Improving tropical cyclone wind and pressure field reconstruction using GK-2A atmospheric motion vectors and machine learning

Enhancing the realism of numerical models is critical for accurately simulating high-impact weather events such as tropical cyclones (TCs), particularly for coastal hazard applications. Model performance is strongly influenced by the accuracy and spatial resolution of the input data. To address the challenges associated with the asymmetric and rapidly evolving structure of TCs, recent studies have increasingly incorporated advanced satellite observations and state-of-the-art machine-learning techniques. One of recent advance is the use of atmospheric motion vectors (AMVs) derived from  satellite imagery. In this study, a dedicated preprocessing framework incorporating quality control, outlier removal, and directional alignment, was developed to refine AMVs for TC wind-field reconstruction. Storm surge simulations driven by these AMV-based winds for TCs (i.e., Lingling, Haishen, and Hinnamnor) demonstrated improved accuracy relative to ERA5 reanalysis at several coastal stations, highlighting their effectiveness in data-sparse oceanic regions. In parallel, a random forest (RF) model was developed to estimate TC pressure fields from wind information. Unlike conventional symmetric parametric approaches, the RF model effectively represents spatial asymmetry, land–sea contrasts, and nonlinear wind–pressure relationships. The model achieves low error rates, particularly within the gale-force wind radius, and performs robustly when driven by real-time satellite wind observations. Overall, the integration of satellite-based observations with machine-learning techniques represents a significant advance toward more physically realistic and operationally valuable numerical modeling, helping bridge the gap between limited observations and complex storm dynamics to improve coastal hazard forecasting and emergency response.