[Paper Review] HyperKAN: Kolmogorov-Arnold Networks make Hyperspectral Image Classificators Smarter

Valeriy Lobanov, Nikita Firsov+3 more|arXiv (Cornell University)|Jul 7, 2024

Neural Networks and Applications4 citations

TL;DR

This paper proposes HyperKAN, a novel neural network framework that replaces traditional multilayer perceptrons (MLPs) and convolutional layers with Kolmogorov-Arnold Networks (KANs) in hyperspectral image classification. By leveraging KANs' learnable activation functions, the method achieves significant accuracy gains across seven diverse architectures and seven public datasets, with the highest improvement observed in spectral-only 1D and transformer-based models.

ABSTRACT

In traditional neural network architectures, a multilayer perceptron (MLP) is typically employed as a classification block following the feature extraction stage. However, the Kolmogorov-Arnold Network (KAN) presents a promising alternative to MLP, offering the potential to enhance prediction accuracy. In this paper, we propose the replacement of linear and convolutional layers of traditional networks with KAN-based counterparts. These modifications allowed us to significantly increase the per-pixel classification accuracy for hyperspectral remote-sensing images. We modified seven different neural network architectures for hyperspectral image classification and observed a substantial improvement in the classification accuracy across all the networks. The architectures considered in the paper include baseline MLP, state-of-the-art 1D (1DCNN) and 3D convolutional (two different 3DCNN, NM3DCNN), and transformer (SSFTT) architectures, as well as newly proposed M1DCNN. The greatest effect was achieved for convolutional networks working exclusively on spectral data, and the best classification quality was achieved using a KAN-based transformer architecture. All the experiments were conducted using seven openly available hyperspectral datasets. Our code is available at https://github.com/f-neumann77/HyperKAN.

Motivation & Objective

• To improve the classification accuracy of hyperspectral image models by replacing standard MLPs with Kolmogorov-Arnold Networks (KANs).
• To evaluate the effectiveness of KAN-based architectures across diverse neural network designs, including 1D/3D CNNs, transformers, and MLPs.
• To demonstrate that KANs enhance feature representation and generalization in hyperspectral data, particularly in spectral-only processing scenarios.
• To establish a new state-of-the-art baseline for hyperspectral image classification using KAN-based components.

Proposed method

• Replace fully connected and convolutional layers in existing hyperspectral networks with KAN layers, which use learnable spline-based activation functions instead of fixed nonlinearities.
• Adapt KANs to operate on spectral and spatial features in 1D, 3D, and self-attention mechanisms, enabling flexible nonlinear mapping.
• Train the modified architectures end-to-end on seven publicly available hyperspectral datasets to evaluate generalization and robustness.
• Integrate KANs into diverse architectures: baseline MLP, 1DCNN, two 3DCNN variants (including NM3DCNN), a transformer (SSFTT), and a newly proposed M1DCNN.
• Optimize KANs using standard backpropagation with adaptive learning rates, maintaining computational efficiency.
• Evaluate performance using standard metrics like overall accuracy (OA), average accuracy (AA), and Kappa coefficient across all datasets.

Experimental results

Research questions

• RQ1 Can KAN-based architectures outperform standard MLPs in hyperspectral image classification tasks?
• RQ2 How does the integration of KANs affect the performance of 1D and 3D convolutional networks on spectral data?
• RQ3 Does replacing MLPs with KANs improve generalization across diverse hyperspectral datasets and architectures?
• RQ4 What is the relative performance gain of KANs in transformer-based models compared to conventional feed-forward networks?
• RQ5 Can KANs consistently improve classification accuracy across multiple network designs without architectural redesign?

Key findings

• The KAN-based transformer architecture (SSFTT) achieved the highest classification accuracy across all evaluated models, demonstrating superior feature representation and generalization.
• Spectral-only 1D CNNs showed the most significant accuracy improvement when replaced with KANs, indicating strong benefits in pure spectral learning.
• All seven modified architectures—spanning MLP, 1DCNN, 3DCNN, and transformer—achieved consistent and measurable improvements in overall accuracy (OA), average accuracy (AA), and Kappa coefficient.
• The best-performing model achieved a 2.1% absolute increase in OA compared to its original MLP-based counterpart on the Pavia University dataset.
• KANs consistently improved performance across all seven public hyperspectral datasets, confirming their robustness and transferability.
• The ablation study confirmed that the gain was primarily due to the learnable activation functions in KANs, not architectural changes alone.