Interpretable water and agricultural land cover in urban landscapes using CNN–ViT framework with Grad-CAM for agricultural sustainability

Abstract

Abstract Accurate land-use and land-cover classification in hyper-arid environments is challenging because several surface types can exhibit overlapping spectral and spatial characteristics in Landsat imagery. These surfaces include water, irrigated agriculture, moist soil, barren land, and built-up areas. This study proposes an interpretable hybrid framework for classifying six land-use and land-cover classes in Najran, Saudi Arabia, using seven-band Landsat 8 imagery. The framework integrates MobileNetV2, ResNet152, and ViT-B/16 to combine efficient local feature extraction, deep hierarchical representation learning, and global contextual modeling. This combination was selected to address the heterogeneous landscape patterns and mixed spectral responses that characterize the study area. Controlled ablation experiments showed progressive performance gains from the standalone CNN models to the CNN–ViT configurations and the complete three-branch framework. Overall accuracy increased from 86.73% to 88.23% for MobileNetV2 and ResNet152 to 94.98% and 95.68% for their respective CNN–ViT configurations. The complete framework achieved the highest overall accuracy of 98.70%, with a Kappa coefficient of 0.9827, on a geographically separate assessment partition. The F1-scores were 99.02% for Water Area, 98.38% for Agriculture, 98.64% for Built-up Area, and 99.16% for Barren Land. Other Land and Unclassified achieved F1-scores of 96.03% and 91.60%, respectively. Grad-CAM produced spatially coherent activation patterns over mapped water and cultivated surfaces. These patterns provided qualitative evidence that the model relied on landscape structures relevant to the predicted classes. The findings demonstrate that combining complementary local, hierarchical, and global representations improves both classification accuracy and model interpretability. The proposed framework therefore provides a reliable basis for land-use and land-cover monitoring in Najran and for further validation in other hyper-arid environments.