Farmland Extraction from UAV Remote Sensing Images Based on Improved SegFormer Model

To further improve the accuracy of extracting farmland spatial distribution information, this thesis proposes an improved SegFormer model for extracting farmland spatial distribution information from unmanned aerial vehicle images. This method first introduces Efficient Channel Attention to optimize each transformer block in the encoder. Then, input the output results of each optimized block into the introduced BiFPN layer for enhanced feature extraction, and input the weighted fused multi-level features from the encoder into the decoder. By aggregating multi-level features through the Multi Layer Perceptron, local and global attention are combined, and then further weighted feature fusion is achieved through BiFPN. Finally, tthe Squeeze Excitation and Efficient Channel Attention was proposed to enhance channel features and improve model performance. The experimental results indicate that the improved SegFormer model’s mean intersection over union and mean pixel accuracy were 96.91 SegFormer model, it has increased by 1.55 union and pixel accuracy for farmland is 98.42 than other semantic segmentation models, effectively extract the extraction accuracy of farmland edges and small farmland from drone images.