J. Eur. Opt. Society-Rapid Publ. 2025, 21, 31 Ó The Author( s), published by EDP Sciences, 2025 https:// doi. org / 10.1051 / jeos / 2025029 Available online at: https:// jeos. edpsciences. org
Journal of the European Optical Society-Rapid Publications
RESEARCH ARTICLE
Dimensionality reduction method based on spatial-spectral preservation and minimum noise fraction for hyperspectral images
Bing Zhou 1, Lei Deng 1,*
, Jiaju Ying 1, Qianghui Wang 2, and Yue Cheng 1
1 Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang, Hebei 050000, China 2 State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System, Luoyang,
Henan 471003, China
Received 24 April 2025 / Accepted 9 June 2025
Abstract. Hyperspectral images contain rich spatial distribution and spectral information of land features, but they also introduce high information redundancy and computational complexity. This paper proposes dimensionality reduction methods that integrate spatial-spectral preservation and minimum noise fraction( MNF) to better analyze and utilize the spatial and spectral information in hyperspectral images. While performing the minimum noise separation transformation, the proposed method aims to preserve the spatial structure of the image as much as possible, maximizing both the signal-to-noise ratio and the spatial structure similarity of the image. The component selection strategy involves grouping components and calculating the average change in the relative position of all pixels in the feature space. The component group that most closely matches the spectral relative position before transformation is selected as the final dimensionality reduction result. Experimental results demonstrate that the proposed method is highly sensitive to noise estimation and requires a relatively accurate noise covariance matrix. The method effectively preserves spatial information, with negligible impact on the accuracy of object detection methods, and outperforms other comparative approaches. It ensures the effectiveness of downstream object detection tasks while significantly reducing computational time. The code of the proposed method is available at https:// github. com / aosilu / spatial-spectral-preservation-MNF.
Keywords: Hyperspectral image, Dimensionality reduction, Minimum noise fraction( MNF), Spatial-spectral preservation.
1 Introduction
Hyperspectral images are characterized by their composition of tens to hundreds of contiguous narrow spectral bands, forming a three-dimensional data structure that seamlessly integrates spatial and spectral information [ 1 ]. From a spectral perspective, each pixel within a hyperspectral image encapsulates a continuous spectral curve, which is intrinsically linked to the unique physicochemical properties of the corresponding ground object. From a spatial perspective, each individual band contributes to the construction of a detailed spatial distribution map, representing the geographical arrangement of ground objects [ 2, 3 ]. The synergistic combination of rich spatial and spectral information endows hyperspectral imaging with unparalleled capabilities, making it an indispensable tool in a wide array of applications. These include, but are not limited to, mineralogical exploration, oceanic and environmental
* Corresponding author. denglei @ aeu. edu. cn monitoring, food quality assessment, and battlefield surveillance [ 4, 5 ].
Hyperspectral images contain numerous contiguous spectral bands, encompassing rich spectral information. However, such high-dimensional data can lead to the“ curse of dimensionality”, making data processing and analysis challenging [ 6 ]. Existing methods for processing hyperspectral images often exhibit high algorithmic complexity, resulting in significant computational load. Data redundancy and high correlation among bands further degrade algorithm performance [ 7 ]. Many researchers have conducted extensive and in-depth studies on dimensionality reduction for hyperspectral images to eliminate redundant information and enable more efficient data processing and analysis. Unsupervised methods have emerged as effective techniques for dimensionality reduction in unknown scenarios, as they do not rely on prior information or expert knowledge.
The commonly used algorithms for dimensionality reduction of hyperspectral images include Principal Component Analysis( PCA), Minimum Noise Fraction( MNF),
This is an Open Access article distributed under the terms of the Creative Commons Attribution License( https:// creativecommons. org / licenses / by / 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.