High-resolution unmanned aerial vehicle remote sensing images have extremely rich semantic and ground feature features, which are prone to problems such as incomplete target segmentation, missing edge information, and insufficient segmentation accuracy in semantic segmentation. To solve the above problems, based on DeepLabV3_plus model, an improved DeepLabV3_ DHC is proposed. First of all, multiple backbone networks are used for down-sampling to collect low-level and high-level features of the image. Second, the atrous spatial pyramid pooling (ASPP) of the original model is replaced by a depthwise separable hybrid dilated convolution, and an adaptive coefficient is added to weaken the mesh effect. After that, the traditional sampling bilinear interpolation method is abandoned and replaced by the learnable dense upsampling convolution. Finally, cascade attention mechanism in low-level features. In this paper, a variety of backbone networks are selected for the experiment, and some images of Longchang City, Sichuan Province are selected for the dataset. The evaluation index uses the average intersection and combination ratio and the average pixel accuracy of the category as the reference basis. The experimental results show that the method in this paper not only has higher segmentation accuracy, but also reduces the amount of computation and parameters.