CBERS 4A/WPM Image Classification for Mapping Land Use and Land Cover in TCRAs

Tropical restoration areas present high spectral heterogeneity and spatial fragmentation, making land use and land cover (LULC) classification challenging, particularly when using medium-resolution satellite imagery. Although spectral mixture analysis, object-based image analysis, and machine learning approaches have been widely explored individually, their integrated application using freely available China–Brazil Earth Resources Satellite 4A Wide-field Panchromatic and Multispectral (CBERS-4A/WPM) imagery for monitoring Environmental Restoration Commitment Agreements (TCRAs) remains limited. This study investigated whether the integration of Linear Spectral Mixture Model (LSMM) fractions, Geographic Object-Based Image Analysis (GEOBIA), and machine learning algorithms could improve LULC classification for mapping TCRAs in the state of São Paulo, Brazil. Spectral, textural, and sub-pixel attributes derived from pan-sharpened CBERS-4A/WPM imagery were combined into a unified feature dataset, and Random Forest, XGBoost, Multilayer Perceptron, and Support Vector Machine classifiers were evaluated. Random Forest achieved the best classification performance, with an overall accuracy of approximately 0.66, demonstrating superior predictive capability compared with the other tested algorithms. Feature importance analysis indicated that LSMM-derived vegetation and shadow fractions contributed significantly to class discrimination. The integration of spectral, spatial, and machine learning approaches provides an interpretable and cost-effective framework for LULC mapping in highly fragmented tropical restoration landscapes, supporting environmental monitoring and restoration assessment.