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.
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.
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