U-Shaped Network Based on Deformable Convolutional Encoder for Semantic Segmentation of Winter Wheat

The distribution of winter wheat in mountainous regions is fragmented and exhibits irregular shapes. Conventional networks tend to lose the fragmented distribution of winter wheat and struggle to accurately predict the distribution edges. To address this issue, this study improves the encoder part of the UNet network by replacing ordinary convolutions with deformable convolutions to achieve flexible extraction of local features. This improvement is beneficial for extracting features from fragmented and irregularly distributed winter wheat. To leverage the advantages of multi-spectral imagery, a bi-temporal remote sensing dataset of winter wheat was created. The bi-temporal data provides more spectral information for the model. In order to model the correlation between the extracted feature channels and the importance of each channel for the segmentation task, a channel attention module called SEblock was added at the bottom of the network. Finally, comparative experiments demonstrate that DUNet effectively locates winter wheat under complex terrain conditions.