Evaluation of machine learning algorithms in land use/land cover change detection using remote sensing and GIS for Moulvibazar, Bangladesh

Due to the adverse effect of climate change, our landscape is changing day by day. Therefore, accurate land use land cover (LULC) classification is essential for the sustainable and useful management of natural resources. This research aims to compare the LULC classification performance of three distinct machine learning techniques—Support Vector Machine (SVM), Random Forest (RF), and Classification and Regression Trees (CART)—within the Google Earth Engine using Landsat 8 satellite imagery for Moulvibazar District. To determine the most precise and high-performing classifier, the confusion matrix, overall accuracy, and Kappa coefficient were systematically computed. The performance measures consistently ranked RF as the best-performing model with an optimal general accuracy of 96.9% and a Kappa Index of 0.96, which indicates very good reliability to the ground truth. Second, the CART classifier had similar results to confirm satisfactory reliability with an overall accuracy of 93.4% and a Kappa Index of 0.91. The SVM classifier also showed good results but was ranked below the previous classifiers with an overall accuracy of 81.9% and a Kappa Index of 0.76, which refers to moderate classification capability. Lastly, this research intends to model different LULC changes for the period of 2014 to 2024 using its best-performing algorithm, RF, for that specific region with multiple bands of Landsat 8 satellite imagery. The water body grew by 1.4%, indicating effective management of the water resource. The decrease of barren land was 2.3%, suggesting that vegetation restoration has taken effect. Small-scale clearance reduced vegetation by 4.4%. But the built-up area grew by 2.0%, possibly due to more sustainable urban planning. Agricultural land use rose by 3.2% due to expanding farming. The district's dynamic land use has boosted the economy and ecosystem. This research will contribute to the scientific community by establishing criteria for optimally choosing both the Machine Learning algorithm and the satellite data source to ensure the development of reliable LULC maps.