Land Cover Classification Using Machine Learning With Image Fusion

This study evaluates pixel-level fusion of Sentinel-1 SAR and Sentinel-2 optical data for land cover classification using machine learning. A region-specific dataset with high-resolution labels for six classes - Water, Vegetation, Paddy, Oil Palm, Urban, and Roadway - was developed. Three models were compared: a Random Forest (RF), a regularized 1D Convolutional Neural Network (CNN), and a stacking ensemble combining both via logistic regression. Pixel-level fusion enhanced class separability across heterogeneous landscapes, enabling all models to exceed 90 % accuracy. RF achieved the highest performance ($\sim 93 \%$), particularly in resolving Urban and Roadway classes, and produced spatially consistent outputs. The CNN showed competitive performance but was limited by lack of spatial context. Stacking offered minimal improvement due to overlapping errors between base models. These results highlight the effectiveness of fusing multi-sensor data at pixel level and the continued strength of interpretable, classical models like RF in structured remote sensing applications.