Ground-based observation is important for solar physics research. However, cloud cover between the Sun and the Earth contaminates and degrades the quality of the acquired solar observation images obtained, thereby affecting the research on solar physics. Therefore, these images need to be screened and categorized. Previously, traditional methods and manual screening were capable of handling this task. However, with the advancement of technology, the observation equipment is now capable of collecting data of terabyte scale. The existing methods are unable to handle the data obtained by the observation equipment efficiently, and thus are no longer applicable. We adopted an improved deep learning method based on U-Net, integrating attention mechanisms, convolutional neural networks and feature fusion mechanisms to solve the problem of cloud coverage assessment in full-disk solar images. We utilized a dataset comprising 4350 pieces of data captured by the Solar Full-disk Multi-layer Magnetograph (SFMM), selecting precision, recall, F1 score, and macro F1 score as evaluation metrics, and evaluated the performance of our model along with the comparative experimental models through five-fold cross-validation. The Macro-Precision of our model reaches 0.9782, the score of our model on Macro-F1 is 0.9782. These metric values demonstrate that our method achieves more accurate classification of ground-based observation data.