Remote sensing ground object images have the characteristics of complex background and numerous varieties, and traditional segmentation algorithms will lead to edge blur, information loss and low segmentation accuracy.To solve the problems, a semantic segmentation algorithm based on the improved DeepLabV3+ network is proposed.Firstly, the improved feature extraction network CHRNet is introduced into the backbone network.Secondly, the Non-Subsampled Contourlet Transform (NSCT) algorithm is used to reconstruct the global pooling operation in the Atrous Spatial Pyramid Pooling (ASPP) module.Finally, the parameter-free attention mechanism SimAM is added in model encoding and decoding stages to enhance feature transfer among modules and improve feature utilization ratio.The experimental results show that Mean Intersection over Union (MIoU) of the improved algorithm is 81.56% on PASCAL VOC2012 data set and 64.2% on WHDLD data set, which are about 4.61 percentage points and 2.8 percentage points higher than those of the original algorithm.The improved algorithm can enhance segmentation accuracy while ensuring segmentation speed.