Clouds play a crucial role in the atmospheric dynamics of the Earth, and precise segmentation of ground-based cloud images is essential for improving the accuracy of weather forecasting. In response to issues such as varying data quality, low data volume, and different capture angles in existing open-source cloud image datasets, a labeled standard ground-based cloud image dataset (Cloud-GT) was constructed using manual annotation and transfer learning methods. The color channel component threshold segmentation method was employed to eliminate sunlight interference. Furthermore, an improved U2Net-based ground-based cloud image segmentation technique was proposed. The model introduced channel attention modules and depth-wise separable convolution modules in the feature extraction unit, which greatly reduces the network model parameters while improving the effective feature extraction of ground-based cloud maps within the network. Finally, comparing and analyzing the method with classical segmentation networks, experimental results indicated that the method achieved classification pixel accuracy, mean class pixel accuracy, average intersection over union, intersection over union, and F1 score of 84.03%, 90.88%, 84.13%, 74.12%, and 89.59%, respectively. In comparison with U2Net, UNet, and FCN, the method demonstrated a significant improvement in performance. In conclusion, the method not only substantially reduced the model parameters but also effectively enhanced segmentation accuracy, which provides the possibility of practical application.