Storm Nowcasting Analysis using LSTM Model

ABSTRACT Thunderstorms pose a significant risk to the island region of Nosy Be, Madagascar, due to their sudden onset and localized impact. Traditional forecasting methods often fall short in providing timely and precise alerts, especially in data-sparse environments. This study presents a deep learning-based approach for short-term thunderstorm nowcasting using Long Short-Term Memory (LSTM) neural networks. Leveraging real-time satellite-derived meteorological features—including latitude, longitude, storm intensity, size, and distance—we developed two specialized LSTM models to predict the probability of storm occurrence within 1-hour and 3-hour windows. The models were trained on labeled datasets and evaluated using metrics such as accuracy, precision, recall, and ROC-AUC, achieving test accuracies of 90.25% (1-hour) and 89.32% (3-hour). Our findings indicate that LSTM networks are well-suited for capturing both temporal and spatial structure and outperform classic machine learning model such as Random Forest and XGBoost in this context. A web interface was engineered, for live user interaction, for prediction by real input handling. The model produces probabilistic predictions, which enable more refined, risk-informed decision making in early warning systems. This work contributes to the development of scalable, location-specific storm prediction frameworks and has significant implications for disaster preparedness in vulnerable regions. Keywords: Storm Nowcasting, LSTM (Long Short-Term Memory), Thunderstorm Prediction, Short-term Weather Forecasting, Meteorological Data, Satellite-derived Features, Time-Series Forecasting, Deep Learning in Meteorology, Convective Storms, Real-time Forecasting, Nosy Be, Madagascar, Machine Learning for Weather Prediction, Spatiotemporal Analysis, Probabilistic Forecasting, Web-based Decision Support System