Fig. 9

Visualization of motion fields extracted by the proposed method. The first row shows the input, the second row presents the predicted motion field at time step T generated by our neural network, and the third row illustrates the subsequent motion field update using Burgers’ Equation. The results demonstrate the model’s ability to capture both large-scale displacement patterns and fine-grained boundary dynamics across diverse precipitation events. Notably, even under sparse precipitation conditions, the extracted fields preserve structural coherence, highlighting the robustness of the learned motion representation.