Explainable vision Mamba–Kolmogorov–Arnold learning for multi-temporal monitoring of vegetation stress and land degradation in drylands

Abstract Monitoring vegetation stress and land degradation in dryland environments is challenging because short-term spectral variability can obscure persistent ecological change. Complex deep-learning architectures also do not necessarily provide higher predictive performance or independently validated explanations. The study developed and evaluated an explainable Vision Mamba (VMamba)–Kolmogorov–Arnold Network (KAN) framework for multi-temporal vegetation monitoring in Badr Al-Janub, Saudi Arabia, using Landsat observations from 2015 to 2025. Multi-temporal vegetation and moisture indices were integrated with terrain variables, while a geographic partition defined training, validation, and locked independent-test subsets. On the locked independent-test set, VMamba–KAN achieved an Overall Accuracy (OA) of 78.76%, frequency-weighted Intersection over Union (FWIoU) of 71.28%, mean Intersection over Union (mIoU) of 20.06%, and Macro-F1 of 25.11%. The CNN baseline achieved higher OA, FWIoU, mIoU, and Macro-F1 values than VMamba–KAN. Cohen's Kappa was 0.424 for VMamba–KAN and 0.404 for the CNN baseline. The Kappa difference is reported separately and is not interpreted as evidence of overall or class-balanced predictive superiority. Temporal analysis identified broad improvement in vegetation condition toward 2023 followed by partial deterioration in 2025 rather than continuous landscape-scale degradation. Integrated Gradients provided descriptive attribution patterns associated with VMamba–KAN predictions. Full-scene softmax probabilities provided descriptive predictive-confidence information and were not interpreted as calibrated uncertainty estimates. The results emphasize temporal persistence, geographic evaluation, class imbalance, predictive performance, and cautious interpretation of attribution evidence.