Dual Attention Pathway for Semantic Segmentation of Agricultural Remote Sensing Images

In the task of segmenting farmland from remote sensing images, challenges often arise due to significant intra-class variability and excessive inter-class similarity among different types of farmland. Common semantic segmentation methods exhibit certain limitations in distinguishing various farmland types and identifying boundaries between fields and roads. Against this backdrop, a novel semantic segmentation network for farmland remote sensing images, named T_segformer, is proposed. Methodology: The Dual-Path High-Low Frequency Feature Extraction Module (DPHL) employs dual attention mechanisms for extracting high and low-frequency features, thus enhancing the model's ability to recognize and segment edges of varying categories in remote sensing images. The Mixed Gated Linear Feedforward Network (MGLFFD) improves the model's feature extraction capability by boosting the feedforward network's attention to and understanding of input data. Results: Implementations on the LoveDA Dataset and the Barley Remote Sensing Dataset reveal that the T_segformer network achieves a mean Intersection over Union (MIoU) of 55.1% and 79.10% and a Mean Accuracy (mAcc) of 66.65% and 86.10%, respectively. Its segmentation performance significantly surpasses that of RSSForm, Segformer, Swin-UperNet, DeepLabV3+, and FCN models in both evaluation metrics and visual prediction outcomes. T_segformer excels in accurately recognizing distinct features within complex remote sensing images, differentiating between similar farmlands, and ultimately enhancing target boundary segmentation accuracy.