Optimizing swin transformer encoder for spatiotemporal desertification classification in an arid environment

Accurate land cover classification in spatiotemporal domains is crucial for monitoring environmental changes and mitigating desertification. Conventional methods often struggle with spatial and spectral limitations, necessitating advanced deep-learning approaches. This study introduces SwinCapT, a novel framework integrating Swin Transformer and Capsule Network for enhanced classification of multi-spectral remote sensing (RS) data. Evaluated on a dataset from Thatta and Badin, Pakistan (2010–2023), SwinCapT outperformed traditional models, achieving 99.8% accuracy, 0.984 precision, 0.983 recall, and a 0.982 F1-score, surpassing Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Residual Networks (ResNet), Visual Geometry Group (VGG16), and Vision Transformer (ViT). The model effectively captured spatiotemporal patterns, identifying critical land cover changes over time. Unlike conventional CNN or ViT models, SwinCapT uses dynamic feature extraction that is sensitive to both spatial and temporal variations, and hence it is well-suited for RS image analysis in environmental monitoring tasks. The comprehensive evaluation methodology established in this study provides a template for advancing transformer-based applications in environmental science while addressing critical challenges in scalability, vegetation health assessment, and operational deployment that have limited previous research efforts.