Genetic algorithm optimization of ensemble learning approach for improved land cover and land use mapping: Application to Talassemtane National Park

• GA-optimized ensemble model achieved high LCLU accuracy of 96.2%. • GA optimization improved accuracy by 10.4% over the best base classifier. • McNemar test confirmed statistical gains over all other classifiers. • Method excelled at mapping spectrally similar Mediterranean land classes. Accurate Land Cover/Land Use (LCLU) mapping is critical for protected area management, yet previous studies of Talassemtane National Park (TNP) in Morocco have produced coarse maps with limited accuracy. This study proposes a Genetic Algorithm (GA) optimization method for ensemble classification to improve LCLU classification accuracy while maintaining computational efficiency. Using Sentinel-2 satellite imagery processed in Google Earth Engine (GEE), six spectral features and six vegetation indices were extracted. Multiple Machine Learning (ML) classifiers including Random Forest (RF), Support Vector Machines (SVM), Naive Bayes (NB), Classification and Regression Tree (CART), Minimum Distance (MinD), and Gradient Tree Boost (GTB), and a Grid Search (GS)-optimized ensemble-were evaluated. The GA optimization framework was implemented to tune the ensemble classifier’s hyperparameters, outperforming the GS method. The GA-optimized ensemble classifier achieved an overall accuracy of 96.2%, significantly higher than all individual classifiers, whose accuracies ranged from 77.5% to 85.8%. McNemar’s statistical test showed that GA-optimized ensemble performance is statistically significantly different across all pairwise comparisons. The GA-optimization also demonstrated strong capability in distinguishing spectrally similar classes typical of Mediterranean environments, such as built-up areas, limestone surfaces and matorral. Analysis of GA parameters revealed optimal ranges for population size (90–100), number of generations (60–70), crossover probability (0.8), and mutation probability (0.1–0.15) that balanced accuracy and computational efficiency. This research demonstrates that GA-optimized ensemble ML represents a robust approach for improved LCLU mapping from RS data, providing valuable insights for conservation planning and protected area management, and is broadly applicable to other environmental and land classification contexts.