Regional Detection of Landslide and Flash Flood Events in the East African Rift

<p>Satellite remote sensing is frequently used for spatial and temporal detection of geomorphic hazards (GH) such as landslides and flash floods. Establishing regional-scale inventories of GH events is crucial to understanding their behavior in both space and time, particularly in the tropics, where GH are under-researched, and impact is disproportionally high. Recently, an increased focus is seen on the use of machine- (ML) and deep learning (DL) methodologies for accurate detection of GH. These methodologies however, have in common that they rely on accurate information on either the GH location (training samples) or the GH timing (pre- and post-event imagery), making them practically unusable in unseen areas without any information on GH occurrences. Here, we aim to develop a methodology that allows for regional multitemporal GH event detection, providing both location and (semi-accurate) timing of GH events without any prior knowledge on GH event occurrence. We additionally test the suitability of our results as input for a more conventional ML-based classifier &#8211; ALADIM. We develop a pixel-based methodology using the open-access, high spatial resolution (10m) Copernicus Sentinel-2 time series from 2016 to 2021. Our methodology uses the peak of the cumulative difference from the mean for a multitude of spectral indices (NDVI, NBR, BI, SAVI, etc.) and allows us to create a map per Sentinel-2 tile that identifies impacted pixels and their related timing. We applied our methodology on six Sentinel-2 tiles in the tropical East African Rift and were able to successfully identify 29 GH events. From these, we chose 12 GH events, with a total of ~ 3900 landslide and flash flood features that occurred in different parts of the time series, in different landscapes and contained different GH event compositions (e.g. GH size, landslide to flash flood ratio). For these GH events, we validated the automatically derived GH event timing, and we used our results to automatically create training samples that served as input for the ALADIM classifier. Estimated GH event timing has on average a ~2-weeks difference from the last available cloud-free pre-event image and a ~ 6-weeks difference from the first available cloud-free post-event image. A general increase in pixel-by-pixel detection accuracy is seen when implementing our output in the ALADIM classifier, with some exceptions for GH event inventories that contain a large amount of small-sized GH features. The detection accuracies are influenced by the amount of cloud cover (less impacted pixels identified in highly cloudy regions), differences in landscapes (low noise levels in pristine forests, and high noise levels in densely cultivated landscapes) and the size distribution of GH events (lower accuracies for GH events that contained a lot of small sized features). Our methodology is working in varying landscapes, shows potential for transferability and in combination with ML-based classifiers allows to better automatize the GH event detection process. Additionally, it is highly optimized in terms of computation time allowing to process large regions of interest, within a relative short time span.</p>