How to Learn More? Exploring Kolmogorov-Arnold Networks for Hyperspectral Image Classification

Ali Jamali, Swalpa Kumar Roy, Danfeng Hong, Bing Lu, Pedram Ghamisi

Abstract

This paper investigates the application of Kolmogorov-Arnold Networks (KANs) to hyperspectral image (HSI) classification. By proposing a "Hybrid KAN" architecture that integrates 1D, 2D, and 3D KAN layers, the study demonstrates that KAN-based models can surpass traditional CNNs and Vision Transformers in accuracy. Furthermore, KANs produce cleaner classification maps with reduced noise, offering a more efficient and robust alternative for complex remote sensing data.

Introduction

Hyperspectral imaging (HSI) captures hundreds of narrow spectral bands for every pixel, providing a detailed "spectral fingerprint" of the Earth's surface. While powerful, this high-dimensional data is incredibly complex to process. Currently, researchers primarily rely on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for classification. However, these models are often "data-hungry" and computationally expensive. Modern Multi-Layer Perceptrons (MLPs) offer efficiency but struggle with the intricate, non-linear relationships inherent in HSI data. This paper explores Kolmogorov-Arnold Networks (KANs) as a promising alternative to bridge this gap, promising superior feature learning with less training data.

Methodology

The core innovation lies in the adaptation of Kolmogorov-Arnold Networks (KANs) for HSI. Unlike standard MLPs, where nodes have fixed activation functions (e.g., ReLU) and edges have linear weights, a KAN eliminates linear weights entirely. Instead, it employs learnable 1D activation functions, typically splines, directly on the edges. This allows the network to learn the optimal non-linear transformation for every connection.

Structure of a Kolmogorov-Arnold Network (KAN) featuring learnable activation functions on edges.

To handle the spatial and spectral dimensions of HSI data, the authors developed a Hybrid KAN architecture:

  1. Dimensionality Reduction: PCA is used to condense hundreds of spectral bands into a manageable number.
  2. Feature Extraction: Three consecutive 3D KAN layers capture joint spectral-spatial features.
  3. Spatial Refinement: A 2D KAN layer further processes the spatial context.
  4. Classification: A final 1D KAN layer maps the features to specific land-cover classes.

The spatial processing is facilitated by convolutional KAN layers, which apply the KAN principle within a sliding window framework to generate output feature maps.

Illustration of the convolutional KAN operation used for spatial feature extraction.

Experimental Setup

The study utilizes three brand-new HSI benchmark datasets captured via UAV over Qingdao, China:

  • QUH-Pingan: 176 bands, 10 classes.
  • QUH-Tangdaowan: 176 bands, 18 classes.
  • QUH-Qingyun: 270 bands, 6 classes.

These datasets are particularly challenging due to high spatial resolution (0.10–0.15 meters) and the complex distribution of natural and artificial land covers.

Results and Discussion

The Hybrid KAN consistently ranked at or near the top across all benchmarks:

  • Tangdaowan Dataset: Achieved a 97.12% Average Accuracy (AA), outperforming ViT (95.30%) and 3D CNN (75.61%).
  • Pingan Dataset: Achieved a 98.48% Overall Accuracy (OA), proving highly competitive against heavyweights like VGG-16 (99.06%) while maintaining better visual map quality.
  • Qingyun Dataset: Reached a top-tier 97.06% OA, significantly higher than the 1D CNN (91.63%) and ResNet-50 (92.80%).

Key insights from the evaluation include:

  • Efficiency and Convergence: KANs leverage the local precision of splines and the compositional power of MLPs, leading to faster convergence than standard CNN-based models like HybridSN.
  • Visual Quality: KAN-based models produced classification maps with significantly less "salt-and-pepper" noise compared to ViTs or CNNs, indicating a more robust spatial understanding.
  • Feature Separation: t-SNE visualizations confirmed that the Hybrid KAN clusters land-cover classes more distinctly in the feature space than its competitors.

Conclusion

The Hybrid KAN shows immense promise for HSI classification, offering high accuracy and robust spatial modeling. While the model exhibits higher parameter complexity than simpler KAN variants, its data efficiency is a significant advantage. Future research will focus on optimizing the architecture to reduce parameter counts and exploring performance on even smaller training sets to further establish KANs as a standard tool in remote sensing.

Summary updated

BibTeX
@article{jamali2024how,
  author = {Ali Jamali and Swalpa Kumar Roy and Danfeng Hong and Bing Lu and Pedram Ghamisi},
  title = {{How to Learn More? Exploring Kolmogorov-Arnold Networks for Hyperspectral Image Classification}},
  journal = {Remote Sensing},
  year = {2024},
  doi = {10.3390/rs16214015},
}

Loading related content...

Figure