Research on road extraction based on high-resolution remote sensing imagery and enhanced UNet2+

In response to the challenges faced in road extraction from high-resolution remote sensing images, this study proposes an improved UNet2+ road extraction method that incorporates Transformer. Building upon the UNet2+ architecture, Transformer modules are explicitly inserted between the 3rd to 5th levels of the encoder and the 3rd to 5th levels of the decoder to enhance the model's ability to capture long-range dependency features. The activation function is optimized to Swish, and ablation experiments are conducted to validate the performance differences among various attention mechanisms (Transformer, CBAM, SE). Simultaneously, the dataset enhancement strategy is improved by integrating Generative Adversarial Networks (GAN) to expand data diversity, with the quality of generated samples evaluated using Fréchet Inception Distance (FID) and Structural Similarity (SSIM). Experimental results demonstrate that, compared to traditional methods and the original UNet2+, this method achieves an 11.12% increase in Intersection over Union (IoU) and a 12.44% improvement in recall on the DeepGlobal road dataset. Moreover, compared to using GAN or Transformer modules alone, their synergistic use further enhances overall accuracy by 3.2%. Additionally, the model's parameter count is controlled at 8.9 million, with an inference time of 0.8s per image (RTX 3080Ti), providing a more advanced technical solution for road extraction and related fields.