Transformer-Based Spatio-Temporal Change Detection Network Using Satellite Image Time Series: A Case Study of Forest Disturbance in Trentino, Italy Following the Vaia Storm

High-resolution Satellite Image Time Series (SITS), combined with advanced deep learning approaches such as transformers, enable reliable change detection by effectively using large amounts of temporal data. While most existing change detection methods overlook temporal dynamics and focus primarily on spatial features, Transformers are well-suited to capture both spatial and temporal features. This paper presents a Transformer-based spatio-temporal architecture for multi-annual change detection that integrates two key components: an Adaptive Swin Transformer Encoder (ASTE) for hierarchical feature extraction and a Feature Difference Module (FDM) to highlight changes for generating change detection maps. ASTE incorporates a Cross Attention Fusion (CAF) module, which integrates convolutional operations for spatial feature extraction and Swin Transformer to extract spatio-temporal dependencies. FDM refines the difference features, enhancing feature interaction for optimization and accurate change detection. The method was tested in the Trentino region of Italy, focusing on deforestation caused by the Vaia storm.