Accurate and temporally consistent land use and land cover (LULC) mapping is important for environmental monitoring, ecosystem evaluation, and sustainable land management, especially in mountainous areas with intricate topology and spectral heterogeneity. Nonetheless, most existing machine learning and deep learning approaches have limited spectral interpretability and temporal instability, which undermine multi-temporal change detection. This research study proposes a feature-aware spectral learning framework of pixel-wise multi-temporal LULC classification based on Sentinel-2 surface reflectance data. In a common experimental environment, four pixel-wise spectral learning models, including Feature-Aware Convolutional Neural Network (FA-CNN), Feature-Aware Kolmogorov–Arnold Network (FA-KAN), Feature-Aware Extended Vision Transformer (FA-ExViT), and the proposed Feature-Aware Spectral Vision Transformer (FS-ViT), are comparatively evaluated. While FA-CNN learns local spectral responses, FA-KAN learns nonlinear feature interactions, and FA-ExViT learns global spectral dependencies with self-attention, whereas FS-ViT learns feature-aware spectral tokenization, adaptive spectral reweighting, and coarse-to-fine hierarchical supervision to increase class separability and temporal consistency. The framework is used for multi-temporal LULC mapping 2019, 2021, 2023, and 2025 of the mountainous Coonoor regions. Findings indicate that FS-ViT is consistently superior to comparison models, achieving overall accuracies between 96.46% and 98.07% having Kappa coefficients above 0.93. Explainable artificial intelligence (XAI) exhibits the physically significant and temporally consistent spectral feature contributions. Multi-temporal change analysis shows a net forest loss of 6.31 km 2 , primarily transitioning to low vegetation, alongside vegetation recovery and limited urban expansion. Overall, the proposed FS-ViT framework offers a decipherable and time-resilient solution for long-term LULC mapping and change detection in environmentally sensitive mountainous landscapes.
Devi et al. (Sun,) studied this question.