Landslide Workflows using GIS & Satellite Imagery

This article was originally posted on my firm's website: https://www.mapmyops.com/landslide-hazard-mapping-and-analysis and is updated there when an edit needs to be made. Hence, I'd recommend you to read this article from the hyperlink above.

________________________________________________________________

The United States Geological Survey defines a landslide as 'a type of mass wasting event' which involves 'the movement of a mass of rock, debris, or earth down a slope'.

As per the latest classification devised by Leroueil and Picarelli (2014), landslides can be categorized under 6 types: Fall, Topple, Slide, Spread, Flow and Slope Deformation.

Figure: Illustration of the major types of landslide movement. Source: USGS Publications

__________________________________________________________________________

The Geological Survey of India (GSI) is the nodal agency which creates a) region specific landslide forecasting models using rainfall and slope as central determinants as well as b) site specific landslide forecasting models using advanced scientific instruments at 8 sites in India currently. As per GSI, 90% of the landslides occurring in India is linked to rainfall - hence, this variable is given extra weightage in predictive models.

India is prone to significant amount of landslide events - 61 have been recorded within the first 8 months of 2021 itself. The Geological Survey of India highlights that "about 0.42 million sq. km or 12.6% of land area, excluding snow covered area, is prone to landslide hazard" - a significant threat. The vicious slant of slope and the pull of gravity is essentially what influences landslides. Factors such as rainfall, loose soil, ground deformation due to earthquakes etc. aids and abets this hazard.

In one of my previous articles, I had covered in detail how the massive rock slide in Chamoli, Uttarakhand unfolded and perhaps, could have been prevented had an early warning system been in place.

__________________________________________________________________________

Landslide Hazard Analysis & Mapping involves the usage of a powerful location analytics platform - GIS (Geographic Information System) and complex datasets such as Radar Imagery to study the terrain, slope and other influencing parameters which, in turn, helps us to identify locations which can be prone to landslide as well as where the landslide has already occurred.

In this article, I will expand into Landslide Hazard Mapping using 3 case-lets which involve the use of GIS a) Creating a landslide risk model b) Identifying stakeholders and infrastructure at most risk from landslides and c) Rapid Detection of Landslide location using Satellite Imagery. I expect these to give you an end-to-end view on the subject matter as well as about the geo-technology used to process it.

In case you do not wish to read this rather detailed article, you may view a condensed summary from the video below-

__________________________________________________________________________

1. Creating a Landslide Risk Model

Context: Existing Landslide Risk models need to be re-evaluated after major events such as earthquakes, wildfires, heavy rainfall etc. because the terrain properties tend to change after such events making certain areas more susceptible thereby making existing models redundant.

For this study, we will create a new, step-by-step, landslide risk model taking into account the recent wildfire in our area of study.

(Much Thanks to Learn ArcGIS for the study material)

First, we will load the layer containing area affected by Wildfire i.e. the Burn Scar layer.

Next, we'll add a B&W Terrain Elevation layer. Lighter Shades indicate areas of high elevation and vice versa.

Below, the B&W Terrain Elevation Layer has been converted into a visually informative 'Slope' layer. Green areas have the lowest slant i.e. Slope whereas Red areas have the highest Slope - these are highly susceptible to landslides.

Next, a B&W Mean Rainfall Layer has been added. Lighter Shades indicate areas of high mean rainfall and vice versa. Areas affected by higher rainfall are more susceptible to landslides as the surface becomes slippery.

The fourth & final layer which we'll use is the Multi-spectral Imagery layer (RGB visualization) captured from Landsat satellite. This layer is useful to derive vegetation density. Sparser the vegetation, the more susceptible it is to landslides.

Now that we have all the necessary data loaded onto the GIS software, we will proceed to create a geo-processing workflow as depicted in the image below. Such workflows encompass the 'methodology' used to derive the output. Moreover, it eliminates the need to do step by step imagery processing. Once the formula is created, the final output can be rendered just by clicking the 'Run' button.

Figure: Geo-Processing Workflow - Landslide Risk Model

So what does this model mean? At first, we are asking the location analytics software (GIS) to - a) convert the base multi-spectral imagery layer into an NDVI (normalized difference vegetation index) layer. This layer is useful to derive vegetation density information (dense vegetation keeps the soil together whereas sparse vegetation means that the soil is prone to breaking up which exposes it to landslide risk).

The Remap buttons are our predetermined classifications of the layer.

Figure: When we 'remap' the rainfall layer as per the table above, we are telling the GIS software to classify and visualize the layer values / output into the specified categories

The Weighted Sum part of the workflow is how we combine the outputs of individual layers.

