the future or even elements that are currently at risk for infrastructure right now. Again, all of the consistent sort of caveats that apply with SAR data are relevant here as well. The same kind of a complex analysis is required here. There is a number of resources available in this training and the resources material that can be used to guide the way in which SAR data can be used. And this is one of those tools that has been highly effective in some settings, but again, there are many systems that is not always the same.

Speaker 2: This is not always as applicable. However, it is important to note that optical and SAR imagery do not have to be used exclusively from one another. Particularly for large landslide events, it can be highly beneficial to use optical and SAR images alongside one another to help build both SAR and optical assessments. And this kind of multi-sensor fusion is perhaps one of the most important and exciting potential further analyses that we can think about as the future of landslide science. So I want to say thank you so much for your attention today.

Speaker 2: And I want to lay the seeds for what's going to come in the final part of this training. Using all the information we've kind of built so far, we're going to talk a bit more about building landslide susceptibility models and how that informs hazard assessments. And then we're going to put all those pieces together to build an understanding of how hazard exposure and impact can all build a much more comprehensive understanding of how landslides and disasters fit together. So again, thank you so much for your time today.

Speaker 2: I'm looking forward to taking any questions. Thank you to my colleague Pukar and to the RSET team for your support. So over to you, Sean.

Speaker 1: Wonderful. Thank you so much, Robert and Pukar. The following are links to resources related to today's training, including the GitHub repository for the Semi-Automatic Landslide Detection, or SALAD, NASA Landslides Research, as well as relevant RSET trainings. In summarizing the concepts covered in part two of the webinar series, An ability to translate two-dimensional observations from orbit into situational awareness on the ground is a critical skill to develop for landslide mapping. Differentiating the source area or headscarp of a landslide with the runout and deposit areas is important for comparative analysis of susceptibility and hazard.

Speaker 1: The number and size of landslides that can be observed strongly depends on the resolution of available imagery. Relevant metadata is needed when mapping landslides. multi-temporal landslide inventories can provide longer-term analysis. Automatic methods for mapping landslides, such as pixel-based and object-based, exploit the typical changes in texture, color, and spectral properties of recently disturbed areas. Automatic methods use a variety of regression, machine learning, or AI-based approaches that can be used to predict other pixels or objects that match the type of input used as training, which is then and output as a prediction of likely landslide locations.

Speaker 1: Training data strongly influences outputs. And SAR-based landslide mapping uses two key approaches, coherence-based and amplitude-based.

Speaker 1: Before we transition to the Q & A session, I want to remind you there will be one homework assignment which you will be able to access from the training page on March 18th. Answers must be submitted by Google Form with a due date of April 1st. To receive a certificate of completion, you must attend all three live webinars and complete the homework assignment by the deadline. You will receive a certificate via email approximately two months after the completion of the course. Below is the contact information for Dr. Emerson and Dr. Amatya, along with links to the RSET website and social media.

Speaker 1: We will now transition to the question and answer portion of today's training. Please enter your questions in the question and answer box and we will get to them in the order that they were received. We will post the question and answer documents to the training web page within a week of today's training session. And once again, Dr. Emerson and Dr. Amatya, thank you so much for the wonderful presentation. We can jump right into it. We've gotten some really great questions so far. So we do have about 30 minutes left.

Speaker 1: So if you have not submitted a question, please do so. And we will try to get through as many of these questions as we can in the next 28 minutes or so. So jumping right in, question number one, is there any way to use satellite data to predict land size or satellite data that can help?

Speaker 2: Thank you for the question. And I just want to say generally, thank you for everyone engaging with this question and answer session. I'm excited to answer some of these. So in this context, prediction can mean a number of things. I guess the first and probably the most common example is predicting the location of where a landslide is likely to occur. And we'll talk a lot more about this. We refer to this as thinking about susceptibility of parts of a landscape to landsliding. We'll talk a lot more about developing susceptibility models in part three of the training.

