Multi-Generator Tropical Cyclone Forecasting Based on Cross-Modal Fusion

Abstract

Accurate prediction of the trajectory and intensity of tropical cyclones (TCs) is crucial for disaster mitigation and prevention. Traditional TC forecasting methods suffer from high research costs, subjectivity, and low prediction accuracy, which pose challenges in meeting the demand for real-time weather forecasting. Recent advancements primarily rely on deep learning methods, which have shown some improvement in prediction performance. However, they have not fully capitalized on the correlation between features of different modalities, thus limiting the potential of multimodal feature fusion. To address this gap in current research, this paper introduces a novel multimodal feature fusion module integrated into a multi-branching model framework based on a Generative Adversarial Network with multiple generators. This framework aims to predict TC information and multiple potential motion trends at future time points. Additionally, a feature fusion loss function is devised to evaluate the effectiveness of the feature fusion module. To evaluate the effectiveness of the proposed model, extensive experiments are conducted on the CMA-BST dataset. Experimental findings reveal that the model's prediction performance surpasses that of existing deep learning methods and often outperforms the official predictor China Central Meteorological Observatory (CMO) across various metrics.

Keywords: Tropical cyclone prediction, cross-modal fusion, multiple generators, Attention mechanism