A data-driven approach for regional-scale fine-resolution disaster impact prediction under tropical cyclones

Tropical cyclones (TCs) pose a significant threat to coastal regions worldwide, demanding accurate and timely predictions of potential disaster impacts. Existing regional-scale impact prediction models, however, are largely limited by the sparsity of modeling data and incapability of fine-resolution predictions in a computationally efficient manner, thus hindering real-time identification of potential disaster hotspots. To address these limitations, we present a data-driven image-to-image TC impact prediction model based on a deep convolutional neural network (CNN) for Zhejiang Province, China, an area of approximately 105,000 km2 consisting of 90 counties. The proposed model utilizes twelve carefully selected predictors, including hazard, environmental and vulnerability factors, which are processed into province-scale 1 km-grid image-format data. An end-to-end encoder-decoder architecture is subsequently designed to extract impact-relevant spatial features from the multi-channel input images, then to construct a spatial impact map of identical size (i.e., $${\sim}105,000$$ km2) and resolution (i.e.,1 km-grid). This gridded impact map is then aggregated spatially to derive county-level impact predictions, which serve as the final layer of the CNN model and are used to evaluate the model’s loss function in terms of mean squared error. This design is informed by the fact that the training data on TC impact, collected from historical events, were recorded at county level. Validation and error analysis demonstrate the model’s promising spatial accuracy and time efficacy. Furthermore, an illustration of the model’s application with Typhoon Lekima in 2019 underscores its potential for integrating meteorological forecasts to achieve real-time impact predictions and inform emergency response actions.