A video rainfall calculation method based on stacked residual time convolutional network

To enhance the density of the "Three Defense Lines" rainfall monitoring network, this study leverages the widely distributed surveillance video network data and proposes a novel rainfall estimation method utilizing video data. By fully exploiting the temporal dynamic features in rainfall video frame sequences, a hybrid CNN+TCN rainfall estimation model is established, integrating the RegNetY backbone network with a stacked TCN architecture. Experimental results demonstrate that the model achieves excellent performance in rain gauge comparisons, with goodness-of-fit metrics (R², N_SE and K_GE) all exceeding 0.976, and optimal error evaluation metrics (M_AE and M_APE) reaching 0.799 mm/h and 3.79%, respectively. The multilayer residual TCN structure effectively enhances temporal feature extraction, maintaining stable rainfall intensity estimation performance under varying frame sequence lengths. This study provides a lightweight technical solution for high-precision rainfall monitoring.