A Comparative Study of Nature-Inspired Optimization Techniques for MLP-Based Sentinel-2 Image Segmentation

Abstract. Accurate land cover classification using Sentinel-2 satellite imagery remains a critical challenge in remote sensing due to spectral complexity and spatial heterogeneity. This study presents a comprehensive evaluation of Multi-Layer Perceptron (MLP) models optimized with nature-inspired algorithms for Sentinel-2 image segmentation. We compare five optimization approaches Ant Colony Optimization (ACO), Whale Optimization Algorithm (WOA), Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Artificial Bee Colony (ABC) to enhance MLP performance for classifying five key land cover types: urban areas, agricultural fields, sparse vegetation, water bodies, and forests. Our optimized MLP architecture achieves superior performance with 90.8% overall accuracy, 90.7% F1-score, 0.883 Cohen’s Kappa, and 0.981 ROC-AUC, representing a 7.2% improvement over the best-performing nature-inspired algorithm (GA/WOA at 83.6% accuracy). Class-specific analysis reveals high accuracy for water bodies (94.2% F1-score) and forests (91.6%), while urban areas (87.4%) and sparse vegetation (82.7%) present greater challenges due to spectral similarities. The study demonstrates that hybrid optimization, combining algorithmic tuning with expert refinement, yields the most robust results for operational land cover mapping. Key findings highlight GA’s effectiveness in handling class imbalance and WOA’s strength in rare class detection. Computational efficiency (2–4 hours training time) further supports the model’s feasibility for large-scale applications. This research advances Sentinel-2 segmentation methodologies while providing practical insights for environmental monitoring, precision agriculture, and urban planning.