From reanalysis to climatology: deep learning reconstruction of tropical cyclogenesis in the western North Pacific

Abstract

Tropical cyclogenesis (TCG) climatology is the key to understanding regional weather extremes and long-term risk, yet their large-scale environmental drivers remain difficult to characterize from observations or traditional physical-based modeling. In this study, we present a deep learning (DL) framework based on an 18-layer residual convolutional neural network (TCG-Net) to reconstruct TCG climatology in the western North Pacific (WNP) basin from climate reanalysis data. The framework addresses two tasks (1) the Past Domain (PD) task that predicts when TCG occurs in the WNP within the next 48 h, and (2) the Dynamic Domain (DD) task that predicts the spatial distribution of TCG at a given date and time. For each task, different labeling strategies are employed to generate negative samples that can help maximize the distinction between TCG and non-TCG conditions. To enhance the model's capability in handling the rarity of TCG data, temporal feature enrichment is further used to incorporate environmental information from the preceding 6 h time steps, which helps improve the representation of each training task. In addition, random under-sampling is applied with class weighting to address the severe imbalance caused by large numbers of negative TCG samples under these labeling strategies. Using NASA's Modern-Era Retrospective analysis for Research and Applications Version 2 (MERRA-2) with a training period from 1980–2016 and a test set from 2017–2022, we show that TCG-Net achieves an overall F1-score of 0.39 for the PD task and 0.33 for the DD task. In the PD task, feature selection experiments reveal that only a subset of environmental variables is required for robust performance, consistent with prior physical studies. In contrast, for the DD task, full-feature models perform better, likely due to their ability to exploit unknown or latent feature interactions. Both tasks reproduce key characteristics of the observed seasonality and spatial TCG distribution when evaluated against the best-track dataset. These results demonstrate that DL-based reconstructions, when coupled with task-specific labeling, temporal enrichment, and imbalance-aware training, can complement physics-based models and vortex-tracking algorithms and provide an efficient pathway for downscaling or projecting TCG climatology from coarse-resolution climate model outputs.