Performance of manual and automatic detection of dry snow avalanches in Sentinel-1 SAR images

Abstract

Radar satellite-borne snow avalanche detection has rapidly grown into an important method for monitoring of avalanche activity over large spatial and long temporal scales. With increased application of Sentinel-1 data for avalanche detection, the need for improved performance evaluation arises. In this study, we make use of a unique dataset of field-based avalanche observations from eight adjacent avalanche paths at lake Tyin in central Norway. The dataset is a complete record of 318 dry slab avalanches that released during the winters 2016–2020, with information on release timing and extent. The dataset thus allows for detailed evaluation of the performance of automatic and manual avalanche detection in Sentinel-1 images, were both techniques make use of relative temporal backscatter intensity increase in case of an avalanche. Both automatic and manual detection underperform compared to previous studies with a probability of detection (POD) of 5.9% (false alarm rate (FAR) of 5.9%) and 11.3% (FAR of 19.27%) respectively. From the low relative backscatter intensity contrast of field-observed dry slab avalanches, it becomes evident that a higher backscatter contrast between avalanches and surrounding is needed for detectability in Sentinel-1 images. Neither the lag time between avalanche release and detection in a Sentinel-1 image, nor local incidence angle can explain the low POD's. Moreover, an analysis of meteorological conditions prior a during avalanche release and/or detection can explain the low POD's, given that differing snow conditions influence radar backscatter intensity. Finally, we cannot rule out that the small dataset of field-observed avalanches and the one-directional slope aspect of the study area influence our results, however, we believe that a physical limit of detectability of dry snow avalanches in C-band radar satellite data is reached.