Improving Aerial Imagery Classification of Supraglacial Features with a DEM-Based Shadow Modeling

Abstract:

The presence of shadows in remotely sensed images of the cryosphere can be highly problematic for classifying snow, ice, debris, and hydrologic features. Many methods have been developed for detecting and removing shadows from high resolution imagery including multi-spectral image analysis, geometric models, or shaded relief methods. These methods though often perform poorly for dark objects, require extensive input information, or do not account for shadows cast on adjacent terrain. Instead, we developed an easily implemented method of removing topographic shadows from remotely sensed images that uses readily available GIS software, corrects for cast shadows, reduces the amount of over-correction, and can be performed on imagery of any spectral resolution. This method was demonstrated on orthomosaiced drone imagery processed with Structure-from-Motion software of a supraglacial stream catchment in southwest Greenland. The landscape has highly variable reflectance values due to supraglacial drainage networks, dark sediment, and bright ice as well as substantial shadowing due to low sun angles and steep topography. Accurately classifying land cover in this region is vital for predicting the surface energy balance of the Greenland Ice Sheet. A supervised classification scheme was applied to the corrected and original image to determine the location of bare ice, sediment, and water resulted in a significant improvement in overall classification accuracy. The correction caused a substantial increase in the spatial coverage of sediment and stream features therefore imagery used in classification that does not correct for shadowing may significantly underestimate the coverage of low albedo features.