Selection of Attributes by Meta-heuristics for Classification of Illegal Mining via SAR Sentinel-1 in the Amazon

Illegal mining classification in the Amazon faces challenges due to the high dimensionality and noise of Sentinel-1 SAR data. This work investigates multi-objective Feature Selection (FS) to optimize the trade-off between predictive performance and model parsimony. A dataset containing 64 attributes was built to compare four feature selection paradigms: Principal Component Analysis (PCA), LASSO, deep methods (LassoNet and DeepFS), and metaheuristic-based wrappers (Simulated Annealing and NSGA-II), evaluated using Random Forest with stratified cross-validation. A control experiment using all 64 variables was included as a baseline. NSGA-II achieved an F1-Score of 0.8843 using only 17 attributes, while a 12-attribute solution preserved 99.4% of the best performance. The results indicate that compact subsets can reproduce the performance of the full model with lower computational cost for environmental monitoring in the Amazon.