【摘要】 针对遥感图像数据本身存在分辨率高、背景复杂和光照不均等特性导致边界分割不连续、目标错分漏分以及存在孔洞等问题，提出了一种基于改进Swin-Unet的遥感图像分割方法。在编码器末端引入空洞空间金字塔池化(Atrous Spatial Pyramid Pooling, ASPP)模块，用于捕获多尺度特征，增强网络获取不同尺度的能力，充分提取上下文信息；将解码器端的Swin Transformer Block替换为残差Swin Transformer Block,不仅保留了原始信息，又能够缓解模型出现梯度弥散现象；在跳跃连接中引入残差注意力机制，可以让模型更加关注特征图中的重要特征信息，抑制无效信息，从而提高模型分割的准确率。在自建数据集上进行实验，结果表明，改进后的网络平均交并比(mean Intersection over Union, mIoU)达到了80.55%,提高了4.13个百分点，证明改进后的网络可以有效提高遥感图像分割的精度。更多还原

【Abstract】 To solve the problems of discontinuous boundary segmentation, target misclassification and missed classification, and holes caused by the characteristics of remote sensing image data itself, such as high resolution, complex background, and uneven lighting in remote sensing image data, a remote sensing image segmentation method based on improved Swin-Unet is proposed. Firstly, an Atrous Spatial Pyramid Pooling(ASPP) module is introduced at the end of the encoder to capture multi-scale features, enhance the network’s ability to obtain different scales, and fully extract contextual information; secondly, replacing the Swin Transformer Block on the decoder side with a residual Swin Transformer Block not only preserves the original information, but also alleviates the phenomenon of gradient dispersion in the model. Finally, introducing residual attention mechanism in skip connections can make the model pay more attention to important feature information, suppress invalid information, and improve the accuracy of model segmentation. After conducting experiments on a self-built dataset, the results show that the mean Intersection over Union(mIoU) of the improved network reaches 80.55%, an increase of 4.13 percentage points, proving that the improved network can effectively improve the accuracy of remote sensing image segmentation.更多还原