Optimisation of Method for Snow Avalanche Detection in SAR images: Supporting the development of snow avalanche mapping and monitoring of Svalbard

Avalanches form a threat to people travelling in mountainous regions as well as for infrastructure and buildings. They cause around 250 fatalities annually worldwide. To limit the number of future fatalities, forecasting services are interested in knowledge on avalanche activity to verify their warning system. They rely on information about the frequency, location and extent of debris fields, as provided by avalanche experts. As avalanche terrain is mostly remote and inaccessible, it can be dangerous or even impossible to obtain necessary field information. This information is especially crucial to gain during strong winds and blowing snow when an increased avalanche danger is present. By applying Synthetic Aperture Radar (SAR), large areas can be monitored at once with both high spatial resolution and high acquisition frequency. It also has the advantage of being daylight- and weather independent. The area of interest, Nordenskiold Land on Svalbard, experiences over four months of polar darkness per year. Consequently, the most applicable technique for avalanche monitoring is SAR. Avalanche debris has an increased surface roughness compared to the surrounding unperturbed snow causing a higher backscatter signal. Therefore, the debris fields appear bright in SAR images. The main goal of this research project is to optimise avalanche detection in SAR images by exploring the option of automatic detection of debris fields. Hence, we present a method to automatically detect avalanche debris fields in SAR images. It is designed and tested on both RADARSAT-2 Ultra Fine (UF) mode and Sentinel-1A Extra Wide swath (EW) mode images. Sentinel-1A has the advantage of obtaining images twice per day over Svalbard and is made available for free. Due to the high costs to acquire RADARSAT-2 data over Svalbard, these images have a low acquisition frequency. The UF mode images are geocoded to a pixel spacing of 3m compared to 40m for the EW mode images. Both modes detect the location of debris fields, but the extent is only clearly distinguishable in the UF mode images due to the fine resolution. In case of automatic detection, the backscatter coefficient of the debris fields is compared to the backscatter coefficient from a reference image. This reference image should be obtained during dry snow conditions or during a snow-free summer. The difference in backscatter coefficient between the two images is determined by subtracting the reference image from the avalanche image. By applying a threshold on the difference image debris fields are successfully located. However, to eliminate areas above the threshold value but not considered as avalanche, a filter is applied. Two filters are tested; a median filter and a Remove Small Objects (RSO) filter. For the RADARSAT-2 UF mode images the best result is obtained by using a median filter and a threshold value of 1.9dB, while for the Sentinel-1A EW mode images a RSO filter in combination with a threshold value of 3.4dB resulted in the optimum detection. None of the designed automatic detection methods resulted in a 100% probability of detection and zero false alarms, but they do confirm that automatic detection of avalanches in these SAR images is possible. They also show that the automatic detection method is depended on the characteristics of the input data. By combining a regular detection of the whole of Svalbard by the coarse Sentinel-1A EW mode images and a more specified forecasting using the fine RADARSAT-2 UF mode images avalanche maps can be created indicating both location and extent of debris fields. These maps can be of great value for avalanche warning services, although further research is necessary before automatically generated avalanche maps can be included in daily operations.Geoscience and Remote SensingCivil Engineering and Geoscience