For example, in the model we've created, we've assigned NDVI layer a higher weightage (2 out of 4 i.e. 50%) whereas slope and rainfall classifications have been assigned same weightage of 25% each. This is because the researchers feel that post a wildfire, it is the vegetation density which will be the most important determinant of a landslide. (Wildfires result in loss of vegetation, making the soil loose and shaky and hence more susceptible to landslides).

With the INT function, we are asking the software to truncate the values to integer form whereas with the Clip function we are asking the software to conduct the analysis in the 'Burn Scar' only i.e. in the wildfire affected region and to remove all other information layers outside the boundary.

Once we run the workflow model we get the output as below - The burn scar has risk scores from 7 to 15. Higher the scores (lighter colors), higher the landslide potential in that particular region.

Figure: Landslide Risk Model Workflow Output

Next, we modify the symbology so that the output is visually engaging. Darker regions are more susceptible to landslides.

We conclude our analysis by identifying spots on a main road within the burn scar where the susceptibility scores are >12 i.e. very high. Larger the circle, higher the score i.e. greater the landslide risk at that point on the road.

I hope you were able to appreciate how GIS software can process information rich data layers to derive a landslide risk model in quick time. The quality of geodata is as much important as the software's prowess (Esri's ArcGIS Pro in this case).

__________________________________________________________________________

2. Identifying at Risk Stakeholders & Infrastructure from Landslide Risk Model

Context: You can assume that this case study is an extension of the above. While in the exercise above we had created a landslide risk model, in our upcoming study we will use it as our base layer to identify stakeholders and infrastructure which are at most risk from landslides and its destructive range.

Most landslide damage doesn't occur where the landslide originates, rather, it occurs when the debris accumulate and accelerate their speed flowing downwards mostly due to heavy rainfall and resulting floods.

(Much Thanks to Learn ArcGIS for the study material)

The Landslide risk layer below becomes our base layer. Darkest portions have the highest risk and vice versa. Slope, Direction of Slope (Aspect) and Soil Type were the determinants used to create this landslide risk model.

Below, we've filtered the map to show only the highest risk landslide areas (dark brown). Also, we've added another information layer - that of blocks (Pin-codes). The larger the circle, the higher the Population at that pin-code.

Next, we've added 'Floodways' layer to our map. These are regions where water accumulates and channels downward in the event of a heavy rainfall ensued by floods.

Below, we've used the GIS software to create a 200 metre buffer around the floodways layer. Termed as 'flood fringe' these are regions which are at risk of being inundated in the event of heavy flooding.

Now's the time for some analysis. We proceed to inspect the individual cities within the map. This city below has a high concentration of both - floodway routes as well as population - an unfavorable combination.

This city below has lesser concentration of both - population and floodways, however, the latter are right on top of major roads. This is concerning because the population may get stranded in the event of heavy floods.

The third and final city below is the least at risk going by the map output. The population is clustered away from the floodways and high risk landslide zones. Also, the main roads are far away too, reducing the risk of stranding.

Next we'll enrich the affected areas with socio-economic and demographic data. Basically, the map view will remain the same but data values i.e. attributes will be added to the layer for us to analyse the risk in a more in-depth manner.

Figure: Numerous categories of authorized datasets are included in Esri's ArcGIS Online GIS platform

By enriching the layer with population data, we get useful information at a click of a button as denoted in the image below

Next, we've filtered the view to show us floodways which intersect major landslide zones. These areas (highlighted in yellow) are especially concerning as they will be exposed to severe debris flooding in the event of a landslide...

The entire yellow stretch is not at highest risk though. Debris flow tends to halt after a certain distance. Hence, we've visualized in green only those areas which are <1 km from the intersection of landslide + floodways layer.

These areas are our prime areas of concern. We proceed to enrich our data further. Now apart from population, we can see additional information such as housing units, property density, road density and number of elderly persons in the highest risk areas. These lend more meaning to our analysis and also helps us to prioritize the preventive measures.

Hence, conducting Spatial Analytics using GIS software helps us to expand from the initial landslide risk model and identify stakeholders and infrastructure at the highest risk.

__________________________________________________________________________

3. Detecting Landslide from Satellite Imagery

Context: While in the two cases above we've seen how to prepare an early warning system for identifying landslide prone areas and at-risk stakeholders, in our third and final case we'll see how to rapidly detect landslide location from satellite imagery once it has occurred.

In the era we live in, one can argue that landslide locations are detected very quickly - footage of flowing debris are quickly captured and circulated in the news and on social media. That being said, knowing the true extent of the landslide, the location of its initiation, its spread and studying the underlying parameters historically, at the time of the disaster and in the weeks thereafter is something beyond the reach of a common man. Moreover, many landslides may go unnoticed when there are no human settlements around and yet, they are equally necessary for researchers to study.

