Automated discrimination of upland and wetland using terrain derivatives

Wetlands are complex ecological systems that result from a number of environmental factors including hydrologic, geomorphic, and biologic processes. As a result of this complexity, wetlands, as a single feature, are often heterogeneous, making traditional image-based classification techniques for wetland mapping problematic. In this study, we propose a wetland classification approach that uses a combination of terrain-based derivatives to account for variability in ecological characteristics. A site was studied near Bolton, Ontario (43.84 degrees N, 79.68 degrees W). Based on research and experience, the authors propose a two-step multisoftware (i.e., geographic information system (GIS), statistical, and image processing) technique, with step one involving the delineation of wetland boundaries using topographic data, followed by the separation of wetlands by type (i.e., marsh, swamp, fen, bog). This paper focuses on the exploration of three wetland boundary delineation methods: (i) a visual derivative image threshold, (ii) a logistic regression model, and (iii) a classification and regression tree ( CART) model. Each method was applied to a number of spatially distributed digital elevation model (DEM) terrain derivatives originating from a photogrammetrically derived 5 m DEM. Of the three methods investigated, the results showed the CART approach provided the best mapping, with a calibration accuracy of 90% correct and a validation accuracy of 84%. The CART approach classified the wetlands based on the degree of terrain complexity, thus allowing for subcomponents of the wetlands to be evaluated in further detail.