A dual-branch spatio-temporal Transformer for enhancing cross-regional transferability of winter wheat extraction using small training datasets

Abstract

Accurate identification of winter wheat from remote sensing imagery is crucial for large-scale agricultural monitoring. Despite the success of Transformer-based deep learning models in various fields, their application in crop identification has been limited by the scarcity of extensive labeled training data. This study proposes a dual-branch spatio-temporal Transformer (DST-Transformer) for winter wheat extraction from Sentinel-2 imagery using a small training dataset. By independently extracting temporal and spatial features, the DST-Transformer effectively delineates crop boundaries and reduces misclassification. Experiments demonstrate its effectiveness with small training datasets, achieving over 90% overall accuracy (OA) and 88.25% mean intersection over union (MIoU) when evaluating on test datasets. The DST-Transformer was further applied to large-scale winter wheat extraction across Shandong Province, China (an area 66 times larger than the training region) to evaluate its cross-regional transferability. Evaluation results showed OA over 92% and MIoU exceeding 85% at all validation sites, highlighting the DST-Transformer’s robustness and strong generalization capability. This study underscores the DST-Transformer’s potential for large-scale crop identification and illustrates the promise of Transformer-based architectures for efficient, high-precision crop mapping with small training datasets, advancing the application of deep learning in agricultural remote sensing.