Snowfall is a complex weather phenomenon that involves multiple scales of snowflakes, complex motion trajectories, and a foggy effect in the long term. However, existing single image snow removal methods and datasets do not take into account these physical properties, which limits the performance of snow removal methods on real snowfall images. To address this issue, we first constructed a snow imaging model that combines snowflakes and fog based on the depth information of the image, and produced a realistic synthetic snowfall image dataset. Then, we designed an end-to-end image restoration network that cascades the defogging and snow removal modules for joint training. The snow removal module adopts a double U-shaped network to solve the problem of multi-scale snowflakes and multi-directional trajectories. Finally, the experimental comparison between our method and other state-of-art methods proves that our method has very good performance in both synthetic datasets and real images, and benefits subsequent visual tasks.