Spatiotemporal Prediction of Radar Echoes Based on ConvLSTM and Multisource Data

Accurate and timely precipitation forecasts can help people organizations make informed decisions, plan for potential weather-related disruptions, protect lives property. Instead of using physics-based numerical forecasts, which be computationally prohibitive, there has been a growing interest in deep learning techniques prediction recent years due to the success these approaches various other fields. These generally use historical composite reflectivity (CR) at surface level predict future time steps. However, relevant factors related motion vertical structure storm have not considered. To address this issue, research proposes multisource ConvLSTM (MS-ConvLSTM) model improve accuracy forecasting by incorporating multiple data sources into process. The was trained on dataset radar echo features, includes only (CR), but also top (ET), vertically integrated liquid (VIL) water, radar-retrieved wind field different elevations. Experiment results showed that proposed outperformed traditional methods terms evaluation metrics, such as mean absolute error (MAE), squared (MSE), probability detection (POD), false alarm rate (FAR), critical index (CSI).