23 February 2022 in Geophysical Research Letters
Forecasting the trajectory and intensity of tropical cyclones (TCs) is important in disaster mitigation as TC usually causes huge damages. However, it remains a substantial challenge due to the limited understanding of TC complexity. Still, TCs have been observed and recorded for several decades, and so can be predicted if viewed as a spatial‐temporal prediction problem with a huge amount of existing data. We propose a novel TC trajectory and intensity short‐term prediction method: Multi‐Modal Spatial‐temporal Networks (MMSTN). It not only predicts the TC's central pressure, winds, and the location of its center, but also forecasts the TC's varied possible tendencies. Experiments were conducted on the China Meteorological Administration Tropical Cyclone Best Track Dataset. Experimental results show that the proposed MMSTN outperformed state‐of‐the‐art methods as well as the official prediction method of the China Central Meteorological Observatory, in intensity prediction and 6 hr trajectory prediction.
8 May 2021 in Multimedia Tools and Applications
1 July 2025 in Nature Communications
Accurate tropical cyclone (TC) forecasting is critical for disaster prevention. While deep learning shows promise in weather prediction, existing approaches demonstrate limited accuracy in TC track and intensity forecasting, hindered by the lack of open multimodal datasets and insufficient integration of meteorological knowledge. Here we propose TropiCycloneNet containing TCN
D
- a open multimodal TC dataset spanning six major ocean basins with 70 years of multi-source data, and TCN
M
- an AI-meteorology integrated prediction model including multiple modules such as Generator Chooser Network and Environment-Time Net. Comprehensive evaluations demonstrate that TCN
M
outperforms both existing deep learning methods and official meteorological forecasts across multiple metrics. This advancement stems from synergistic optimization of our meteorologically-informed architecture and the dataset’s comprehensive spatiotemporal coverage. The released resources and method can attract more researchers to the field, thereby accelerating data-driven tropical cyclone prediction research.
8 May 2026 in Journal of Geophysical Research: Atmospheres
Tropical cyclones (TCs) are destructive weather systems that can trigger flooding, landslides, and other severe hazards. Accurate estimation of key TC attributes is critical for disaster risk management. In particular, the maximum sustained wind speed (MSW) and radius of maximum wind (RMW) are fundamental parameters for operational forecasting and wind‐field modeling. However, existing approaches struggle to capture and fuse the diverse physical information embedded in multi‐modal TC data. We propose Multi‐Modal Hybrid Guided Network (MHG‐Net), a multi‐task deep learning framework that integrates satellite imagery with auxiliary physical factors via soft parameter sharing and correlation graph embedding. This design enables effective inter‐task collaboration while explicitly modeling underlying physical correlations essential for accurate TC estimation. Using Himawari‐8/9 satellite observations and best‐track data from IBTrACS, MHG‐Net achieves low mean absolute errors of 6.49 knots for MSW and 7.08 nmi for RMW. Beyond performance gains, the learned correlation graphs and attribution analyses provide insights into how different physical factors influence TC attributes evolution.
25 June 2026 in The Computer Journal
Achieving accurate air quality predictions depends on a variety of factors, including meteorological conditions, the spatial distribution of pollution sources, and the geographic environment of the target area. This complexity poses challenges in capturing the intricate spatiotemporal dependencies inherent in air quality data. To address this issue, this work proposes a spatiotemporal graph network called GA$^{2}$Transformer, as a novel Transformer-based encoder–decoder architecture, to capture fine-grained spatiotemporal dependencies of air quality data. The proposed GA$^{2}$Transformer integrates a multilevel discrete wavelet decomposition module for extracting trend and periodicity clues at multiple levels from air quality data, and a graph-assisted attention module for modeling graph-structured spatiotemporal dependencies of air quality data. Extensive experiments on the real-world collected Beijing dataset show the validity of the proposed GA$^{2}$Transformer in air quality prediction tasks, outperforming other used methods.