Improving Land Cover Classification with Pixel-level Fusion SAR and Optical Images Using Hybrid Machine Learning

Abstract

This study evaluates land cover classification in Sungai Petani using Sentinel-1 SAR, Sentinel-2 optical data, and their fusion. Six classes were mapped: water vegetation, paddy, oil palm, urban, and roadway. A 1D CVV and a Random Forest were trained on single-sensor and fused feature sets, and a logistic stacking meta-classifier combined CNN and RF probabilities. Single-sensor baselines reach 0.76 accuracy for Sentinel-1 (macro F1 0.75) and 0.88 for Sentinel-2 (macro F1 0.85). Fusion improved performance to 0.91 accuracy for both CNN and RF (macro F1 = 0.91), with clear gains in the built classes. The stacked meta-model achieved the best results at 0.93 accuracy and macro F1 0.93, raising urban F1 to 0.83 and roadway F1 to 0.89 while maintaining high scores for water and crops, including oil palm at 0.98 F1. Results show that Sentinel-1 backscatter and Sentinel-2 spectra are complementary, and that fusion with ensemble learning yields more balanced and reliable maps than either sensor alone.