IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task Estimation

TL;DR

IDOL tackles distribution shifts in tropical cyclone multi-task estimation by learning identity tokens that encode physical invariances. It combines a Holland-model–based task dependency flow, a dark knowledge graph–driven correlation bridge, and an identity-aware estimator to jointly predict wind speed, pressure, and core sizes. The approach integrates prior physics into feature distributions, yielding improved accuracy, stability, and distribution alignment on PDTC and DigitalTC datasets, with strong ablations and MI/KDE evidence supporting the physical identities. IDOL demonstrates practical robustness for real-time TC monitoring and forecasting, with code available for reproducibility.

Abstract

Tropical Cyclone (TC) estimation aims to accurately estimate various TC attributes in real time. However, distribution shifts arising from the complex and dynamic nature of TC environmental fields, such as varying geographical conditions and seasonal changes, present significant challenges to reliable estimation. Most existing methods rely on multi-modal fusion for feature extraction but overlook the intrinsic distribution of feature representations, leading to poor generalization under out-of-distribution (OOD) scenarios. To address this, we propose an effective Identity Distribution-Oriented Physical Invariant Learning framework (IDOL), which imposes identity-oriented constraints to regulate the feature space under the guidance of prior physical knowledge, thereby dealing distribution variability with physical invariance. Specifically, the proposed IDOL employs the wind field model and dark correlation knowledge of TC to model task-shared and task-specific identity tokens. These tokens capture task dependencies and intrinsic physical invariances of TC, enabling robust estimation of TC wind speed, pressure, inner-core, and outer-core size under distribution shifts. Extensive experiments conducted on multiple datasets and tasks demonstrate the outperformance of the proposed IDOL, verifying that imposing identity-oriented constraints based on prior physical knowledge can effectively mitigates diverse distribution shifts in TC estimation.Code is available at https://github.com/Zjut-MultimediaPlus/IDOL.

Figures (15)

• Figure 1: Diverse Distribution Shifts in TC Multi-Task Estimation. The $\mathtt{v}$, $\mathtt{p}$, $\mathtt{ri}$ and $\mathtt{ro}$ represent the wind speed, pressure, inner-core and outer-core size of TC, respectively.
• Figure 2: Solution to various distribution shifts: Regulating the feature space to task identities by prior physics.
• Figure 3: The framework of the proposed IDOL. $\mathtt{v}$, $\mathtt{p}$, $\mathtt{ri}$, and $\mathtt{ro}$ denote the TC wind speed, pressure, inner-core size, and outer-core size, respectively. DevEnc, CorEnc, and GMM refer to the proposed Development Encoder, Correlation Encoder, and Gaussian Mixture Model, respectively.
• Figure 4: (a) Distribution visualization of test set estimation results based on KDE. (b) Variance comparison between physical identity (PID) tokens and the features extracted by DeepTCNet.
• Figure 5: Estimation performance of pressure and outer-core size on test samples under distribution shifts: (a) Estimation results under covariate shift and (b) Mean absolute errors under label and concept shifts across different TCs and output attributes. Each vertical red box indicates a pair of samples exhibiting distribution shift.