Modeling and performance evaluation of machine learning for historical and future land use and land cover dynamics in Chattogram District, Bangladesh

Introduction Land use and land cover (LULC) change is rapidly redrawing the environmental landscape of coastal Bangladesh. Methods This study integrates Landsat-based classification, machine learning-based approaches, and Markov Chain simulation to assess historical and future LULC dynamics in Chattogram District. Results Random Forest classification achieved overall accuracies of 95.96%, 90.45%, and 94.25% for 2006, 2015, and 2025, respectively. Results reveal substantial cropland expansion (+32,375 ha) and continued growth of built-up land, accompanied by declines in vegetation, waterbodies, and bare land. Random Forest demonstrated strong predictive performance, achieving an out-of-bag (OOB) accuracy of 92.63% and a skill score of 0.9171. By 2035, cropland and built-up land are projected to increase to 43.95% and 4.63%, respectively, while vegetation, waterbody, and transitional land-cover classes continue to decline. Discussion These findings provide empirical evidence to support sustainable land-use planning and environmental management in coastal-urban systems.