Spatiotemporal Prediction of Radar Echoes Based on ConvLSTM and Multisource Data

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