Mapping of winter wheat with UNet-Transformer fusion model based on Sentinel-2 A imagery and field observation data

Accurate high-resolution winter wheat mapping remains challenging due to crop spectral similarity and complex agricultural landscapes. This study proposes a novel framework integrating object-oriented multiscale segmentation with deep learning to enhance winter wheat identification. Our method leverages object-oriented multiscale segmentation to optimize image annotation and feature representation, which is then integrated with semantic segmentation networks for precise crop classification. Using 24 scenes of Sentinel-2 A imagery and 7280 field observation points in the Guanzhong Plain, we systematically compared three deep learning architectures (SegNet, DeepLab V3+, and UNet). The results initially demonstrated the superiority of UNet, which achieved best winter wheat identification accuracies of 86.04%. The corresponding F1-scores were 90.66% for UNet. Compared to UNet, the UNet-transformer model achieved 5.12% higher identification accuracy and 1.94% higher of F1-score. UNet-Transformation based identified winter wheat areas are 7935.94 km², 7670.15 km², and 8695.18 km² in 2018, 2020, and 2023 years in the Guanzhong plain. The obtained TA values of 94.87%, 92.44%, and 98.21% when compared to government statistical data, indicating that the winter wheat map produced at 10 m resolution was reliable overall. These findings demonstrate that the proposed integration of object-oriented multiscale segmentation and advanced deep learning offers a robust practical solution for large-scale, high-resolution crop mapping.