Speaker 2: So yeah, I encourage you to follow on. In terms of predicting the timing of a landslide, so thinking more about the specific hazard itself, prediction in this case would require forecast information, so information that is not just observational, what is happening with rainfall right now. We would want to predict something about the likely rainfall to come, and that is not something that is observed from satellite data. So we would obviously have to rely on model data, whether it's rainfall forecast data for a global rainfall forecast or a more local rainfall forecast.

Speaker 2: Either way, that information is typically course resolution and lower fidelity than a lot of observational data. And so prediction of landslides, whether through a kind of threshold-based model or otherwise, is certainly going to rely upon a calibration with forecast information rather than kind of observational data. We make a big distinction between our model, which we refer to as the Landsat Hazardous Awareness or LASA model, which is a situational awareness model that uses near real-time data.

Speaker 2: So this is not a prediction model. This is a forecast. So this is a, sorry, it is not a prediction model. It is a situational awareness model. And those are two distinct things. We have implemented a forecast mode for that, but it is still in a kind of like, I guess the best way to describe it is a early stages of developing accuracy metrics. And so I think, again, local prediction of landslide locations with satellite data is, I would argue, not something we're doing right now. And I think it's more situational awareness than it is Low answer, I apologize, but yeah, thanks for the good question.

Speaker 2: Hey, wonderful.

Speaker 1: Question number two. Sorry if I'm jumping ahead, but I'm wondering if we will discuss predicting versus assessing the potential threat of landslide areas. So in advance of an actual event versus post facto and realizing obviously the need for additional data such as, for example, soils, geology, digital elevation models, et cetera?

Speaker 2: Great question, and no apology necessary. You are, in fact, though, jumping ahead. We're looking at this in part three of this training, so thinking about predicting versus assessing potential threat is kind of exactly what we're thinking of in terms of sustainability analysis, and we will talk about that in part three of this training. Okay, wonderful.

Speaker 1: Question number three, which state-of-the-art convolutional neural network and artificial intelligence models are being used to map landslides with optical satellite imagery?

Speaker 2: Thanks for the question.

Speaker 4: So right now, like our team uses various different iterations of UNET, right? So it is not state of the art by today's standard. Recent literatures have shown transformer-based models to be better.

Speaker 3: At mapping landslides compared to CNN.

Speaker 4: So one example that participants here can go to is Recently, IBM and NASA has released a foundation model called and it has been.

Speaker 3: Updated to version two.

Speaker 4: So this kind of foundation models are trained on large amount of data, right? So what you can do is like, you can fine tune this model by showing this model a small amount of data, which suggests the model what a landslide looks like, right? So that can be used to quickly fine tune the model and then map landslides. So if you look at like QTV's recent paper, there is a downstream task where they looked at mapping landslides and they did some comparisons with different methods and found that the foundation models do work better compared to other methods.

Speaker 3: Thank you.

Speaker 2: Wonderful.

Speaker 1: Thank you so much, Pukar. Question number four, are there any papers or framework to train a convolutional neural network to automatically map landslides while also writing all the metadata of landslide inventory, a landslide inventory ought to have?

Speaker 4: So yes, there are open source public reference training data sets that can be used to train a deep learning based model. In terms of like metadata, I would say it's a little bit difficult because what the neural network is doing is like just mapping landslides, right? So it might only have like some attributes related to a polygon. So we would suggest caution looking at metadata that is generated automatically. So we would recommend some manual intervention while generating online looking for metadata.

Speaker 3: Robert, do you want to add something?

Speaker 2: No, I totally agree with you, Prakhar. I think there's just one, yeah, of course, that the Metadata analysis is something that we, again, we encourage manual way to ensure that there is like reliability and consistency. For the individual landslide polygons, obviously you could theoretically code up a model that could produce a file that have specific attributes attached to those things. But again, it would be important to sort of check through those. As I mentioned, polygons produced would obviously be important in any context and unsupervised landslide inventory production is certainly something we would be like, make sure you check it on both ends, put it that way.

Speaker 4: Yes, and I'd also like add some links to this answer.

