Advanced terrain-adaptive tropical cyclone wind field modeling using deep learning for infrastructure resilience planning

Abstract

Tropical cyclones pose significant threats to the resilience of coastal communities, underscoring the need for reliable wind field models to support robust hazard analyses. Parametric wind models (PWMs), despite their computational efficiency, often fall short in capturing intricate wind-terrain interactions, leading to inaccurate resilience evaluations for spatially-distributed civil infrastructure systems situated in complex terrains. This study introduces an innovative approach that integrates the strengths of numerical wind models to handle intricate terrain features into PWMs through a deep learning-based Convolutional Neural Network for Terrain Modification (CNN-TM). The CNN-TM model, trained over 3 million km² of numerically simulated high-resolution wind fields, enhances terrain representation in PWMs by generating 450 m-resolution terrain-modified wind fields for both wind speed and direction. The accuracy and efficiency of this integration are validated across multiple scales: grid (∼0.2 km²), patch (∼506 km²), and region (∼34,000 km²). Applications during Typhoon Hagupit (2020) in Zhejiang Province, China, demonstrate its practical effectiveness across a 105,000 km² area. By leveraging deep learning to synergize numerical and parametric models, the CNN-TM model addresses limitations of traditional PWMs and provides a robust tool for resilience-oriented decision-making for infrastructure systems in coastal regions characterized by complex terrains.