Multisource evidential classification of surface cover and frozen ground

Abstract

An evidential reasoning classification system (MERCURY©) has been developed to address some of the methodological limitations or conventional image analysis strategies applied to larger, multisource data sets. The software can process any number of variables at different levels of measurement without restrictions of underlying statistical models, and it can also incorporate measures of uncertainty into the analysis. A new frequency-based technique to compute evidence from training data has been implemented, and a modified decision rule is introduced for use with Dcmpstcr-Shafer orthogonal combination of evidence. These ideas were tested for a complex environment in northern Canada where the occurrence of permafrost (perennially frozen ground) was classified using field and remotely-sensed measures of landcover, equivalent latitude (potential insolation), and terrain aspect. Evidential classification of permafrost from field observations was 85 per cent in agreement with soil probe field determination of frozen ground at over 400 field sites. When the same variables were obtained from image processing and evidential classification of SPOT imagery and a digital elevation model, a permafrost classification accuracy of 82 per cent was obtained. These tests show a successful monitoring application with nominal variables that is difficult or impossible to consider using conventional statistical and image analysis techniques. They indicate a possibility for the detection and monitoring of permafrost using digital remote sensing data alone.