Urban Land Cover Classification Using Earth Observation (EO) Data and Machine Learning Models

Rapid global urbanization presents complex challenges that require accurate and reliable urban land cover information. This chapter uses two decision tree models to map urban land cover by leveraging data from Sentinel-1 and Sentinel-2 imagery. The first decision tree model, employing 33 predictor variables, faced constraints in classification accuracy due to significant correlations among the predictors. In contrast, the second model streamlined the approach by utilizing a concise set of eight uncorrelated variables, achieving better land cover map accuracy. To further enhance the models&s; effectiveness and transparency, we employed explainable machine learning techniques to elucidate the contribution of individual features to the model decisions. These methods include feature importance, SHAP (SHapley Additive exPlanations) values, and BreakDown profiles. The two decision tree models showcased the value of a data-centric strategy for evaluating model performance. Despite using an interpretable decision tree method, the models exhibited certain limitations, notably their tendency to oversimplify complex data structures and overfit, resulting in only moderately accurate land cover maps. These shortcomings underscore the need to improve the decision tree models. We propose a direction that focuses on the potential for integrating additional data sources and employing more sophisticated decision tree algorithms. This evolution aims to create more robust and interpretable decision tree models, improving urban land cover mapping tools.