Reanalysis data fusion-based tropical cyclone formation prediction network

Abstract

Abstract This paper proposes a tropical cyclone formation prediction network based on pyramid attention feature extraction and multi-scale feature fusion, aiming to enhance the prediction of whether tropical cloud cluster (TCC) precursors intensify to tropical storm (TS) strength by integrating multi-source reanalysis data. First, reanalysis data from ERA5 and NCEP/NCAR are represented as images at different scales and labeled according to tropical cyclone (TC) formation events derived from the TCC and IBTrACS datasets. The labels include information on whether TC formation occurred and its location. Then, a feature extraction module based on a Pyramid Attention Mechanism (PAM) is designed to extract features related to TC formation. Next, the features at different scales are input into a PAM-based feature fusion module that dynamically generates weights for different features to perform weighted fusion and unify the feature scales. Finally, a lightweight Convolutional Neural Network (CNN) is designed as the prediction module to predict TC formation occurrence and location. Experimental results over five independent data splits show that at a lead time of 24 hours, the proposed method achieves a Probability of Detection (POD) of $$0.865 \pm 0.012$$ , a False Alarm Ratio (FAR) of $$0.326 \pm 0.016$$ , and a location prediction Root Mean Square Error (RMSE) of $$0.795 \pm 0.033$$ grid units ( $$\approx$$ 437 km), demonstrating competitive performance in balancing POD and FAR. Ablation studies confirm the contribution of each proposed module to overall performance.