[Paper Review] Global Tropical Cyclone Intensity Forecasting with Multi-modal Multi-scale Causal Autoregressive Model

Xinyu Wang, Kang Chen+4 more|arXiv (Cornell University)|Feb 16, 2024

Tropical and Extratropical Cyclones Research6 citations

TL;DR

MSCAR is a global tropical cyclone intensity forecasting model that fuses satellite and ERA5 data through a multi-scale causal autoregressive framework, introducing SETCD as a large multimodal TC dataset and achieving state-of-the-art short-term forecasting in global and regional tests.

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.

Motivation & Objective

• Motivate improved TC intensity forecasting by leveraging multi-modal, multi-scale data and causal temporal relationships.
• Introduce SETCD, the longest global TC dataset with 43 years of satellite and ERA5 data across 72 variables.
• Develop MSCAR to integrate spatial multi-scale features with temporal causality for autoregressive TC intensity forecasts.
• Demonstrate MSCAR's state-of-the-art performance globally and across basins, with real-time robustness and ablation insights.

Proposed method

• Fuse satellite imagery and ERA5 data via a Feature Pyramid Network to capture multi-scale spatial features.
• Represent spatial information as tokens and apply a Causal Cross-Attention mechanism to incorporate temporal causality between history and spatial context.
• Use an Autoregressive Decoder that iteratively forecasts future TC intensities in latent space, enabling arbitrary-step forecasts.
• Train MSCAR with MAE loss and regularization (L1 and EMA) on SETCD, using past 4 steps to predict the next 4 steps (N=4, M=4).
• Provide SETCD with 4,668 TCs and 282,505 instances across 1980–2022, derived from GridSat-B1 and ERA5 with IBTrACS centers.

Experimental results

Research questions

• RQ1 Can a multi-modal, multi-scale, causally informed autoregressive model improve global TC intensity forecasts beyond existing methods?
• RQ2 What is the impact of including ERA5 and satellite data, as well as historical TC intensity sequences, on forecast accuracy?
• RQ3 How does MSCAR perform across global and regional basins, and under near real-time data scenarios?
• RQ4 What is the contribution of each architectural component (FPN, CC Attention, AR Decoder) to forecasting performance?
• RQ5 Is SETCD a comprehensive and beneficial resource for future global TC research?

Key findings

• MSCAR achieves state-of-the-art short-term TC intensity forecasts versus baselines (global and regional tests).
• In global short-term forecasts (2019–2020), MSCAR shows improvements over conventional models across multiple horizons (up to 24h).
• Ablation shows ERA5 inputs yield substantial gains (9.39%–25.09%), historical intensity sequences are crucial for short horizons, and CC Attention outperforms self-attention in capturing causal spatiotemporal relations.
• Real-time-style tests (using analysis data as ERA5 substitute) show MSCAR maintains comparable performance to non-real-time results, indicating robustness for near real-time use.
• Component ablations confirm the importance of FPN over U-Net for multi-scale fusion and the necessity of causal cross-attention over self-attention for physical plausibility and performance.