Speaker 1: Okay, terrific. Thank you both, Robert and Pukar. Question number five. Do different satellites have different measurements and resolutions for Landsat? What are some good satellites for measuring Landsat in your opinion? I keep hearing Landsat, Sentinel-1, etc., but not sure of the differences. Are there any recommended reading resources to know the differences?

Speaker 2: So great question. And we appreciate the interest in learning more about these different elements. There are a wide variety of different satellites. And yeah, the resolution certainly varies significantly in the different benefits. I've mentioned the answer to this question. I think it might be important to review the fundamentals of remote sensing training that ARCET offers, as well as the introductory information from ARCET. post that in the final Q & A. I'll note that typically we work widely with Landsat and Sentinel-1 datasets, and part of that is because even though they are what we refer to as moderate resolution outputs, sort of 30 meter resolution pixels, they're very long archives for those datasets, which means we can go back a long time to look at the location of different pixels and say, you know, what's changed over time.

Speaker 2: And they are very well calibrated with respect to each other, as well as specific sources on Earth. More recently, commercial data has become an important part of Landsat mapping. We extensively used, for example, data from PlanetScope, as well as other commercial sources that NASA has obtained access to through the Commercial Satellite Data Acquisition Program. These tools typically are of higher resolution and often have a higher return rate, although in terms of the longer archives that are offered by these, I think, like Landsat and Sentinel, they have the same long archives.

Speaker 2: I will note that the question off here focused on Sentinel-1, and so I should make a comment that Sentinel-2 is the optical satellite offered by the Copernicus Sentinel Group in Europe. Sentinel-1 is the star satellite, so Landsat and Sentinel-1 are not exactly analogous. Landsat and Sentinel-2, however, those are kind of directly comparable datasets, and Sentinel-1, rather, is a star-based assessment. So just a note there.

Speaker 1: Okay, great, Rob, and thank you for that clarification. Question number six. Could you use a single algorithm combining object-based image analysis and a pixel-based method simultaneously to enhance the end result?

Speaker 3: Yes, thanks for the question.

Speaker 4: So in a basic sense, like pixel-based and object-based are intertwining with each other because we are using pixel-based features that gets turned.

Speaker 3: Into an object to map landslides, right?

Speaker 4: However, there are like a couple of studies that I've seen where object-based image analysis and deep learning based methods are combined. So what they're basically doing is like taking an output from OBIA and then use it into, or supply that to a deep learning model to map landslides. And they have shown that the combination of this like to work better than single instance.

Speaker 3: I can provide the link here as well later.

Speaker 1: Okay, thank you Pukar. Question number seven, what is a reasonable training ratio manually mapped slide area versus overall area of interest or solid approach?

Speaker 4: This is a very difficult question to answer. The general consensus is the larger the training data, the better. But it depends upon what you're trying to do, right? For example, if you're trying to do some rapid response landslide mapping, then you'll have.

Speaker 3: To make do with a few amounts of training data.

Speaker 4: But in my test of using SALAD for rapid response, I have seen acceptable results using about 50 landslide samples. So it really depends upon what you're trying to do. If you're trying to do a large-scale study where you need to be very accurate, then I would suggest using more training data.

Speaker 3: Robert, do you want to add something to this?

Speaker 2: No, I think that's an excellent answer. Again, I think this is very context-dependent. I guess the larger point is that if the landslides in the vicinity of the area, in the area you're interested in, are all spectrally very similar and have all very similar shapes and structures, your training data set may need to be smaller. But if there is a wide variety of different landslides, again, this may be something that you would want to have a larger training data set that would encapsulate the more diverse potential kind of landscapes and the potential sort of spectral indices, particularly if there's a very large distinction between bedrock and shallow landslides that are kind of either incorporating superficial material or deeper material that might have different spectral signatures.

Speaker 2: So that's, again, something to think about if you have, just based on initial inspections, a wide variety of kind of obvious landslides that you'd want to distinguish between.

Speaker 1: Hey, thank you, Rob and Pukar. Question eight, specifically for Pukar. Not sure why over-segmentation would not be a bad result from the model. Can you please explain further?

