I recently took an online course on CLASlite offered by the Carnegie Institution for Science. CLASlite is a software designed to automatically detect deforestation, forest degradation and forest regrowth. It uses the AutoMCU algorithm to detect the percentage of various types of vegetation, allowing to automatically detect deforestation using Landsat images.
In one of my blog posts, I used Landsat image to see the Rockfall scar on Annapurna mountain. I detected the scar manually, but imagine if we could do this automatically. Manually delineating landslide takes a lot of time. We need a lot of historical landslide data to build better landslide hazard maps and landslide prediction models. Being able to automatically detect landslide would be very useful to the geo-disaster community.
A useful index used in remote sensing is the Normalized Difference Vegetation Index (NDVI). This is used to detect vegetations. Most of the hilly region in the Himalayan region is covered with vegetation. When a landslide occurs in such a vegetated area, the vegetation will be swept away, leaving fresh bare soil. Thus, could a change in NDVI be used as an indication of landslide?
Not necessarily. Deforestation would also have a similar effect in NDVI, thus it would be difficult to differentiate landslide from deforestation or another normal bare surface. However, there is a peculiar difference between a fresh landslide and existing normal bare surface. Fresh landslide generally consists of no vegetation, but an existing bare surface would contain some non-photosynthetic vegetations like chopped woods, dead leaves or other dry plans. This is when I thought AutoMCU can come in hand.
Unlike NDVI, AutoMCU outputs three levels of indicators: Bare Substrate (BS), Photosynthetic Vegetation (PV) and Non-photosynthetic Vegetation (NPV). NDVI can only detect photosynthetic vegetation, however AutoMCU can distinguish Bare Substrate from Non-photosynthetic Vegetation. I wanted to try this out myself, so I used the Landsat 8 images near the recent Jure landslide site in Sindhupalchowk, Nepal.
Figure 1 (a) shows the pan-sharpened true color raw image of a site showing a curved road with small landslide scars towards the south-east. It also shows some bare land areas, probably but normal bare lands which are not necessarily landslides. Figure 1 (b) shows calibrated true color reflectance which will be used to compute NDVI and AutoMCU algorithms.
Figure 1 (c) shows the NDVI of the same location. We can see vegetation are shown in dark and bare areas are white. However, both the landslide as well as normal bare lands are white making them indistinguishable from each other. On the other hand, the output of AutoMCU (Figure 1 (d)) shows the landslide as Bare Substrate (red) and the normal bare area as Non-Photosynthetic Vegetation (blue). This should be because the normal bare lands would generally consist of some dead leaves or other dry vegetations.
Next, I used the following criteria to segment landslide:
where, BS = bare substrate; PV = Photosynthetic Vegetation; NPV = Non-Photosynthetic Vegetation
In Figure 2 (c) and (d) we can see that the landslide areas are well identified by these criteria. There are a few extra mismatches, but most of the normal bare areas are correctly not detected as a landslide. We should be able to further improve these criteria and add other criteria extra (like DEM and slope data) or Change detection. Using the output of this AutoMCU criterion before manual landslide delineation could reduce the manual landslide delineation time by order of magnitude because it would only involve identifying the wrong classifications.
In one of my blog posts, I used Landsat image to see the Rockfall scar on Annapurna mountain. I detected the scar manually, but imagine if we could do this automatically. Manually delineating landslide takes a lot of time. We need a lot of historical landslide data to build better landslide hazard maps and landslide prediction models. Being able to automatically detect landslide would be very useful to the geo-disaster community.
A useful index used in remote sensing is the Normalized Difference Vegetation Index (NDVI). This is used to detect vegetations. Most of the hilly region in the Himalayan region is covered with vegetation. When a landslide occurs in such a vegetated area, the vegetation will be swept away, leaving fresh bare soil. Thus, could a change in NDVI be used as an indication of landslide?
Not necessarily. Deforestation would also have a similar effect in NDVI, thus it would be difficult to differentiate landslide from deforestation or another normal bare surface. However, there is a peculiar difference between a fresh landslide and existing normal bare surface. Fresh landslide generally consists of no vegetation, but an existing bare surface would contain some non-photosynthetic vegetations like chopped woods, dead leaves or other dry plans. This is when I thought AutoMCU can come in hand.
Unlike NDVI, AutoMCU outputs three levels of indicators: Bare Substrate (BS), Photosynthetic Vegetation (PV) and Non-photosynthetic Vegetation (NPV). NDVI can only detect photosynthetic vegetation, however AutoMCU can distinguish Bare Substrate from Non-photosynthetic Vegetation. I wanted to try this out myself, so I used the Landsat 8 images near the recent Jure landslide site in Sindhupalchowk, Nepal.
Figure 1 (a) shows the pan-sharpened true color raw image of a site showing a curved road with small landslide scars towards the south-east. It also shows some bare land areas, probably but normal bare lands which are not necessarily landslides. Figure 1 (b) shows calibrated true color reflectance which will be used to compute NDVI and AutoMCU algorithms.
Figure 1 (c) shows the NDVI of the same location. We can see vegetation are shown in dark and bare areas are white. However, both the landslide as well as normal bare lands are white making them indistinguishable from each other. On the other hand, the output of AutoMCU (Figure 1 (d)) shows the landslide as Bare Substrate (red) and the normal bare area as Non-Photosynthetic Vegetation (blue). This should be because the normal bare lands would generally consist of some dead leaves or other dry vegetations.
Next, I used the following criteria to segment landslide:
BS > 35% AND PV < 50%
where, BS = bare substrate; PV = Photosynthetic Vegetation; NPV = Non-Photosynthetic Vegetation
In Figure 2 (c) and (d) we can see that the landslide areas are well identified by these criteria. There are a few extra mismatches, but most of the normal bare areas are correctly not detected as a landslide. We should be able to further improve these criteria and add other criteria extra (like DEM and slope data) or Change detection. Using the output of this AutoMCU criterion before manual landslide delineation could reduce the manual landslide delineation time by order of magnitude because it would only involve identifying the wrong classifications.
No comments:
Post a Comment