Comparative evaluation of machine learning algorithms for land use land cover mapping in Sankhuwasabha District, Eastern Nepal

Accurate land use/land cover (LULC) mapping in the Himalayan region remains challenging because of its rugged terrain, steep elevation gradients, heterogeneous landscapes, and high spectral variability. This study aims to evaluate the six commonly used supervised classification algorithms in Sankhuwasabha District of Eastern Nepal using open-source Landsat and Sentinel-2A satellite imagery. The algorithms considered for evaluation are Maximum Likelihood Classification (MLC), Support Vector Machine (SVM), Parallelepiped (PP), Artificial Neural Network (ANN), Random Forest (RF), and Minimum Distance (MD). Sankhuwasabha District in eastern Nepal was chosen as the evaluation site due to its complex terrain (elevation range 247–8485 m) where field visit was conducted in the north–south and east–west transect to collect the training and test samples (1100 sample points). Additional 207 ground-truth data were added from the high-resolution Google Earth Imagery of the year 2023. The images were classified into water body, cultivated area, vegetation, snow/ice, barren land and built up area. For the accuracy assessment of the classified image and evaluation of the algorithms, producer, user and overall accuracy, and kappa coefficient was used. Results show that RF provides the highest overall accuracy for both the Sentinel 2A (88.73%, Kappa = 0.81) and Landsat (84.12%, Kappa = 0.73) followed by SVM (72.73%) for Sentinel and ANN (72.27%) for Landsat, while PP provides the lowest overall accuracy value for the both satellite data. Although, RF algorithms better fits for the complex mountain landscapes, the appropriate algorithm is sensor-algorithm dependent rather than the resolution of image alone. By systematically comparing multiple machine learning and conventional classifiers across two widely used satellite sensors in a highly heterogeneous Himalayan landscape, this study provides guidance for selecting appropriate classification algorithms for accurate LULC mapping in mountainous regions.