Published Online:October 2026
Product Name:The IUP Journal of Computer Sciences
Product Type:Article
Product Code:IJCS011026
DOI:10.71329/IUPJCS/2026.20.4.7-20
Author Name:Ramesh Babu Jonnalagadda, Sagar K G, Mrinmoy Matilal, Gagan Nagesh, Thejas K J and Zunaira Khan
Availability:YES
Subject/Domain:Engineering
Download Format:PDF
Pages:7-20
Hyperspectral images (HSIs) capture rich spectral and spatial information across hundreds of contiguous wavelength bands, making them highly valuable in remote sensing, precision agriculture, medical imaging, and environmental monitoring. However, HSIs are inherently susceptible to various noise sources, including Gaussian noise, shot noise, stripe noise, and impulse noise introduced during image acquisition. Traditional denoising methods based on mathematical models struggle with the complexity and dimensionality of HSI data. Over the past decade, deep learning— particularly convolutional neural networks (CNNs)—has emerged as a powerful data-driven solution for HSI denoising. This paper critically reviews recent CNN-based approaches to HSI denoising, examining architectures ranging from simple 2D CNNs and 3D CNNs to advanced hybrid models combining CNNs with attention mechanisms. The key research gaps, such as limited interoperability between noise types, high computational cost, and the scarcity of real-world clean-noisy data pairs, are identified. The paper also proposes a lightweight CNN architecture for practical implementation, along with evaluation metrics, dataset choices, and future directions. This review is intended to serve as a structured and accessible reference for undergraduate students and early researchers entering the field of deep learning-based image processing.
Hyperspectral imaging (HSI) is a technique that captures visual information in both visible light range, and also across hundreds of narrow and contiguous spectral bands spanning ultraviolet, visible, near-infrared, and shortwave infrared wavelengths.