Abstract Hyperspectral images (HSI) record detailed spectral data over numerous narrow wavelength bands, producing data with high dimensionality often containing redundant, irrelevant, or noisy bands. This high dimensionality poses challenges for classification tasks, including increased computational burden, over-fitting, and degraded model performance. Intrinsic dimensionality (ID) estimation is therefore crucial to recognize the true count of informative dimensions inherent in the data, guiding the selection of suitable dimensionality reduction (DR) and manifold learning techniques that preserve meaningful spectral features while discarding noise and redundancy. In this work, a comprehensive set of ID estimation methods is applied to the HSI (PRISMA and LongKou datasets) to determine its optimal ID. Based on these estimates, some of the significant linear and nonlinear DR techniques are employed to obtain low-dimensional, discriminative feature spaces for improved classification efficiency and interpretability. To address the challenges of inter-class spectral similarity and intra-class variability commonly found in HSI, a Siamese Neural Network (SNN) framework is adopted. Unlike conventional classifiers, the SNN learns a similarity-based embedding space using contrastive loss, enabling effective pair-wise class separation even in reduced dimensions. Experimental results demonstrate that combining accurate ID estimation with appropriate DR techniques significantly enhances classification performance. The SNN further improves robustness to spectral overlap and noise, offering a reliable and computationally efficient pipeline for hyperspectral image analysis. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Comprehensive evaluation of linear and nonlinear dimensionality reduction techniques for HSI classification. </ItemContent> <ItemContent> Multiple intrinsic dimensionality estimators applied to determine optimal subspace sizes. </ItemContent> <ItemContent> Linear and nonlinear methods showed comparable classification performance across datasets. </ItemContent> <ItemContent> Siamese Neural Network effectively handled varied DR outputs using contrastive loss. </ItemContent> <ItemContent> Combination of dimensionality reduction and Siamese Neural Network was more important than the choice of DR method alone. </ItemContent> </UnorderedList>

From dimensions to decisions: A Siamese Neural Network for comprehensive analysis of intrinsic dimensionality estimation and dimensionality reduction in hyperspectral imaging

期刊论文
发表日期：

2026年1月21日

查看全文

Ashish Kumar,
Rahul Dev Garg

摘要

Abstract

Hyperspectral images (HSI) record detailed spectral data over numerous narrow wavelength bands, producing data with high dimensionality often containing redundant, irrelevant, or noisy bands. This high dimensionality poses challenges for classification tasks, including increased computational burden, over-fitting, and degraded model performance. Intrinsic dimensionality (ID) estimation is therefore crucial to recognize the true count of informative dimensions inherent in the data, guiding the selection of suitable dimensionality reduction (DR) and manifold learning techniques that preserve meaningful spectral features while discarding noise and redundancy. In this work, a comprehensive set of ID estimation methods is applied to the HSI (PRISMA and LongKou datasets) to determine its optimal ID. Based on these estimates, some of the significant linear and nonlinear DR techniques are employed to obtain low-dimensional, discriminative feature spaces for improved classification efficiency and interpretability. To address the challenges of inter-class spectral similarity and intra-class variability commonly found in HSI, a Siamese Neural Network (SNN) framework is adopted. Unlike conventional classifiers, the SNN learns a similarity-based embedding space using contrastive loss, enabling effective pair-wise class separation even in reduced dimensions. Experimental results demonstrate that combining accurate ID estimation with appropriate DR techniques significantly enhances classification performance. The SNN further improves robustness to spectral overlap and noise, offering a reliable and computationally efficient pipeline for hyperspectral image analysis.

Research highlights

Comprehensive evaluation of linear and nonlinear dimensionality reduction techniques for HSI classification.

Multiple intrinsic dimensionality estimators applied to determine optimal subspace sizes.

Linear and nonlinear methods showed comparable classification performance across datasets.

Siamese Neural Network effectively handled varied DR outputs using contrastive loss.

Combination of dimensionality reduction and Siamese Neural Network was more important than the choice of DR method alone.