Filtering sources of false alarms from automatically detected avalanches in Sentinel-1 SAR images

Abstract

This study analyses automatically detected avalanche deposits in Sentinel-1 SAR images by the SatSkred algorithm from the three winter seasons 2018/19, 2019/20 and 2020/21. SatSkred allows for near-real-time automatic avalanche detection, but previous studies indicate great over-detection due to misinterpreted signatures in the change detection images. By inspecting scenarios in the dataset with high avalanche activity, we aim to discern patterns for false positive detections and suggest concepts for improvement, which the developers can use to increase the true detection accuracy. We have inspected smaller areas in a total of 16 RGB-change detection images. By developing a Greenness Indicator, we have quantified the band values in RGB-composites to explain the relative change in backscatter. Further, this was used to analyse the greenness in non-avalanche terrain, avalanche terrain, detection area, and a 40 meters buffer zone in potentially true and false change detection scenarios. The greenness in non-avalanche terrain correlated highly with the avalanche terrain (corr = 0.85), while the greenness in the detection area correlated highly with the buffer zone (corr = 0.74). We found no dependence between the avalanche terrain and the detection area. Our results indicate that a contrast of minimum 0.25 between the buffer zone and detection area excluded 83\% of the potentially false detections while including 92\% of the potentially true detections. Following, a contrast of minimum 0.48 between the avalanche terrain and the detected area excluded 78\% of the potentially false detections and included 73\% of the potentially true detections. The thresholds were further tested in two case studies with changing meteorological conditions. Our case studies revealed high sensitivity to the threshold of 0.48 between the avalanche terrain and the detected area, indicating a high potential for false positive errors. Our findings indicate a potential for false detections to be filtered by the proposed threshold value between the detections and the surrounding buffer areas.