Hybrid Integration of CNN and Kolmogorov–Arnold Networks for Multispectral Image Super-Resolution

Abstract

This letter proposes KAN-based multispectral image super-resolution method (KMSR), a novel deep learning framework for multispectral image (MSI) super-resolution (SR) that integrates convolutional neural networks (CNNs) and Kolmogorov–Arnold networks (KANs). The model employs a KAN-block to enhance nonlinear spectral–spatial representation, combined with CNNs for local feature extraction. Supported by channel weighting, residual connections, and global skip links, the approach effectively improves reconstruction accuracy and spectral consistency. Experiments on Sentinel-2 imagery demonstrate that KMSR outperforms state-of-the-art methods in both quantitative metrics and visual quality, providing a robust and efficient solution for MSI SR.