Work

A Hybrid Deep Learning Framework for Tropical Cyclone Intensity Estimation With Spatiotemporal Features and Physical‐Persistence Adjustment

Year: 2026

Type: article

Abstract: This study proposes SOPHIENet (TC Satellite‐Observed Physical‐Persistence and Spatiotemporal Hybrid Intensity Estimation Network), a hybrid framework for Tropical Cyclone (TC) intensity estimation. SOPHIENet integrates a convolutional neural network, the Convolutional Block Attention Module, and a physical‐persistence adjustment mechanism. By extracting spatiotemporal features from multi‐temporal satellite infrared (IR) imagery, SOPHIENet captures TC cloud‐structure evolution. The embedded attention module refines channel‐wise and spatial representations, improving the feature‐learning capability of the deep learning (DL) component. A key innovation is the physical‐persistence adjustment mechanism inspired by the Dvorak technique, which converts historical TC intensity‐evolution characteristics over the Western North Pacific (WNP) into multi‐timescale constraints on intensity change. These constraints suppress physically unreasonable short‐term fluctuations in the initial DL estimates, improving the temporal consistency, physical plausibility, and reliability. In an independent operational evaluation using FY‐4B IR observations for 43 TCs (20,144 samples) over the WNP during 2023–2024, SOPHIENet achieved a mean absolute error (MAE) of 2.8 m/s. A temporally collocated comparison of 1,841 samples from 26 TCs in 2024 yielded an MAE of 2.7 m/s for SOPHIENet against China Meteorological Administration best‐track 2‐min maximum sustained wind (MSW), compared with 3.4, 3.6, and 4.0 m/s for D‐PRINT, AiDT, and Advanced Dvorak Technique v9.0, respectively, against Joint Typhoon Warning Center best‐track 1‐min MSW. The comparison is not strictly homogeneous because of differences in wind‐averaging conventions, best‐track references and TC center inputs. More broadly, SOPHIENet also showed competitive performance relative to existing DL approaches and stable performance across TC intensity categories, with slight underestimation for super typhoons and greater uncertainty during rapid intensity evolution.

Source: Journal of Geophysical Research Machine Learning and Computation

DOI: 10.1029/2026jh001526

Authors: Yue Wang, Xiaoqin Lu, Xinyan Lyu, Lin Dong

Institutions: Shanghai Typhoon Institute, China Meteorological Administration

Language: English

Topic: Tropical and Extratropical Cyclones Research

Subfield: Atmospheric Science

Field: Earth and Planetary Sciences

Domain: Physical Sciences

SDG: Climate action

Open Access status: gold

Estimated APC paid: $2,500