A Recognition Method for Weak Tropical Cyclones Based on a Two-Stage Artificial Intelligence Model Architecture with Cloud Cluster Structural Decomposition Features

Artificial intelligence (AI)-based recognition technology for tropical cyclones (TCs) using satellite imagery has been extensively utilized in meteorological monitoring. While this technology exhibits high recognition accuracy for mature TCs, it nevertheless encounters considerable challenges in identifying cloud clusters associated with weak TCs and those in the genesis stage. The core underlying reason for this limitation lies in the underdeveloped nature of cloud cluster systems for weak and nascent TCs, which are characterized by irregular morphology, incomplete vortex structures, and loose cloud cover—features that hinder accurate identification. This study presents a TC cloud cluster recognition method that integrates a two-stage AI model architecture and cloud cluster structural decomposition features. The first-stage AI model implements the screening and localization of candidate cloud clusters, facilitating the identification of a set of targets with TC cloud cluster-like characteristics from complex backgrounds; this target set comprises both TC and non-TC cloud clusters, while determining the central position of each candidate cloud cluster. Using the localized central positions of the cloud clusters, we perform Fourier decomposition to further extract the symmetric and asymmetric features of the cloud clusters, thereby constructing decomposition features that depict the cloud cluster structure. Then, the second-stage AI model is adopted to classify the extracted structural decomposition features of cloud clusters and identify which cloud clusters correspond to TC structures. Through the combination of the two-stage AI model and the utilization of unique structural features of TC cloud clusters, the proposed method can efficiently identify TC cloud clusters while maintaining a low false detection rate. In particular, it markedly enhances the accuracy and reliability of recognizing weak TCs and those in the genesis stage, thus offering technical support for the early monitoring and warning of TCs.