Automatic validation of Sentinel-1 borne snow avalanche detections

Abstract

Snow avalanches threaten human lives, settlements and roads in snow covered mountainous areas. For avalanche forecasting, knowledge of the spatio-temporal occurrence of avalanche activity is critical. Automatic avalanche detection algorithms have been developed to enable consistent avalanche activity monitoring for large regions. The Satskred avalanche detection algorithm developed by NORCE applies synthetic aperture radar (SAR) data from the Sentinel-1 satellite constellation and detects avalanches through a relative increase in energy scattered back to the radar from avalanche debris. Field validation of all automatically detected features is desirable, but not achievable due to weather- , light- , and avalanche danger-conditions as well as avalanches occurring at remote locations.

In this thesis, an algorithm is presented for automatic comparison of the Satskred avalanche detections to crowd-sourced avalanche observations from regObs, the Norwegian public registry for snow-, weather-, flood-, and ice observations. Thereby, the validation set of field observed avalanches grows with every registered observation and validation of detected features can be performed without further manual intervention. To evaluate whether a detection matches an observed avalanche, the comparison algorithm initially filters detections by time period to ensure temporal similarity. Then, the detection is evaluated with regards to distance, slope aspect and membership of the same drainage basin region as the observation to ensure spatial similarity. If the detection fulfills all the similarity requirements, it is considered to likely represent the same avalanche.

Studying a 120 x 86 km area centered over Tromsø in Northern Norway, 308 avalanche observations from 2014 - 2019 were automatically compared to a set of avalanche detections from the same area and time period. The field observations were used as a truth-set and the resulting probability of detection (POD) for the Satskred algorithm was 25.3% (78 out of 308). Further analysis identified trends of larger POD for wet- than dry avalanches, and an increasing POD with avalanche size. A large proportion of avalanches entered to the regObs database are dry slab avalanches, which was found to partly explain the low POD.