[Paper Review] Enhancing AI-Based Tropical Cyclone Track and Intensity Forecasting via Systematic Bias Correction

Peisong Niu, Haifan Zhang+7 more|arXiv (Cornell University)|Mar 20, 2026

Tropical and Extratropical Cyclones Research0 citations

TL;DR

BaguanCyclone combines probabilistic center refinement and region-aware intensity forecasting to surpass AI baselines and NWP in track and intensity predictions across six basins, achieving 16% tracking error reduction and 34% intensity precision improvement, with real-time CMA deployment in 2025.

ABSTRACT

Tropical cyclones (TCs) pose severe threats to life, infrastructure, and economies in tropical and subtropical regions, underscoring the critical need for accurate and timely forecasts of both track and intensity. Recent advances in AI-based weather forecasting have shown promise in improving TC track forecasts. However, these systems are typically trained on coarse-resolution reanalysis data (e.g., ERA5 at 0.25 degree), which constrains predicted TC positions to a fixed grid and introduces significant discretization errors. Moreover, intensity forecasting remains limited especially for strong TCs by the smoothing effect of coarse meteorological fields and the use of regression losses that bias predictions toward conditional means. To address these limitations, we propose BaguanCyclone, a novel, unified framework that integrates two key innovations: (1) a probabilistic center refinement module that models the continuous spatial distribution of TC centers, enabling finer track precision; and (2) a region-aware intensity forecasting module that leverages high-resolution internal representations within dynamically defined sub-grid zones around the TC core to better capture localized extremes. Evaluated on the global IBTrACS dataset across six major TC basins, our system consistently outperforms both operational numerical weather prediction (NWP) models and most AI-based baselines, delivering a substantial enhancement in forecast accuracy. Remarkably, BaguanCyclone excels in navigating meteorological complexities, consistently delivering accurate forecasts for re-intensification, sweeping arcs, twin cyclones, and meandering events. Our code is available at https://github.com/DAMO-DI-ML/Baguan-cyclone.

Motivation & Objective

• Address the limitations of coarse-resolution reanalysis data (0.25° ERA5) for TC tracking and intensity forecasting.
• Develop a unified framework that refines center location with sub-grid precision and improves localized intensity estimates.
• Investigate data-centric causes of intensity forecasting gaps in AI models and design region-aware solutions.
• Demonstrate real-world operational deployment and validate performance across major TC basins.

Proposed method

• Introduce BaguanCyclone with two integrated modules: probabilistic center refinement and region-aware intensity forecasting.
• Model cyclone center as a continuous probability density and recover sub-grid coordinates via an expectation operation.
• Use kinematic initialization, probabilistic density mapping with a truncated Gaussian kernel, and neural correction to refine track.
• Partition the atmosphere into regions around the TC core and apply a region-aware intensity predictor (Swin-transformer-based) to capture localized maxima.
• Couple the tracking and intensity modules so the refined center queries the regional intensity map for final forecasts.
• Demonstrate joint deployment in inference to produce coherent track and intensity predictions.

Experimental results

Research questions

• RQ1 Can probabilistic, sub-grid center refinement overcome grid-locking from 0.25° input data to improve TC track accuracy?
• RQ2 Does a region-aware, high-resolution intensity forecasting approach better capture extreme TC intensities than coarse-grid or standard AI losses?
• RQ3 Does coupling track refinement with region-aware intensity yield coherent improvements over independent modules?
• RQ4 How does BaguanCyclone perform across multiple basins and complex TC scenarios (re-intensification, multi-TC interactions, meandering) compared to NWP and AI baselines?

Key findings

• BaguanCyclone achieves a 16% average improvement in tracking error over AI baselines across six basins.
• BaguanCyclone delivers a 34% average reduction in forecast error for intensity relative to AI baselines.
• In real-world deployment (WP basin, CMA), average track error was 85.61 km and average intensity error was 6.20 m/s.
• BaguanCyclone outperformed ECMWF HRES in WP for all lead times with an average 12.23% gain in track/intensity performance.
• NI basin saw up to a 50% reduction in intensity forecast error compared with the second-best model.
• During Typhoon Co-May (July 2025), the system supported evacuations of 97,000 people, demonstrating operational value.