[Paper Review] KAN-FIF: Spline-Parameterized Lightweight Physics-based Tropical Cyclone Estimation on Meteorological Satellite

Jiakang Shen, Qinghui Chen+5 more|arXiv (Cornell University)|Feb 12, 2026

Tropical and Extratropical Cyclones Research0 citations

TL;DR

KAN-FIF introduces a lightweight, physics-guided multimodal framework using Kolmogorov–Arnol’d networks to estimate tropical cyclone MSW and RMW on edge devices, achieving major parameter and latency reductions with maintained accuracy.

ABSTRACT

Tropical cyclones (TC) are among the most destructive natural disasters, causing catastrophic damage to coastal regions through extreme winds, heavy rainfall, and storm surges. Timely monitoring of tropical cyclones is crucial for reducing loss of life and property, yet it is hindered by the computational inefficiency and high parameter counts of existing methods on resource-constrained edge devices. Current physics-guided models suffer from linear feature interactions that fail to capture high-order polynomial relationships between TC attributes, leading to inflated model sizes and hardware incompatibility. To overcome these challenges, this study introduces the Kolmogorov-Arnold Network-based Feature Interaction Framework (KAN-FIF), a lightweight multimodal architecture that integrates MLP and CNN layers with spline-parameterized KAN layers. For Maximum Sustained Wind (MSW) prediction, experiments demonstrate that the KAN-FIF framework achieves a $94.8\%$ reduction in parameters (0.99MB vs 19MB) and $68.7\%$ faster inference per sample (2.3ms vs 7.35ms) compared to baseline model Phy-CoCo, while maintaining superior accuracy with $32.5\%$ lower MAE. The offline deployment experiment of the FY-4 series meteorological satellite processor on the Qingyun-1000 development board achieved a 14.41ms per-sample inference latency with the KAN-FIF framework, demonstrating promising feasibility for operational TC monitoring and extending deployability to edge-device AI applications. The code is released at https://github.com/Jinglin-Zhang/KAN-FIF.

Motivation & Objective

• Motivate timely tropical cyclone monitoring on resource-constrained edge devices.
• Reduce model size and inference latency while maintaining or improving accuracy.
• Model cross-modal dependencies between infrared imagery and temporal TC features with physics-based constraints.
• Demonstrate edge deployment feasibility on FY-4 satellite processor hardware.

Proposed method

• Replace traditional CNN/MLP components with Kolmogorov–Arnold Network (KAN) layers to dramatically reduce parameters.
• Fuse temporal sequence data and infrared imagery through a hybrid encoder with physics-based, differentiable polynomial constraints.
• Incorporate a physics-guided constraint module to model high-order polynomial relationships between MSW and RMW.
• Use KAN-based shared feature extraction (KAN-LSTM and KAN-CNN) for temporal and spatial features respectively.
• Implement a KAN-based attention mechanism to encode spatial distances and content features with center-aware attention.
• Deploy a deployment-ready variant by removing LSTM and enforcing static computation graphs for edge hardware compatibility.

Experimental results

Research questions

• RQ1 Can KAN layers achieve similar or better TC attribute prediction accuracy with far fewer parameters?
• RQ2 How effective are physics-based, high-order polynomial constraints in coupling MSW and RMW predictions?
• RQ3 Is edge-device deployment feasible with real-time or near-real-time latency on FY-4 satellite hardware?
• RQ4 Does multimodal fusion of sequential features and infrared imagery improve MSW/RMW estimation over single-modality models?

Key findings

• KAN-FIF achieves 94.8% parameter reduction (0.99 MB vs 19 MB) and 68.7% faster per-sample inference (2.3 ms vs 7.35 ms) compared with Phy-CoCo.
• KAN-FIF reduces MSW MAE by 32.5% (3.21 kt vs 4.76 kt) and MSW RMSE by 31.9% (4.31 kt vs 6.33 kt) relative to Phy-CoCo.
• RMW MAE is slightly improved (8.83 nmi vs 8.89 nmi) with MSW/RMW RMSE improvements (11.66 vs 12.24 RMSE for RMW).
• Offline deployment on the FY-4 processor with Qingyun-1000 board achieves 14.41 ms per-sample inference latency, showing edge feasibility.
• Removal of LSTM for deployment caused marginal accuracy loss, indicating a viable static-graph edge deployment path.