Comparative Assessment of Machine Learning Classifiers for Land Use/Land Cover Mapping in a Tropical Savanna Landscape

Accurate land use/land cover (LULC) information is essential for environmental monitoring, agricultural planning, and sustainable land management, particularly in rapidly changing tropical savanna landscapes. This study comparatively assessed the performance of Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Tree (CART) classifiers for LULC mapping in Sabon Gari and Zaria, Kaduna State, Nigeria, using Sentinel-2 spectral bands and derived spectral indices. Sentinel-2 imagery acquired between November 2024 and March 2025 was processed to generate ten spectral bands and four indices (NDVI, NDWI, NDBI, and BSI), which served as predictor variables. Classification was performed for six land cover classes, namely built-up areas, savanna vegetation, dense vegetation, farmland, water bodies, and rock outcrops. Model performance was evaluated using spatially independent validation data and confusion matrix-based metrics. The results revealed notable differences among the classifiers. RF achieved the highest classification accuracy, with an Overall Accuracy of 91.13% and a Kappa coefficient of 0.894, followed by CART (84.62%; κ = 0.832) and SVM (82.89%; κ = 0.809). Savanna vegetation (42.2%) and farmland (35.3%) were identified as the dominant land cover classes in the area. Feature importance analysis showed that shortwave infrared and red-edge bands contributed most strongly to class discrimination, while derived spectral indices provided additional discriminatory power. Visual assessment further demonstrated that RF produced more spatially coherent and realistic land cover patterns than SVM and CART. The findings highlight the effectiveness of Sentinel-2 data for LULC mapping and identify RF as the most suitable classifier for tropical savanna environments characterized by heterogeneous land cover assemblages.