A Multimodal Spatio-Temporal Attention Mechanism for Regional Meteorological Disaster Risk Assessment

Meteorological disaster risk assessment is a crucial component for disaster prevention and mitigation. Traditional methods for meteorological disaster risk assessment struggle to effectively integrate multi-source heterogeneous data and capture complex spatio-temporal dependencies. This paper proposes a Multimodal Spatio-Temporal Attention (MSTA) mechanism for regional meteorological disaster risk assessment. This model can simultaneously process meteorological time series data, remote sensing images, and text data by using intra-modal spatio-temporal attention, cross-modal attention, and adaptive modal weight mechanisms to achieve effective fusion of multimodal information. Experimental results show that compared with existing methods, the proposed MSTA model achieves significant improvements in meteorological disaster risk assessment tasks in Sichuan Province with an accuracy of 84.5%, a warning lead time of 3.3 days, and F1 score of 0.83 for extreme weather event prediction. Additionally, this paper analyzes the performance of different LSTM variants on this task, demonstrating the necessity and effectiveness of attention mechanisms. The proposed method provides new insights and technical support for regional meteorological disaster risk assessment with significant theoretical value and application prospects.