Deep Learning Reconstruction of Tropical Cyclogenesis in the Western North Pacific from Climate Reanalysis Dataset

TL;DR

This paper addresses reconstructing tropical cyclone genesis climatology in the Western North Pacific from climate reanalysis data using a deep learning framework. It adopts a ResNet-18 CNN and two labeling strategies (Past Domain and Dynamic Domain) to train on MERRA-2 and IBTrACS data (1980–2022). The model captures the seasonal cycle and spatial patterns of TCG at $0.5^{\circ}$ resolution, and sensitivity analyses show that a subset of environmental channels suffices for reconstruction. The study demonstrates potential for climate downscaling and downstream TCG prediction from large-scale environments, with public code available for reproducibility.

Abstract

This study presents a deep learning (DL) architecture based on residual convolutional neural networks (ResNet) to reconstruct the climatology of tropical cyclogenesis (TCG) in the Western North Pacific (WNP) basin from climate reanalysis datasets. Using different TCG data labeling strategies and data enrichment windows for the NASA Modern-Era Retrospective analysis for Research and Applications Version 2 (MERRA2) dataset during the 1980-2020 period, we demonstrate that ResNet can reasonably reproduce the overall TCG climatology in the WNP, capturing both its seasonality and spatial distribution. Our sensitivity analyses and optimizations show that this TCG reconstruction depends on both the type of TCG climatology that one wishes to reconstruct and the strategies used to label TCG data. Of interest, analyses of different input features reveal that DL-based reconstruction of TCG climatology needs only a subset of channels rather than all available data, which is consistent with previous modeling and observational studies of TCG. These results not only enhance our understanding of the TCG process but also provide a promising pathway for predicting or downscaling TCG climatology based on large-scale environments from global model forecasts or climate output. Overall, our study demonstrates that DL can offer an effective approach for studying TC climatology beyond the traditional physical-based simulations and vortex-tracking algorithms used in current climate model analyses.

Figures (11)

• Figure 1: Illustration of a TCG data labeling strategy based on the dynamical domain approach, for which a positive TCG label at one location is surrounded by 8 negative TCG labels for the sampling strategy.
• Figure 2: a) The pipeline of the DL design for reconstructing TCG climatology from the MERRA-2 dataset, and (b) the core DL model based on the ResNet-18 architecture used for TCG reconstruction in this study.
• Figure 3: Precision ($P$, blue columns), Recall ($R$, red columns), and F1 score (black columns) for the TCG prediction as obtained from the ResNet-18 model with the test set as a function of the temporal data enrichment window at an interval of 6 hours, using a) the PD sampling strategy, and b) the DD sampling strategy.
• Figure 4: (a) The $P$ score for the TCG prediction of the ResNet-18 model with the test set as a function of the temporal data enrichment window using a range of the RUS ratio and class weight (solid colors) for the PD sampling strategy, (b)-(c) similar to (a) but for the $R$ and F1 scores, respectively. Dashed black line denotes the values used for our best-tuned model. Weight balanced assigns fixed importance to each class based on frequency whilst weight dynamics adaptively adjusts sample or class importance.
• Figure 5: Similar to Figure \ref{['fig:RUS_sensitivity_PD']} but for the DD sampling strategy.