Aiming at the field of power automation, this paper considers the problem of general extraction effect caused by the unbalanced distribution of ground object categories and the large differences in scene styles in different domains in the process of extracting key feature information from remote sensing images, and an improved Deeplabv3+ semantic segmentation network is proposed. The instance-batch normalization (IBN) module is used in the backbone network ResNet101 to enhance the model’s generalization ability for remote sensing images with large differences in styles. The squeeze-and-excitation (SE) module is added to the network to strengthen important channel information, alleviating the problem of information loss. The joint loss function of Dice Focal is adopted as the loss function, and dice loss can alleviate the impact of imbalanced category distribution on the extraction of small targets. Focal loss can not only make the model pay more attention to objects that are difficult to classify, but it can also improve the instability of network training caused by dice loss. Experimental results show that in comparison with original Deeplabv3 model, the improved Deeplabv3 improves F1-Score by 7.78% and Intersection over Union (IoU) by 5.78%. In comparison with other mainstream semantic segmentation models (including FCN, UNet, and SegNet), the improved Deeplabv3+ also achieves better segmentation accuracy in ground objects extraction.