Aiming at the problems of high cost and low precision of traditional bare rock extraction methods in karst areas, this paper constructs a bare rock extraction method based on improved DeepLabV3+. This method first uses CA-DC-MobileNetV3 to replace DeepLabV3+ backbone network Xception in the encoder for feature extraction, which greatly reduces the amount of model parameters. Secondly, the features extracted by the encoder are enhanced through the feature pyramid network and the coordinate attention mechanism to obtain more small target information and reduce the loss of image details. Finally, in the atrous spatial pyramid pooling module, the features of the convolutional layers with different dilation rates are fused to improve the utilization of information. The results show that the method in this paper performs best in the bare rock extraction tasks in different scenarios, the number of model parameters is about 1/13 of that of DeepLabV3+, and the intersection ratio and F₁-Score are 72.46% and 84.04% respectively. Compared with the DeepLabV3+ model, the above two indicators have improved by 4.62 and 3.19 percentage points, respectively, and are superior to other commonly used semantic segmentation models, improving the accuracy of bare rock extraction.