Timing landslide and flash flood events from SAR satellite: a regionally applicable methodology illustrated in African cloud-covered tropical environments

Abstract

Abstract. Landslides and flash floods are geomorphic hazards (GHs) that often co-occur and interact. They generally occur very quickly, leading to catastrophic socioeconomic impacts. Understanding the temporal patterns of occurrence of GH events is essential for hazard assessment, early warning, and disaster risk reduction strategies. However, temporal information is often poorly constrained, especially in frequently cloud-covered tropical regions, where optical-based satellite data are insufficient. Here we present a regionally applicable methodology to accurately estimate GH event timing that requires no prior knowledge of the GH event timing, using synthetic aperture radar (SAR) remote sensing. SAR can penetrate through clouds and therefore provides an ideal tool for constraining GH event timing. We use the open-access Copernicus Sentinel-1 (S1) SAR satellite that provides global coverage, high spatial resolution (∼10–15 m), and a high repeat time (6–12 d) from 2016 to 2020. We investigate the amplitude, detrended amplitude, spatial amplitude correlation, coherence, and detrended coherence time series in their suitability to constrain GH event timing. We apply the methodology on four recent large GH events located in Uganda, Rwanda, Burundi, and the Democratic Republic of the Congo (DRC) containing a total of about 2500 manually mapped landslides and flash flood features located in several contrasting landscape types. The amplitude and detrended amplitude time series in our methodology do not prove to be effective in accurate GH event timing estimation, with estimated timing accuracies ranging from a 13 to 1000 d difference. A clear increase in accuracy is obtained from spatial amplitude correlation (SAC) with estimated timing accuracies ranging from a 1 to 85 d difference. However, the most accurate results are achieved with coherence and detrended coherence with estimated timing accuracies ranging from a 1 to 47 d difference. The amplitude time series reflect the influence of seasonal dynamics, which cause the timing estimations to be further away from the actual GH event occurrence compared to the other data products. Timing estimations are generally closer to the actual GH event occurrence for GH events within homogenous densely vegetated landscape and further for GH events within complex cultivated heterogenous landscapes. We believe that the complexity of the different contrasting landscapes we study is an added value for the transferability of the methodology, and together with the open-access and global coverage of S1 data it has the potential to be widely applicable.