Speaker 4: Yeah, so landslides vary in shape and sizes within a satellite imagery, right? So one set of segmentation parameters, which is, able to segment all landslides to its perfect shape might not truly exist. You know, yeah, you can try various combination of segmentation parameters to get at it, but just to be clear, like it would be a very time consuming task, right? So as, as logic suggests, like, since, uh, at the end of the salad procedure, we are taking those objects, which were classified as landslide and merging them together.

Speaker 4: Over segmentation tends to be a lot easier because smaller things can be combined and then transform into bigger things. Whereas if you have like a large object which encompasses landslides and surrounding areas, it might be very difficult for you to break, right? So I'm not, I should rephrase it, maybe saying like over segmentation is not a good result, but it might be a preferred result in a sense that we can, easily combine that to form a bigger landslide area than having to break a over-segmented object.

Speaker 3: Thank you.

Speaker 1: Okay, Pukar, thank you very much. Question number nine. Is there a correlation study for repeat solid versus change detection investigations for the same area of interest?

Speaker 4: Yeah, so we haven't done a study to be specific, but salad and change detection basically works into works or like is useful for different different types of landslide mapping right where salad as it only uses a post event imagery. It cannot separate old and new landslides. So if you are to use it for for rapid response. that might create a problem because you get information about all the landslides in a satellite imagery. So rapid responders might not know which one is new, which one is.

Speaker 3: Old, right?

Speaker 4: Whereas with change detection, you leave all those old landslides out and be only detecting new landslides, which might be.

Speaker 3: Very beneficial for rapid responders.

Speaker 4: So in terms of study, we don't have it, but these are basically two different.

Speaker 3: Kinds of systems that can be used in two different scenarios.

Speaker 1: Okay, great. Thank you. Thank you, Pukkar. Question number 10. How can someone who isn't familiar with machine learning get started on using SALAD? Do I need any special software? Do I use Google Earth for satellite imagery?

Speaker 2: That's a great question. And I'm, Pukkar, I'll refer to you in a second. I just wanted to say that all the resources that refer to SALAD are available on the NASA's product page, which I've linked here. You can find papers on the model as well as the information that In general, you can sort of, the code is also available there. In terms of Google Earth, because you can't directly download Google Earth imagery, or at least if you're taking screenshots, for example, of Google Earth imagery, it's not going to be geo-referenced.

Speaker 2: That's not preferred. We would suggest using imagery from, you can have a download of the data that would be from either Sentinel, Landsat, or other data that you have available to you. And then in terms of special software, all of the code is available to run on open source platforms, including Python. So again, Pukkar, I defer to you on specifics, but I've missed it.

Speaker 3: Yeah, you're spot on. Like, Salad is created as an open source software, so if there's Python available, it should work.

Speaker 4: It might be a little bit difficult to install because it has some special packages like off your toolbox, but if anybody wants to try it and it's facing problems, please don't hesitate to send me an email.

Speaker 3: I'll try my best to solve that.

Speaker 1: Terrific. Thank you, Robin. Pukkar, question number 11. How often and at what point are ground truthing and or any on-site observation or calibration data or evidence obtained considered useful or necessary at what relative cost?

Speaker 2: So again, I think that's a really important question to ask. I would suggest that perhaps the alternative, again, let me start this by saying, it's going to depend on your application. So if you're in the emergency response space, Your initial question is, how much time do I need to spend figuring out what's going on? And low latency analysis is going to be most important. And that's where satellites can come in really handy. Producing automatic mapping is going to give you a sense of what is out there and where you might want to then prioritize your kind of localized assessments.

Speaker 2: These would be particularly important for things like infrastructure impacts or new landslide deposits that might have dammed rivers. And the stability of landslide dams, for example, is something that geoengineering experts think a lot about. So again, in terms of ground truthing, I think I would probably place this the other way around in that satellite data can be used as an initial guide to where to then go into the field to produce more detailed assessments. The other element is obviously there are some landslides that occur of smaller sizes than can be detected with satellite observations, and particularly this is the case if you're using relatively moderate resolution data like Landsat.

