Endeavours into a more automated workflow for regional scale landslide and flash flood event detection in the tropics using IMCLASS

Geomorphic hazards such as landslides and flash floods (hereafter called GH) often co-occur
and interact imposing significant impacts in the landscape. Particularly in the tropics, where
GH are under-researched while impact is disproportionally high, establishing regional-scale
inventories of GH events is essential to better understand the behaviour and the patterns in
GH event occurrence. Robust AI-based detection tools such as the IMCLASS classifier
provide an excellent solution to accurately determine the location of GH events. However,
they rely on accurate training samples and require some knowledge on the timing of the event.
This information is regularly unavailable when exploring for new GH events in inaccessible
areas such as the tropics. Here we present our first endeavours into an automated workflow
for detecting unknown events in the tropics using the IMCLASS detection tool associated to
an unsupervised building of training samples using time series of Copernicus Sentinel 2
imagery. Per pixel, we investigate the cumulative difference from the mean over time for a
multitude of spectral index time series (e.g. NDVI, BI, SAVI) and their related z-score time
series. The method allows us to distinguish GH-affected and non-affected pixels based on the
prominence of the peak, and determine an approximate timing based on the location of the
peak within the timeseries. Both information are then used as input for the IMCLASS
classifier. The method is highly optimized in terms of computation time allowing to process
large regions of interest. Preliminary results over Uvira, DRC and the Mahale Mountains,
Tanzania, have shown to be encouraging and provide insight into a more automated workflow
applicable on the regional scale where event occurrence and timing is yet unknown. Further
steps will consist of adapting the workflow to different landscape, topography and climatic
regions.