CVPR Findings2026
iTCTSL: Interpretable Tropical Cyclone Track and Intensity Forecasting via Task Sensitive Learning
Pan Mu, Yuchao Zhu, Shiqi Zhang, Hanting Yan, Jinglin Zhang, Cong Bai
Abstract
Tropical Cyclones (TCs) cause significant damage, especially in coastal regions. Despite the critical role of topography in TC evolution, its effects remain underrepresented in existing deep learning-based TC forecasting approaches. Moreover, current methods lack a systematic investigation into how environmental factors relate to TC track and intensity forecasting, which limits model interpretability. To address the above issues, we propose an interpretable TC track and intensity forecasting model via Task Sensitive Learning (iTCTSL). First, iTCTSL integrates a dynamic topography perception mechanism that improves forecasting accuracy in complex topographies by modeling the relationship between topography and TC directional evolution through factor disentanglement. Moreover, a task-sensitive environmental routing module is designed to enhance model interpretability, characterized by formulating the quantification of meteorological variable contributions as a differentiable optimal transport problem and using a soft selection strategy to capture task-customized environmental features. Finally, a dynamic multi-task coordination mechanism is proposed to adaptively coordinate sub-tasks during the optimization process. iTCTSL demonstrates remarkable performance in all TC forecasting scenarios and greater effectiveness on coastal TCs, where topographic influences are more significant, and further shows strong global generalization capability. We also provide interpretations of the relationships between meteorological factors and forecasting tasks captured by deep learning.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on3
- MGTCF: Multi-Generator Tropical Cyclone Forecasting with Heterogeneous Meteorological DataCheng Huang, Cong Bai, Sixian Chan, Jinglin Zhang et al.AAAI 2023 · 19 citations
- TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone ForecastingShiqi Zhang, Pan Mu, Cheng Huang, Jinglin Zhang et al.AAAI 2025 · 6 citations
- Joint Scheduling of Causal Prompts and Tasks for Multi-Task LearningChaoyang Li, Jianyang Qin, Jinhao Cui, Zeyu Liu et al.CVPR 2025
Related papers
- TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change AwarenessCheng Huang, Pan Mu, Cong Bai, Peter AG WatsonICML 2025
- CausalX: A Unified and Causally-Interpretable Plug-and-Play Model for Multi-modal Spatio-Temporal ForecastingShiqi Zhang, Pan Mu, HantingYan, Yuchao Zhu et al.ICML 2026
- IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task EstimationHanting Yan, Pan Mu, Shiqi Zhang, Yuchao Zhu et al.NeurIPS 2025
- AutoCTS+: Joint Neural Architecture and Hyperparameter Search for Correlated Time Series ForecastingXinle Wu, Dalin Zhang, Miao Zhang, Chenjuan Guo et al.SIGMOD 2023 · 36 citations
- TROPHY: A Topologically Robust Physics-Informed Tracking Framework for Tropical CyclonesLin Yan, Hanqi Guo, Thomas Peterka, Bei Wang et al.IEEE VIS 2023 · 8 citations
