Optimizing Landslide Detection and Validation Through Sentinel-1 Radar Imagery: Case Studies of Hokkaido and Hiroshima in Japan

Satellite-based (optical and radar) imaging techniques are frequently employed for landslide detection with short time intervals, especially for emergency response, risk assessment, and hazard mitigation. While radar imagery overcomes many limitations associated with optical imagery, it is still constrained by various noise and distortions, potentially resulting in the underestimation of significant landslides and an increased prevalence of smaller landslides. The primary objective of this study is to develop a methodology that can enhance the detection accuracy of different-sized landslides using radar imagery. We utilize multi-temporal imagery from the C-band Sentinel-1 satellites, which are freely available on the Google Earth Engine platform, to analyze the rainfall-triggered landslides in Hiroshima and earthquake-induced landslides in Hokkaido, Japan, across different time frames. We investigate the change in the backscattering intensity coefficient of radar images with VV (vertical transmit-vertical receive) polarization and test the effectiveness of speckle filtering and range-doppler terrain correction using the SNAP software. We explore different cutoffs of the backscattering intensity coefficient, which increase the performance of the area under the ROC curve, to determine the threshold that best captures significant changes associated with landslide events. We choose the same threshold for distinguishing landslide and non-landslide areas in the backscattering intensity maps with and without speckle filtering. For validation, we quantitatively compare our satellite-based detection results with the high-resolution optical-based landslide inventories in terms of the spatial overlap, sizes, and scaling properties of landslides. The results indicate that filtering Sentinel-1 images enables more robust detection of larger landslides and simultaneously reduces false negatives for smaller landslides. These additional processing steps hold promise for more accurate and efficient landslide monitoring at various spatial scales and potentially over time.