Boreal biome wetland classification using multi-seasonal EO data on gee and machine learning optimization/XAI modelling

The objective of this study was to map wetlands representing 2017-2022 conditions in two areas in Canada's Boreal Forest, specifically Nunavut's taiga shield ecozone and Saskatchewan's boreal shield ecozone. Wetland classification was performed by leveraging machine learning (ML) modelling on the Google Earth Engine (GEE) JavaScript API, and Python programming for model cross validation/optimization and explainbility. Robust wetland coverage estimates were derived by summarizing the predictions of several ML models (i.e., an ensemble) trained on different sample subsets. Moreover, this study employed two key components to improve the wetland mapping modelling: (1) the processing of multi-temporal, multiseasonal (summer and fall) satellite imagery, since timeseries observations capture shifting eco-hydrological conditions, and (2) the use of Grey Level Co-occurrence Matrix (GLCM) textural variables, which are understudied in the GEE literature. Models were interpreted using Shapley explainable AI (XAI) methods. Overall accuracies of > 95% suggest this is a promising methodology.