Mapping and Characterizing Wetlands and Wetland Dynamics in the Highlands of Ethiopia using Random Forest Classification

Wetlands are recognized worldwide for the critical ecosystem services they provide and their role in maintaining livelihoods Earth observation can be utilized to assist in mapping and monitoring wetland ecosystems. This research evaluated satellite-based multispectral data (Landsat-5 TM), radar (ALOS-PALSAR) data, and terrain metrics in characterization and mapping of the Dabus Marsh, in the highlands of Ethiopia. Using the Random Forest (RF) classifier, wetland types were classified based on plant community composition and structure. RF produces independently constructed classification trees using bootstrapped samples of the original data. the output class at each pixel is the class selected by the majority of the classifications. RF models built with multi-source data yielded 94.4% and 92.9% overall classification accuracy for the dry and wet season, respectively. Seasonal differences in wetland aerial extent were only 5-6%, a level that was considered too low to be significant and mainly attributed to model errors.