The study area covers three major Polish cities, Warsaw, Krakow, and Wroclaw, analysed on a 1 km2 regular grid. Ground truth labels are derived from the CORINE Land Cover 2018 dataset using a dominant-class assignment rule. Two models are trained and compared: a tuned Random Forest and a cross-validated Artificial Neural Network. Model performance is evaluated using overall accuracy, Cohen’s Kappa, per-class F1 scores, and Moran’s I applied to the spatial distribution of prediction errors. The spatial error analysis is included because classification errors in geographic data are rarely independent: if a model fails in one location, neighbouring cells are often affected by the same structural problem, and identifying such clustering provides diagnostic information beyond what aggregate accuracy metrics reveal.