Automatic avalanche mapping using texture classification of optical satellite imagery

Detection and characterization of an avalanche is important for making avalanche inventories as well as for the management of emergency situations. Problems related to field measurements include remote or inaccessible terrain, poor weather, and avalanche danger. Earth observation satellites therefore represent a potentially important source of information. We present a framework for automatic detection of avalanches in very-high resolution optical satellite imagery. The approach builds upon an initial texture segmentation stage using directional filters to enhance avalanche snow in order to separate it from other relevant area cover types, such as rough and smooth snow surfaces, trees and rock. The directional filters are oriented in the same direction as the terrain aspect, which is estimated from a digital elevation model. In the training stage, filter responses corresponding to the same texture class are used to form so-called textons. We then perform pixel based classification of the image based on the distribution of textons within a sliding window. In the second stage of the detection algorithm, pixels are grouped according to texture class to form image objects, which are finally post-classified based on extracted object features. We have assessed the mapping abilities of our algorithm on a set of QuickBird images of Norwegian mountain areas. The automatically derived avalanche maps have then been compared to manually drawn avalanche outlines made by experts. The preliminary results show that we are able to locate most of the fresh avalanches in the image with few false detections, but that the outline of the avalanche is not always adequately determined.