Speaker 2: You might find landslides that are still impactful that are below the resolution you can observe. So again, going into the field to observe those is important. I'll note that in general, we found a lot of benefit to working with researchers and scientists who then go into the field to assess those. I think, again, the application is what matters here. If you're interested in producing an accurate and reliable inventory of the long time periods that has been done effectively, for example, in Puerto Rico, the USGS has gone in and mapped landslides on the ground level to ground truth the estimates that were collected initially from satellite.

Speaker 2: But again, I think if you're looking to build accessibility models, for example, again, that might be something that would be relevant if you're looking for a very localized assessment. So it's context dependent, I think, again. But we would suggest that satellites are more of a potential guide for where to go to do that ground truth thing. And then based on the evaluation of where the most impacted potential areas would be, that could be a way to determine the cost benefit.

Speaker 1: Great. Thank you, Rob. Question number 12. Question specifically to Pukar. Is the multi-temporal landslide inventory data of Nepal mentioned in the presentation open source? What is its accuracy compared to manually extracted data? Also, just want to double check, are both the training and prediction based on the entire landslide polygons rather than just the source area?

Speaker 4: So yes, like whatever inventories are produced by the.

Speaker 3: Landslide team, they are open source.

Speaker 4: Some might have gone through Corrections some that were released in rapid response might not have been but in terms of these particular Multi-temporal inventories for NEPA like these are available in the NSIDC high mountain Asia deck so I can put in a link here later So these inventories have gone through manual corrections before we released it and yes training For salad generally like is done using complete landslide polygons. We never use partial or point.

Speaker 3: Perfect.

Speaker 1: Thank you, Pukkar. Question number 13, quite interesting. When training an AI model, as you said, the model is only as good as its training data. I can imagine that finding a good landslide data set for training, especially due to issues like cloud cover, is not easy. Could you find a high-quality data set that would be a strong candidate for training and then generate 1,000 synthetic data sets based on it, modifying the images, altering the landslide geometry, etc.? to reduce the time needed to find the training data and speed up the entire process?

Speaker 2: So I'm just going to offer an answer here. I think this is a really excellent research question. And I think understanding whether we can generate inventories without the input of kind of initial manual basis, that is something that we are currently exploring. And I don't know if you want to say anything else. Yes.

Speaker 4: So, so in this, in this context, like there is a method called data augmentation where like you augment the data set that you already have and create synthetic data out of it, right? What can that be like flipping the image, rotating the image, you know, like introducing noise to an image. So it's a common technique used in deep learning to increase the amount of landslide samples there are. If you look at like any deep learning based landslide mapping papers, I think you will share of the term data augmentation.

Speaker 4: So it is possible.

Speaker 1: Okay, terrific. Thank you, Rob and Pukar. Question 14. Is the geo AI segmentation method effective for monitoring active and recorded landslides?

Speaker 2: I don't know exactly what this refers to, Pukar.

Speaker 3: Yeah, it's a little bit confusing to me as well.

Speaker 4: So when we hear the top monitoring, we go into different realm of actually measuring movement. So geo AI segmentation is generally used for mapping landslides. So for monitoring, my preference would be utilizing synthetic aperture radar. You can monitor growth of the landslides using geo AI segmentation.

Speaker 1: Or reactivation. Okay, terrific. Moving on, question number 15. You mentioned INSAR is good for slow-moving landslides. What is a good method to detect a fast-moving landslide? What about a fast-moving landslide but at a smaller scale?

Speaker 2: Well, I think what we've been talking about in terms of optical mapping and kind of generic cell mapping is ultimately is exactly what we were kind of referring to for fastening landslides. So we would, again, think of shallow landslides and federal landslides as landslides occurring. If satellite overpasses, you're very unlikely to capture that happening literally as it happens. So you're just taking before and after. And so I think that would be, again, mapping the aftermath using optical data would be kind of our typical advice for that.

