Cloud amount calculation methods for ground-based cloud play an important role in accurate weather forecasting and climate model construction in the meteorological field. At present, the method for calculating cloud amount during the day has been relatively complete. However, due to the influence of light, ground-based cloud detection at night is difficult to distinguish the outline and size of clouds, which increases the difficulty of calculation, and there are few related studies. To address this problem, we propose a night cloud amount calculation method based on transfer learning. First, the nighttime cloud image is processed with contrast enhancement; secondly, transfer learning technology is used to transfer the weight parameters of the U-Net network model to the night cloud detection task; finally, on the basis of the original network model, two-layer skip connection is used to enhance the extraction of nighttime cloud layer by the model. So as to improve the accuracy of cloud calculation. Experimental results show that compared with the FCN model and the original U-Net model, the night cloud amount calculation based on transfer learning performs better in cloud image segmentation. Its detection speed is only half that of the U-Net model, and the average calculated error is reduced from 11.32% for the FCN model and 10.56% for the U-Net model to 8.62%, significantly improving the accuracy of the calculation results.