Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Global Tropical Cyclone Intensity Forecasting with Multi-modal Multi-scale Causal Autoregressive Model

Xinyu Wang, Kang Chen, Lei Liu, Tao Han, Bin Li, Lei Bai

TL;DR

This paper tackles the challenge of global tropical cyclone intensity forecasting by proposing MSCAR, a multimodal, multi-scale autoregressive model that injects causal temporal relationships via CC Attention and fuses satellite and ERA5 data through a Feature Pyramid Network. It introduces SETCD, the longest global TC dataset with 72 variables spanning 1980–2022, enabling robust learning across scales and environments. Experiments on SETCD demonstrate state-of-the-art short-term forecasting globally and regionally, with substantial error reductions and strong real-time robustness when ERA5 is replaced by analysis data. The work provides publicly available code and data, facilitating broader adoption and future TC research, and highlights the practical impact of integrating causal, multimodal information for disaster risk reduction.

Abstract

Accurate forecasting of Tropical cyclone (TC) intensity is crucial for formulating disaster risk reduction strategies. Current methods predominantly rely on limited spatiotemporal information from ERA5 data and neglect the causal relationships between these physical variables, failing to fully capture the spatial and temporal patterns required for intensity forecasting. To address this issue, we propose a Multi-modal multi-Scale Causal AutoRegressive model (MSCAR), which is the first model that combines causal relationships with large-scale multi-modal data for global TC intensity autoregressive forecasting. Furthermore, given the current absence of a TC dataset that offers a wide range of spatial variables, we present the Satellite and ERA5-based Tropical Cyclone Dataset (SETCD), which stands as the longest and most comprehensive global dataset related to TCs. Experiments on the dataset show that MSCAR outperforms the state-of-the-art methods, achieving maximum reductions in global and regional forecast errors of 9.52% and 6.74%, respectively. The code and dataset are publicly available at https://anonymous.4open.science/r/MSCAR.

Global Tropical Cyclone Intensity Forecasting with Multi-modal Multi-scale Causal Autoregressive Model

TL;DR

This paper tackles the challenge of global tropical cyclone intensity forecasting by proposing MSCAR, a multimodal, multi-scale autoregressive model that injects causal temporal relationships via CC Attention and fuses satellite and ERA5 data through a Feature Pyramid Network. It introduces SETCD, the longest global TC dataset with 72 variables spanning 1980–2022, enabling robust learning across scales and environments. Experiments on SETCD demonstrate state-of-the-art short-term forecasting globally and regionally, with substantial error reductions and strong real-time robustness when ERA5 is replaced by analysis data. The work provides publicly available code and data, facilitating broader adoption and future TC research, and highlights the practical impact of integrating causal, multimodal information for disaster risk reduction.

Abstract

Accurate forecasting of Tropical cyclone (TC) intensity is crucial for formulating disaster risk reduction strategies. Current methods predominantly rely on limited spatiotemporal information from ERA5 data and neglect the causal relationships between these physical variables, failing to fully capture the spatial and temporal patterns required for intensity forecasting. To address this issue, we propose a Multi-modal multi-Scale Causal AutoRegressive model (MSCAR), which is the first model that combines causal relationships with large-scale multi-modal data for global TC intensity autoregressive forecasting. Furthermore, given the current absence of a TC dataset that offers a wide range of spatial variables, we present the Satellite and ERA5-based Tropical Cyclone Dataset (SETCD), which stands as the longest and most comprehensive global dataset related to TCs. Experiments on the dataset show that MSCAR outperforms the state-of-the-art methods, achieving maximum reductions in global and regional forecast errors of 9.52% and 6.74%, respectively. The code and dataset are publicly available at https://anonymous.4open.science/r/MSCAR.
Paper Structure (34 sections, 4 equations, 5 figures, 11 tables)

This paper contains 34 sections, 4 equations, 5 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. TC Intensity Forecast
  4. TC dataset
  5. Methodology
  6. Problem Definition
  7. SETCD
  8. MSCAR
  9. Feature Pyramid Network (FPN)
  10. Causal Cross-Attention (CC Attention)
  11. Autoregressive Decoder (AR Decoder)
  12. Experiments
  13. Experimental Setup
  14. Datasets
  15. Implementation Details
  16. ...and 19 more sections

Figures (5)

  • Figure 1: (a) The overall architecture of the MSCAR. The diagram illustrates the process of forecasting the TC intensity at the $j$th time step in the future based on the previous four-time steps. ${K}$ represents the tokens for spatial information, and ${R}$ represents the tokens for temporal information. These tokens are pairwise matched using CC Attention, shown by the dashed box. This process generates ${H}$, which is then passed through the AR Decoder for ${j}$ iterations. Finally, the AR Decoder produces the TC intensity for the future ${j}$ step. (b) The diagram illustrates the calculation method of CC Attention (c) The AR Decoder takes the initial ${H}$ and iteratively updates it to obtain the TC intensity.
  • Figure 2: The global TC colored by their Saffir-Simpson Hurricane Wind Scale (SSHWS) intensity categories from 1980 to 2022.
  • Figure 3: The FPN structure framework diagram we used in the experiment. "Res" represents the ResNets module resnet, "k" denotes the kernel size, "s" indicates the stride, and "p" represents the padding.
  • Figure 4: The figure illustrates the computation process of the Cross Attention module, where "Attention" represents the multi-head attention calculation module, and "FFN" stands for the feed-forward neural network.
  • Figure 5: Visual comparison of TC intensity forecasts. "GT" represents the ground truth of TC intensity. "6h PRE" signifies the 6-hour forecast, while "12$\sim$24h PRE" denotes the forecast results for 12 hours, 18 hours, and 24 hours.