Speaker 4: Yeah, so in terms of SAR, like in terms of monitoring of fast moving landslides or like detecting, you can use pixel tracking, right, as opposed to INSAR. And also for detecting like change detection based method like log ratio, amplitude based change detection would be better. And the scale like typically depends on the resolution. So it depends upon what satellite image you are using.

Speaker 2: Okay, terrific.

Speaker 1: Question number 16. According to the rules of perspective, what parameters are allowed for foreshortening images taken with SAR?

Speaker 2: I'm probably going to ask if the person who asked this question could slightly rephrase it, because I'm not sure I fully understand what's being asked here. But unless I'm wrong, I don't.

Speaker 4: Yeah, no, me too, me too.

Speaker 3: Okay.

Speaker 1: I would point, whoever asked this question, we have a series, probably up to 10 SAR RSET trainings. So if you're interested just generally about SAR, the physics, applications, et cetera, we have some very thorough, everything from fundamentals to very advanced trainings on SAR. So if your question is more general in scope, then we do encourage you to visit some of those trainings, and we'll certainly place some links to these before we post it on the training website. Question number 17, will you be presenting more insights on fusing optical and SAR-based analyses?

Speaker 2: We'll talk a little bit about that in part three, but in general, I think the consensus here is that that is still an evolving field, and I think the science and research in that space is not something that is... I want to be careful about presenting exactly the details of that, just because I think we are one of a group of a number of different researchers involved in Trying to determine how to do that effectively. And I think the best thing we can do is point you in the office question to a number of recent research papers that have studied this and we can hopefully like introduce some of those elements.

Speaker 2: But yeah, in general, I think this is an evolving area. And so I wouldn't want to be tripping what is being done.

Speaker 1: Okay, terrific. Question 18. InSAR requires large satellite baselines to achieve high fringe sensitivity. Usually Sentinel-1 has a relatively small orbital tube of just 150 meters. As far as I know, the optimal baseline size for good fringe sensitivity is around two kilometers. Are there any satellite systems that have this baseline? Would you have some literature that you could share regarding this use case?

Speaker 2: I love a really complicated and detailed question like this. This is excellent. So off the top of my head, I do not have those numbers available, but I will look them up and figure out which is the other ones. I know that for the InSAR baseline site analysis, other tools that have been used include the AIOS suite of satellite data set, satellite data. And upcoming NISAR obviously will be part of that analysis. But yeah, off the top of my head, I don't know those numbers. So I'll figure that out.

Speaker 2: Regardless, you know differently.

Speaker 3: I think that's going to be- No, we'll have to look it up.

Speaker 2: Yeah.

Speaker 1: Okay, question number 19. Is it possible to predict a landslide from non-optical detection, such as INSAR, from interferograms?

Speaker 4: I particularly haven't worked in this field, but I've seen papers which have done so.

Speaker 3: So I'm pretty certain we can.

Speaker 2: Yeah. And I think typically with a lot of these, the cell-based analysis are often useful to take landslides that have no spectral observation. You can't see with optical data that I'm either moving slowly or have previously moved. And as a result, I'm forming an important part of analyses in parts of the world, particularly where there's very limited spectral difference between the optical nature of landslides and non-landslide-affected zones. So yeah, we've not specifically pointed to any of those today, but these certainly exist, and we'll make sure we have some references in the final answers here.

Speaker 1: Okay, wonderful. Question number 20 has to do with LIDAR. Have you used LIDAR data for looking at post-landslide forest structural changes? Is it possible to incorporate point cloud information into the automated approach?

Speaker 2: Go ahead, Thomas.

Speaker 3: Robert, go ahead.

Speaker 2: Okay. Well, I was going to say, first of all, is within the context of NASA data, we have not explicitly used LIDAR on the NASA LIDAR dataset. From ISAT, for example, we've determined is not necessarily aligned with the way we want to map landslides. I do, however, know that LIDAR from vehicles or drones or from ground-based assessment is widely used in that assessment, and particularly in terms of, like, you know, the vegetation change as well. I previously worked in groups where this has been widely done.

