HyperKAN: Kolmogorov-Arnold Networks Make Hyperspectral Image Classifiers Smarter.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Multilayer Perceptrons (MLPs) are standard in neural network classification.
• Kolmogorov-Arnold Networks (KANs) offer a novel alternative with potential for enhanced accuracy.
• Pixel-wise classification of hyperspectral images is crucial for remote sensing applications.

Purpose of the Study:

• To investigate the efficacy of KAN-based networks for hyperspectral image classification.
• To compare KAN performance against traditional MLPs.
• To evaluate KAN replacements for linear, convolutional, and attention layers in existing architectures.

Main Methods:

• Comparative analysis of baseline MLP and KAN networks with varied hidden layer neuron counts.
• Modification of six state-of-the-art neural networks (1DCNN, 2DCNN, two 3DCNNs, NM3DCNN, SSFTT) by integrating KAN layers.
• Experimental validation using seven diverse, publicly available hyperspectral datasets.

Main Results:

• KAN-based networks consistently outperformed baseline MLP networks across all tested architectures.
• Significant improvements in pixel-wise classification accuracy were observed for all modified networks.
• The KAN-based transformer architecture (SSFTT) yielded the highest classification accuracy.

Conclusions:

• KANs represent a superior alternative to MLPs for hyperspectral image classification tasks.
• Replacing conventional layers with KAN counterparts enhances the performance of various neural network architectures.
• KAN-based transformers show exceptional promise for achieving state-of-the-art results in hyperspectral image analysis.