Evaluation of CA-Markov and MLP-based machine learning approach for long-term land use land cover prediction in Himalayan mountain basins

Land Use Land Cover (LULC) processes strongly impact ecological sustainability, particularly in fragile mountain basins. Reliable forecasting of these changes is essential for supporting land management, hazard mitigation, and regional planning. However, LULC prediction remains constrained by environmental and socio-economic drivers, as well as data inconsistencies, which introduce uncertainties into model projections. This study comparatively assesses the predictive capabilities of Cellular Automata-Markov (CA-Markov) and Multi-Layer Perceptron (MLP) for projecting future LULC patterns in the Jhelum Basin. Preliminary LULC maps were generated using a Support Vector Machine (SVM) classifier to develop training and transition datasets. The MLP was further optimized to predict pixel-based transition potentials using historical LULC maps and a set of environmental and socio-economic drivers identified through Cramér’s V analysis. Topographic variables, particularly elevation (0.7175) and slope (0.6520) showed greatest influence on transition potential. CA-Markov demonstrated high accuracy for static classes but overestimated dynamic categories, including cropland (+ 45.92%). In comparison, the MLP model exhibited higher accuracy for both stable and transitional classes. The findings were validated using area deviation, Kappa statistics (MLP: Kstandard = 0.80, Klocation = 0.82; CA-MC: Kstandard = 0.90, Kno = 0.89), ground truth comparison (Kappa: MLP = 0.85, CA-MC = 0.83), and spatial agreement analysis, indicating that the MLP demonstrated superior performance. This study highlights the importance of driver sensitivity and modelling choices, while future work may integrate climatic variables with CNN-based or hybrid approaches to better capture climate-sensitive land-use transitions and improve long-term LULC prediction robustness in Himalayan Mountain basins.