Short-Term Multi-Hazard Prediction Using a Multi-Source Data Fusion Approach

The increasing frequency and intensity of natural disasters necessitate advanced prediction techniques to mitigate potential damage. This study presents a comprehensive multi-hazard early warning framework by integrating the multi-source data fusion technique. A multi-source data extraction method was introduced by extracting pressure level and average precipitation data based on the hazard event from the Cooperative Open Online Landslide Repository (COOLR) dataset across multiple temporal intervals (12 h to 1 h prior to events). Feature engineering was performed using Choquet fuzzy integral-based importance scoring, which enables the model to account for interactions and uncertainty across multiple features. Three individual Long Short-Term Memory (LSTM) models were trained for hazard location, average precipitation, and hazard category (i.e., to detect the potential of natural disasters). These models were trained on varying temporal scales from 12 to 1 h prior to the event. These individual models achieved the performance of Mean Absolute Error (MAE) 2.2 and 3.2, respectively, for the hazard location and average precipitation models, and an F1-score of 0.825 for the hazard category model. The results also indicate that the LSTM model outperformed traditional Machine Learning (ML) models, and the use of the fuzzy integral enhanced the prediction capability by 8.12%, 2.6%, and 6.37%, respectively, for all three individual models. Furthermore, a rule-based algorithm was developed to synthesize the outputs from the individual models into a $3\times 3$ grid of multi-hazard warnings. These findings underscore the effectiveness of the proposed framework in advancing multi-hazard forecasting and situational awareness, offering valuable support for timely and data-driven emergency response planning.