Spatio-Temporal Gated Recurrent Unit (ST-GRU) Model for High-Fidelity Meteorological Video Forecasting

Abstract

 Complexity, multi-scale, and atmospheric process dynamism are some of the issues with meteorological video forecasting. This study presents a Spatio-Temporal Gated Recurrent Unit (ST-GRU) architecture for high-resolution sequential satellite video prediction in order to address the drawbacks of the aforementioned numerical techniques and previous deep learning models. The model uses channel-wise normalization, spatial-global gating, and multi-scale dilated convolutions in a recurrent neural network to explicitly represent both local and global spatiotemporal dependencies. With a weighted RMSE of 0.188 and a mean MSSSIM greater than 0.93 based on the experimental results in the proprietary satellite video dataset, ST-GRU outperforms all top baselines, including ConvGRU, PredRNN, TrajGRU, and PhyDNet, in terms of accuracy, consistency, and structure preservation. The necessity of all the architectural elements—multi-scale context fusion, global gating, and normalization—for reliable predictions during regime changes and uncommon convective phenomena has been further confirmed by ablation investigations. According to the aforementioned findings, the suggested model has a workable architecture and is appropriate for use in real-time forecasting since it can faithfully replicate the evolution of weather changes over time. The development of high-accuracy, data-driven weather prediction using extensive video data is supported technically by this study.