结合CNN和Transformer的遥感图像土地覆盖分类方法

Land cover classification method for remote sensing images using CNN and Transformer

利用遥感图像进行语义分割是一种有效的土地覆盖分类方法。然而由于主流框架存在边缘分割不准确、缺乏全局信息导致错误分类等问题,阻碍了其在土地覆盖分类中的应用。针对以上问题,提出了一种用于遥感图像土地覆盖分类的卷积神经网络(Convolutional Neural Networks,CNN)和Transformer混合网络CTHNet,结合了CNN的局部细节提取能力和Transformer的全局信息提取能力。同时设计了自适应融合模块,融合来自对应级别的CNN和Transformer特征,自适应融合模块的输出进入分割头得到最终的预测结果。最后,结合边界检测分支为语义分割提供边缘约束。在两个公开的土地覆盖分类数据集上的实验结果表明,该方法优于当前主流的方法,分别实现了90.53%和64.33%的平均交并比(mIoU),对遥感图像中的大目标和边界也有更好的识别效果。

Semantic segmentation using remote sensing images is an effective land cover classification method.However,its application in land cover classification is hindered by the problems of inaccurate edge segmentation and lack of global information leading to misclassification in mainstream frameworks.To address these problems,a Convolutional Neural Networks(CNN)and Transformer hybrid network CTHNet for land cover classification of remote sensing images is proposed,which combines the local detail extraction capability of CNN and the global information extraction capability of Transformer.The adaptive fusion module is also designed to fuse the CNN and Transformer features from the corresponding levels,and the output of the adaptive fusion module enters the segmentation head to get the final prediction results.Finally,the boundary detection branch is combined to provide edge constraints for semantic segmentation.Experimental results on two publicly available land cover classification datasets show that the method outperforms current mainstream methods,achieving 90.53%and 64.33%of the mean Intersection over Union(mIoU),respectively,and also has better recognition of large targets and boundaries in remote sensing images.