KAN-FIF: Spline-Parameterized Lightweight Physics-based Tropical Cyclone Estimation on Meteorological Satellite

Jiakang Shen, Qinghui Chen, Runtong Wang, Chenrui Xu, Jinglin Zhang, Cong Bai, Feng ZhangarXiv 2026 · arXiv

TL;DR

KAN-FIF leverages Kolmogorov-Arnold Networks (KANs) to create an ultra-lightweight, physics-guided model for estimating tropical cyclone intensity and size directly on meteorological satellites. By replacing traditional multi-layer perceptrons (MLPs) with spline-parameterized KAN layers, it slashes parameter counts by 94.8% and speeds up inference by nearly 70% while significantly improving prediction accuracy.

What Problem Does This Paper Address?

Monitoring tropical cyclones (TCs) traditionally relies on downlinking massive amounts of satellite imagery to Earth for centralized processing, introducing critical delays in disaster response. While edge computing on satellites could enable real-time, in-orbit tracking, existing deep learning models are too computationally heavy and memory-intensive for resource-constrained hardware.

Furthermore, current physics-guided models rely on linear feature interactions that fail to capture the complex, high-order polynomial relationships between TC attributes (such as intensity and size). This limitation results in bloated model sizes and poor hardware compatibility, making them impractical for deployment on edge devices.

Methods

The authors introduce the Kolmogorov-Arnold Network-based Feature Interaction Framework (KAN-FIF). Unlike traditional Multi-Layer Perceptrons (MLPs) that use fixed node activation functions, KANs utilize learnable, spline-parameterized activation functions on the network edges. This architectural shift allows the model to capture intricate nonlinearities with exceptional parameter efficiency.

Key components of the framework include:

  • KAN-based Shared Feature Extraction: Parallel pathways using KAN-CNN for multi-scale spatial pattern extraction from infrared imagery and KAN-projections for temporal sequence data.
  • KAN-Attention: Replaces standard linear+tanh attention mechanisms with KAN layers to better encode spatial distances and cloud patterns.
  • Physics-Guided Constraint Modeling: Implements bidirectional residual connections governed by KAN layers to fit high-dimensional polynomial relationships between a cyclone's Maximum Sustained Wind (MSW) and its Radius of Maximum Wind (RMW).
  • Edge-Device Optimization: To ensure compatibility with the static computation graphs required by satellite Neural Processing Units (NPUs), the authors refactored the architecture to replace NPU-incompatible LSTM layers with static KAN projections.
Figure 1: The comparison of frameworks between the MLP- based and KAN-based models. (a) Conventional MLP-based model (Phycoco) (b) our KAN-based lightweight model
Figure 3: The architecture of the KAN-FIF learning framework: (a) The Task-specific Feature Extraction Module uses KAN layer and center-aware attention to extract the task features of MSW and RMW respectively (b) The Physical Constraint Module is designed to conduct constraints among task-specific features ; (c) The KAN Fusion Decoder fuse the features from(a)(b)(d) and obtain the final output ; (d) The Shared Feature Extraction Module take the multi-channel image and the temporal sequence data as input and obtain the shared feature between TC tasks; (e) the architecture of KAN layers; (f) the architecture of Multiscale Conv

Datasets

The model was evaluated using the Tropical Cyclone Multi-Modal Dataset (TCMM) spanning 2015–2022, which includes:

  • Satellite Imagery: Infrared brightness temperatures from four channels of the Himawari-8 satellite.
  • TC Track Records: Spatiotemporal evolution features (latitude, longitude, central pressure, storm category, and time since naming) sourced from the IBTrACS repository, structured in 3-hour interval sequences.

Key Insights and Ideas

  • The Power of Splines over MLPs: Physical laws governing meteorological phenomena are inherently non-linear and polynomial. Because KANs optimize inter-layer function parameters using spline bases, they can represent these complex physical equations far more efficiently than standard neural networks.
  • Mitigating Negative Transfer: In multitask learning, optimizing for both intensity (MSW) and size (RMW) often causes performance degradation due to conflicting gradients. KAN-FIF’s explicit, differentiable physics-constraint module allows the tasks to mutually inform each other, turning a potential bottleneck into an accuracy gain.
  • Hardware-Driven Refactoring: The paper demonstrates that removing recurrent layers (like LSTMs) to satisfy edge-device constraints does not have to ruin performance; flattening temporal sequences into KAN layers preserves the necessary temporal context without violating NPU operator constraints.

Key Results

  • Massive Parameter Reduction: KAN-FIF reduced the model parameter footprint from 19MB (baseline Phy-CoCo) to just 0.99MB (a 94.8% reduction). An ultra-compressed variant (KAN-FIF-s) shrunk this further to 0.22MB with minimal loss in accuracy.
  • Faster Inference: On standard hardware, per-sample inference time dropped from 7.35ms to 2.3ms (68.7% faster).
  • Superior Accuracy: KAN-FIF achieved a 32.5% lower Mean Absolute Error (MAE) for MSW prediction (3.21 knots vs. 4.76 knots) compared to the state-of-the-art baseline.
  • Hardware Deployment Validation: The deployment version was tested offline on the Qingyun-1000 development board (simulating an FY-4 series satellite processor equipped with an Ascend 310 NPU). It successfully achieved a per-sample inference latency of 14.41ms, proving its readiness for operational, in-orbit edge computing.
Figure 4: Visual comparison between KAN-FIF and Phy-CoCo models under different tropical cyclone structures

Limitations and Future Work

While offline hardware validation on the NPU was highly successful, the authors note that actual on-orbit deployment introduces additional engineering hurdles. These include data pipeline integration, radiation hardening, strict power management constraints (<10W budget), and command-and-control system synchronization.

Future work will focus on expanding the framework to handle larger multimodal spatiotemporal datasets and optimizing dynamic KAN configurations tailored for diverse meteorological satellite deployment scenarios.

Summary updated

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.
BibTeX
@misc{shen2026kan,
  author = {Jiakang Shen and Qinghui Chen and Runtong Wang and Chenrui Xu and Jinglin Zhang and Cong Bai and Feng Zhang},
  title = {{KAN-FIF: Spline-Parameterized Lightweight Physics-based Tropical Cyclone Estimation on Meteorological Satellite}},
  year = {2026},
  eprint = {2602.12117},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.12117},
}
Figure