A semi-supervised multi-temporal landslide and flash flood event detection methodology for unexplored regions using massive satellite image time series

Landslides and flash floods are geomorphic hazards (GH) that often co-occur and interact and frequently lead to societal and environmental impact. The compilation of detailed multi-temporal inventories of GH events over a variety of contrasting natural as well as human-influenced landscapes is essential to understanding their behavior in both space and time and allows to unravel the human drivers from the natural baselines. Yet, creating multi-temporal inventories of these GH events remains difficult and costly in terms of human labor, especially when relatively large regions are investigated. Methods to derive GH location from satellite optical imagery have been continuously developed and have shown a clear shift in recent years from conventional methodologies like thresholding and regression to machine learning (ML) methodologies given their improved predictive performance. However, these current generation ML methodologies generally rely on accurate information on either the GH location (training samples) or the GH timing (pre- and post-event imagery), making them unfit in unexplored regions without a priori information on GH occurrences. Currently, a detection methodology to create multi-temporal GH event inventories applicable in relatively large unexplored areas containing a variety of landscapes does not yet exist. We present a new semi-supervised methodology that allows for the detection of both location and timing of GH event occurrence with optical time series, while minimizing manual user interventions. We use the peak of the cumulative difference to the mean for a multitude of spectral indices derived from open-access, high spatial resolution (10–20 m) Copernicus Sentinel-2 time series and generate a map per Sentinel-2 tile that identifies impacted pixels and their related timing. These maps are used to identify GH event impacted zones. We use the generated maps, the identified GH events impacted zones and the automatically derived timing and use them as training sample in a Random Forest classifier to improve the spatial detection accuracy within the impacted zone. We showcase the methodology on six Sentinel-2 tiles in the tropical East African Rift where we detect 29 GH events between 2016 and 2021. We use 12 of these GH events (totalizing ∼3900 GH features) with varying time of occurrence, contrasting landscape conditions and different landslide to flash flood ratios to validate the detection methodology. The average identified timing of the GH events lies within two to four weeks of their actual occurrence. The sensitivity of the methodology is mainly influenced by the differences in landscapes, the amount of cloud cover and the size of the GH events. Our methodology is applicable in various landscapes, can be run in a systematic mode, and is dependent only on a few parameters. The methodology is adapted for massive computation.