Use of Data-Driven and Knowledge-Driven Methods In Disaggregating Soil Survey Maps, Okanagan Basin, Southern British Columbia.

One approach to creating raster soil attribute maps over large areas (>100,000 ha) where legacy soil mapping exists is to disaggregate individual soil series through predictive methods then assign attribute values from associated soil databases.  In this study, we utilized five different methods to disaggregate some 25 dominant soil series from harmonized 1:50,000 scale soil maps for the Trout Creek watershed within the Okanagan Basin.  Individual soil series were assigned to grid cells of a 25 m digital elevation model.  We used two data driven probabilistic methods – logistic regression and weights of evidence, two knowledge driven methods – fuzzy logic and expert (pedological) rules and one hybrid approach using a fuzzy logic calculation but using the output from weights of evidence to help define input curves.  We used ArcSIE as the inference engine for fuzzy logic, expert rule set and hybrid approaches.  For each approach we used the same set of environmental covariates (various terrain derivatives, biogeoclimatic zone and surficial geology maps) and training data obtained by mining selected polygons of the soil maps and through field observations.  Accuracy of the predictions was assessed using a set of field validation data. All approaches yielded reasonable overall accuracy although results varied considerably in predicting the spatial locations and extent of individual soil series.