(Much thanks to RUS Copernicus for the study material)

The process of detecting landslides entails analyzing two radar images over the same area and close to each other in terms of timeline - one before the event and another after the event.

As you would imagine, the next steps involves observing discernible changes in the surface to see if it demonstrates properties indicative of a landslide.

Radar imagery is much more powerful that optical imagery - it penetrates the weather and can capture data even during night due to the presence of active sensor onboard the satellite. Additionally, the data captured is information-ally complex and rich which helps us to compare and contrast the observations and detect irregularities i.e. perform change detection.

Below is how radar imagery looks like. Very bland and unappealing when compared to an optical image. However, the information captured here is immense and using the right geo-processing tools, several interesting phenomenon can be unearthed.

The processing chain of satellite imagery is also much more complex and technical than what we discussed in our previous two cases - hence we will not delve into that. Just know that a lot of processing effort goes towards making images comparable and analysis-worthy.

Figure: Depiction of one part of the complex processing chain (Interferometric Processing) to detect landslides from radar imagery.

Once the processing is done and coregistered image is create, we visualize it by using a RGB composite (figure below). Within, we can see 4 colors predominantly - yellow pixels (where both pre and post imagery values are the same), red pixels (where there is some extra information in the pre-image as compared to the post-image), green pixels (where there is some extra information in the post-image as compared to the pre-image) and black areas (Low intensity / No data values - likely to be water bodies). The green pixels should pique our interest as the surface property has significantly changed in the newer radar image when compared to the older one.

Below is the RGB composite of the co-registered image. The green area might be a candidate for the landslide location.

Below is the Raw Image and RGB visualization tiles over the same area. Higher intensity (white triangle) is spotted in the newer July image.

The bright green area in a section of the imagery below gives us an indication that there has been a significant change in the second image - this could be the potential landslide location and we must inspect it further to validate our assumption.

As a next step, we create an interferogram. Again, without expanding into the technical aspect, what we essentially do is to overlay one image on top of the other and analyse the change in multiple data values (phase and coherence) over the same surface / pixel.

Figure: Phase (Left) and Coherence (Right) visualization from interferogram

Loss of phase (as seen in the highlighted section) will be present if there is certain land deformity (indicative of landslide). To explain 'Phase' as a concept in a simple way - imagine seeing a calm lake and the waves flowing towards you in consistent patterns. Now if you throw a stone into the water, the ripples will continue to be wave-like but of different dimensions and the rhythm disturbs the natural patterns. Loss of phase essentially means throwing a stone into calm waters and monitoring the disturbance as seen in the image above which is indicative of significant surface changes such as earthquake, landslides etc.

To interpret the image on the right above, pixels with very low coherence values are black in color and pixels with very high coherence values are white in color. If the pixel values in pre and post images tally, then it is said to be high coherence whereas if there is a deviation in the values, which in the case of a landslide is highly anticipated, then there will be lower coherence as seen in the image above.

By using various variables to confirm our theory, we have located the landslide using satellite imagery. Like both the cases discussed in this article previously, this is a real use-case as well.

Use the slider below (image is hyperlinked to the slider) to see that the imagery output tallies with Google Earth imagery which shows the exact location of the Fagraskógarfjall landslide in Iceland -

To summarize, Detection and Validation of Landslide was done using four different observations a) High intensity values in post landslide image, b) RGB validation of the same (High Greens), c) Low Coherence in the bundled image and d) Loss of Phase in the bundled image.

(Please note that each of these observations need to individually affirm the presence of landslide for us to be completely sure / validate our hypothesis about the potential landslide location.)

Thus, analyzing Satellite imagery helps us to derive valuable landslide information in quick time. It is also cost-effective when compared to on-ground, instrument-intensive methods of landslide detection.

(That being said, I had tried to replicate the Case 3 tutorial on recent Indian landslide incidents -Mizoram, Idukki etc. After spending considerable hours and days in doing the processing work, my output somehow didn't validate the landslide occurrence. I suspect, the satellite imagery model works best on large landslides. Also, the selection of post incident image could play a crucial role. Radar backscatter over flooded areas post incident might jeopardize the model from giving the best results.)

Interesting isn't it?

__________________________________________________________________________

Intelloc Mapping Services | Mapmyops is engaged in selling products which capture geodata (Drones & Drone Services), process geodata (Geographic Information System) as well as enhances geodata (Imagery + Analytics & PoI Solutions). Together, these help organizations to benefit from Geo-Intelligence for purposes such as operations improvement, project management and digital enabled growth.

Write to us on [email protected].

Regards, Arpit