An LSTM based model for the prediction of INSAT-3DR satellite images for nowcasting

Abstract

The demand for accurate and early warning systems for severe weather occurrences is increasing due to the growth in both the frequency and intensity of extreme meteorological phenomena. Accurate advance prediction of such catastrophic events remains a significant challenge for weather forecasters, making disaster mitigation difficult. To enhance nowcasting accuracy, multiple data sources including meteorological (radar, satellite, and surface observations) and geographical information (elevation, exposure, vegetation, hydrological characteristics, and anthropogenic features) are increasingly integrated into forecasting frameworks. This study, presents a deep learning–based architecture employing Long Short-Term Memory (LSTM) neural networks for short-term satellite image prediction using the INSAT-3DR data, which plays a crucial role in weather nowcasting applications. Experimental evaluation is carried out using the Thermal Infrared-1 (TIR-1) channel. The LSTM-based framework processes satellite imagery by randomly segmenting images into pixel blocks of 5 × 5, 20 × 20, and 50 × 50 resolutions. Three model variants CNN–LSTM, Bidirectional LSTM, and Vanilla LSTM are implemented to predict future satellite images at lead times of 30 minutes, 1 hour, 1.5 hours, and 2 hours. The best-performing model for the TIR-1 band achieves an average normalized mean absolute error of approximately 1.40%, demonstrating high prediction accuracy and confirming the effectiveness of the proposed approach for INSAT-3DR satellite image nowcasting.