Speaker 2: And so again, yes, in terms of whether it's done or not, it certainly is. In terms of an automated approach to this, I think I have to look at the papers, Yeah, I have.

Speaker 4: Seen papers which are using lighter DMS to map landslide automatically.

Speaker 3: So is it possible.

Speaker 4: I think they will using Features that are slow roughness, etc.

Speaker 3: But I think it's possible.

Speaker 1: Okay, wonderful. And two questions left question 21 What automatic techniques are recommended for fast slides, given that the frequency of daily or longer resolution images is low for the required response times, no more than 48 hours?

Speaker 2: So, again, this, I think, is, you're actually right, that emergency response timelines are often very short. In our experience working with this data, and particularly working through NASA's disaster program, use of commercial data has been widely effective and incumbent in combination with the SALAD method. we can and have produced landslide inventories on short timelines of that kind of time span. Those have then been part of a kind of a wider kind of support response activation that the disaster response coordination system at NASA is involved with.

Speaker 2: If you want more information about that, I will certainly provide it in this response. But yeah, I think, Prakash, if you have anything to add there, I know that you've been part and parcel of some of those responses.

Speaker 4: Yeah, it typically depends upon like when the satellite imagery becomes available, right? At NASA, due to commercial satellite data acquisition program, we have access to commercial satellites, so we can do it.

Speaker 3: Pretty fast given cloud cover is not there. If cloud cover is there, we do use.

Speaker 4: Synthetic aperture radar based amplitude Change detection method to generate what we call landslide proxy heat maps, which gives you an idea of like where landslide might happen.

Speaker 3: I think Robert showed it in the slides.

Speaker 2: Okay, wonderful.

Speaker 1: And last question 22 is solid only for mapping landslide areas like feature extractions.

Speaker 4: Yeah, so solid in score even will be a based mapping system right here, it is mapping landslides because we are providing features. which helps us to detect landslides but it can be used also for general purpose classification like land cover mapping I think Robert showed an example of solid being used for flood damage.

Speaker 3: Mapping as well.

Speaker 4: So yeah, if you can bring in features which might be useful at detecting certain things, then it can be used for classification or mapping that as well.

Speaker 1: Okay, fantastic. We were able to get through all the questions. So thank you for everybody that did submit a question. Really terrific one. So thank you for everybody for doing that. Once again, Dr. Emerson and Dr. Amatya, wonderful, absolutely excellent presentations. I don't know if you've been following the chat at all, but I certainly have. I know that there's a lot of people you know, very pleased and very excited about what they were able to learn today. So once again, really thank to both of you for such a terrific presentation, as well as providing so many terrific answers to these really great questions.

Speaker 1: As we wrap up today, I did want to give an opportunity for both Dr. Emerson and Dr. Amatia to provide some closing comments. Maybe, Robert, maybe we can start with you.

Speaker 2: Thank you, Sean. And just want to say thank you to everyone involved in this training course today. And those who have attended and participated, I'm really delighted to be here and to support your learning. And I'd love to get to hear more from you either through our email or in the chat going forward. I look forward to training for part three. And I just want to say thank you to my co-host, Prakara Macha, and then obviously to the Asset team whose tremendous work has led to this. So thank you.

Speaker 1: Thank you.

Speaker 3: Thanks, everyone, for listening to our program or talk. If you have any questions, please don't hesitate to email me.

Speaker 4: I'll try my best to get you the answers and I'll see you on part three of this training. And thanks to Robert and the RSET team for allowing me to participate in this. Thank you.

Speaker 1: Terrific. And I also wanted to single out some colleagues from the RSET program, Natasha Johnson-Griffin, Selwyn Hudson-Odoi, and Sarah Cutshall. They were working, they have been working and will continue to support this training, all the parts before and after. And I just want to thank all of them for all the terrific job they're doing to make this run so smoothly and so successfully. And finally, I just want to thank everybody that joined today. Wherever you're joining from, I really appreciate it.

Speaker 1: We look forward to seeing you for the third and final part next Tuesday. So stay safe and we'll see you next week.

The Story