A novel spatio-temporal vision transformer model for improving wetland mapping using multi-seasonal sentinel data

Wetlands mapping using remote sensing data is a challenging task due to the spectral similarity of wetlands, the fragmented nature of these landscapes, and seasonal variations in wetlands. To address these limitations, this study proposes a novel spatio-temporal vision transformer (ST-ViT) model for an accurate wetland classification using seasonal data. The ST-ViT model was trained using multi-seasonal Sentinel-1 (S1) and Sentinel-2 (S2) data acquired during the spring, summer, and fall of 2020 in a study area located in Newfoundland and Labrador, Canada. The performance of the ST-ViT model was evaluated against the validation dataset, achieving an overall accuracy (OA) of 0.950 and F1-score (F1) of 0.934, outperforming other deep learning models such as random forest (RF), hybrid spectral network (HybridSN), etc. The model demonstrated strong classification capabilities among most wetland classes, with some challenges in distinguishing between spectrally similar classes like bogs and fens. Moreover, the integration of spatio-temporal features enabled the reduction of feature mixing between wetland classes, particularly during different seasons. The ST-ViT model provides an accurate wetland distribution map in different seasons, supporting critical decision-making processes related to wetland conservation and environmental monitoring.