Reconstructing Pre-Satellite Tropical Cyclogenesis Climatology Using Deep Learning

TL;DR

This study presents a deep learning approach to reconstruct historical TC activity in the western North Pacific (WNP) basin, with a main focus on the pre-satellite era, and demonstrates that DL can effectively capture the main characteristics and changes in TCG climatology during the post-satellite era.

Abstract

A reliable tropical cyclone (TC) climatology is the key to assessing historical and future changes in TC activities. While global TC records have been systematically maintained since the early 1940s, substantial uncertainties remain for the pre-satellite era during which TC observations relied mostly on scattered aircraft reconnaissance and sporadic ship reports. This study presents a deep learning (DL) approach to reconstruct historical TC activity in the western North Pacific (WNP) basin, with a main focus on the pre-satellite era. Using data feature enrichment tailored for tropical cyclogenesis (TCG), we demonstrate that DL can effectively capture the main characteristics and changes in TCG climatology during the post-satellite era. With additional cross-validations, the reconstruction of TCG climatology is then extended to a pre-satellite period (1940-1960) during which TC base-track datasets are most uncertain. Our DL reconstruction reveals a significant missing of TCG in the current best-track data between September and November during the pre-satellite era. Such a TCG undercount in the best track data occurs mainly around 10-15$^\circ$N in the central WNP, while coastal regions show better consistency with DL reconstruction. These findings not only highlight the potential of DL for improving historical assessments of TC activity, but also advance our understanding of TCG processes by identifying key environmental conditions conducive to TC formation. The DL approach presented herein can be applied to other ocean basins, climate proxies, or reanalysis datasets for future TC climate studies.

Figures (5)

• Figure 1: (a)-(h) Spatial distribution of five-year average of TCG probability predictions (blue shading) from the ResNet-18 DL model during the test period from 2017--2022, based on different data enrichment windows ranging from 6 to 48 hours prior to the first moment of a Tropical Depression record in the best track data; and (i) observed TCG density (gray shading) for the same 2017--2022 period that is derived from the best-track data. Here, the observed TCG density is defined as the total number of actual TCG events reported in each grid box during the test period, divided by the total number of TCG events recorded in the best-track dataset for the same test period. Note that different shading scales are used for the best-track data and DL predictions to better display the contrast in areas of maximum TCG probability in the DL model.
• Figure 2: Monthly distribution of TCG frequency detected in the WNP basin from the test data (2017--2022), using the ResNet-18 model with 8 different time windows from 6 to 48 hours (blue columns). Black columns denote the actual TCG frequency obtained from the best-track database IBTrACS during the same period. The error bars represent the statistics of DL predictions using different data enrichment windows.
• Figure 3: (a)-(b) Similar to Fig. \ref{['fig:dist_1722']}, but using the 1980--1985 period as the test data for the best-track TCG density (gray shading, left panel) and the DL-reconstructed TCG density using a 18-hour data enrichment window (blue shading, right panel), (c)-(d) the difference between the 1980-1985 period and the 2017--2022 period, and (e) similar to Fig. \ref{['fig:dist_monthly']} but for the TCG seasonality during the 1980--1985 period (left panel) as obtained from the best-track data (black columns) and DL reconstruction (blue columns), and (f) the relative difference between the 1980-1985 and the 2017-2022 periods (right panel).
• Figure 4: Similar to Fig. \ref{['fig:dist_8085']}a,b,e but for the TCG density (shaded) and frequency (columns) distribution during the pre-satellite 1940--1960 period, using the training data from MERRA-2 and ERA5 during 1980--2020. (d)-(f). Similar to (a)-(c) but for the pre-satellite 1960-1970 period.
• Figure 5: a) A data enrichment design to generate TCG labels needed for training the ResNet-18 model, where the asterisk denotes the first timestamp in the best-track data that indicates the existence of a Tropical Depression (TD) stage; and b) the DL model design based on the ResNet-18 architecture that is used for TCG reconstruction in this study (adopted from Le_etal2025).