An Explainable and Optimized Machine Learning Framework for Multisensor Remote Sensing-Based Land Use and Land Cover Classification

Abstract

Land use and land cover (LULC) mapping supports environmental monitoring and spatial planning, but classification performance often suffers in feature spaces that combine multiple data sources with redundant predictors. This study compares five feature selection methods (mRMR, ReliefF, Boruta, RFE, HSIC-Lasso) with three tree-based classifiers (RF, XGBoost, LightGBM) for LULC classification in Kahramankazan, Ankara, Türkiye. Combining Sentinel-1, Sentinel-2, topographic, climatic, and socioeconomic data yielded 240 candidate features. Hyperparameters were tuned with Optuna under spatial cross-validation, and SHAP was used to interpret feature contributions. HSIC-Lasso + LightGBM achieved the highest F1 (0.9655) and OA (0.9657) using only 31 features, 84% fewer than Boruta + RF, which reached a comparable F1 with 196 features. McNemar's test with Holm–Bonferroni correction confirmed equivalence with the best Boruta variants and superiority over the mRMR- and ReliefF+RF baselines. HSIC-Lasso also completed the pipeline 16.3 times faster than Boruta, supporting compact embedded selection in multi-source LULC workflows.