Deep spatio-temporal learning for multi-hazard events: A ConvGRU multi-label classification approach

The forecasting of multi-hazards is a vital, though underinvestigated, area of disaster risk management. The traditional studies have mainly focused on single-hazard forecasting, thus leaving its utility in real-world and realistic scenarios. This study, in turn, presents a spatio-temporal multi-label classification model, a framework designed expressly to capture the complex interrelationships between a range of hazards. The methodological framework used disaster occurrence data from the Open Federal Emergency Management Agency (OpenFEMA) database and converted the raw records of disasters into a multi-label dataset. Pressure-level reanalysis data is extracted from Climate Data Store (CDS) based on the multi-hazard event. Spatial data is extracted in 25
59 grid format in different temporal dependencies (12 h, 8 h, 6 h) at the 850 hPa pressure level. The model architecture combines convolutional neural networks (CNNs) with spatial attention mechanisms and gated recurrent units (GRUs) that model the temporal sequences. This combination enables multi-hazard predictions by utilizing the spatial and temporal data. Experimental analysis reveals that the proposed model outperformed the baseline variants, i.e., 2D CNN, Convolutional Long Short-Term Memory (ConvLSTM), and Convolutional Gated Recurrent Unit (ConvGRU) without attention. The proposed model achieved per-class accuracy up to 0.8868, the subset accuracy is 0.55, and the Hamming loss up to 0.127, which are 3.88%, 13.59% and 21.12% performance improvements over the baseline models respectively. In addition, the use of various lead times and the fusion of multiple lead times (12 h+8 h+6 h) significantly improves the predictive capability. The proposed framework has high potential for disaster preparedness and early warning systems in the real world. It proposes a flexible and efficient method of dealing with the growing complexity of multi-hazard environments.