Advanced rainfall nowcasting using 3D convolutional LSTM networks on satellite data

Abstract

This paper introduces an innovative method for rainfall nowcasting using a deep learning model that combines 3D Convolutional Neural Networks (3D-CNN) with Long Short-Term Memory (LSTM) model. The primary objective is to improve the accuracy and timeliness of short-term rainfall predictions. The 3D-CNN component is responsible for extracting spatial features from complex weather data, while the LSTM component captures temporal dependencies across time steps. This hybrid architecture, referred to as the 3D-Conv-LSTM model, has demonstrated high effectiveness for nowcasting applications. The model processes weather data stored in Network Common Data Form (NetCDF) files and integrates satellite imagery to enhance forecast precision. This dual-data approach enables the model to learn intricate spatiotemporal patterns and relationships often missed by traditional techniques. Through extensive experimentation and validation, the proposed model exhibits superior performance in predicting precipitation events compared to conventional methods. The model achieved a Mean Squared Error (MSE) of 0.0003, Peak Signal-to-Noise Ratio (PSNR) of 42.11, Root Mean Square Error (RMSE) of 0.019, and a Structural Similarity Index Measure (SSIM) of 0.99, indicating excellent prediction quality. Furthermore, the computation time for training and inference was recorded 18 min, demonstrating the model’s efficiency. These results confirm a significant improvement in forecast accuracy, which is critical for disaster preparedness and resource management in weather-sensitive regions.