Distributed snow process modelling: An image processing approach

An approach to spatially distribute a snow process model by segmenting images of land cover, terrain and snow properties is reported. A small 1.7 ha study area with an existing database was selected for this preliminary evaluation. The methodology was carried out over a relatively flat valley bottom at Camp Grayling, Michigan. Meteorological measurements on two sides of the area showed only small differences, so uniform meteorological variables were assumed over the site. Initial snow cover conditions were reconstructed and were distributed over the area using snow maps and sparse snow pit measurements. One metre resolution terrain, soil, vegetation and snow type maps were individually processed into class maps. These layers were then combined to produce a segmented class map, where the attributes from the data layers were known for each class. A one-dimensional model of snow processes was run for each class, then the results were mapped back into images. Shallow snow conditions provided high sensitivity of ablation patterns to meteorological conditions over a 72 h period. The model performance was assessed by comparing predicted and observed ablation patterns. The error in total snow-covered area was less than 9%. However, the location errors were greater (predicted snow where no snow was observed and observed snow where no snow was predicted). Extensive error analysis was not justified because of the lack of multiple point measurements of snow properties.