Skip to content
Open access

Unit-Simplex-Inspired Graph Attention Network for Robust Endmember Determination in Hyperspectral Imagery

Sep 2026 · Remote Sensing · Vol 18, pp. 3270 · 0 citations · 56 references

TL;DR

A novel unit-simplex-inspired graph attention network (GAT) to identify endmembers based on the pure pixel assumption has competitive performance and improved robustness under heterogeneous scenarios, despite its limitations in dealing with highly mixed data.

Abstract

Endmember extraction (EE) in an unmixing chain aims at determining source component spectra constituting pixels from hyperspectral imagery. However, the influence of spectral variability, nonlinear manifolds, and outliers on datasets can seriously degrade EE’s reliability even if pure pixels exist. Improving EE’s accuracy and applicability across heterogeneous scenarios remains a challenge. In this paper, we propose a novel unit-simplex-inspired graph attention network (GAT) to identify endmembers based on the pure pixel assumption. First, the geometric relationship between pixels in a constructed graph is incorporated into a network architecture which integrates spectral feature learning with GAT-driven adaptive information aggregation and inference among neighboring pixel nodes. Second, a unit-simplex maximization rule is exploited to guide the network’s training without relying on data reconstruction, producing pixel purity representations. The normalized purity of pixels with respect to different land covers depends on their geometric positions and relations on the data manifold. Finally, candidate endmember bundles consisting of pixels with high purity are recognized, and the purest pixels in each endmember bundle are identified as the endmembers because their purity feature vectors construct the largest unit simplex. The pixel purity-based characterization of endmembers’ spatial distributions enables geometrically meaningful discrimination between endmember and non-endmember nodes in a graph, also improving EE’s robustness to interfering factors. Extensive experiments conducted on different types of simulated datasets and three real hyperspectral images demonstrate that the proposed method has competitive performance and improved robustness under heterogeneous scenarios, despite its limitations in dealing with highly mixed data.

Read PDF

Similar papers

Open access Oct 2026

An Abundance-Guided Mamba Network for Hyperspectral Unmixing

Results indicate that AGMNet is particularly effective for abundance recovery while retaining physically interpretable spectral reconstruction, although further improvement of endmember spectral-shape consistency remains necessary.

Zhong-Yi Tang, Xiao-Qiang Wang, Bing-Chun Li · 0 citations
Open access Oct 2026

UGSA-Net: Uncertainty-Gated Attention and Self-Supervised Alignment for Calibration-Aware Hyperspectral Image Classification

UGSA-Net is proposed, an ambiguity-guided spatial–spectral transformer that infers a per-pixel ambiguity field without direct label supervision that exceeds 98.9% OA on every scene over three runs, and is comparable to competing architectures retrained under the same protocol.

Houari Youcef Moudjib, Hai-Bin Duan, Zain Anwar Ali et al. · 0 citations
Open access 2026

Spectral Gradient Field-Guided Bayesian Adaptive Graph Convolution for Hyperspectral Image Classification

Graph-based deep learning methods play a significant role in hyperspectral image classification (HSIC). Nonetheless, the paucity of prior information to construct graph model structures continues to severely challenge the classification of images characterized by significant spatial heterogeneity. To achieve effective...

Heng-Yi Zheng, Hong-Jun Su, F. Bovolo et al. · 0 citations
2026

Sparse Tensor Attention for Hyperspectral Unmixing

Hyperspectral unmixing (HU) is a fundamental task for resolving the mixed pixel problem by decomposing a hyperspectral image (HSI) into constituent endmembers and their corresponding abundances. Although deep-learning-based HU methods have achieved promising performance, most rely on manually designed network architect...

Jing-Ran Wang, Feng-Chao Xiong, Zhiyuan Wen et al. · 0 citations
Oct 2026

Enhancing Mamba with discriminative graph convolution for hyperspectral image classification

Accurate hyperspectral image (HSI) classification requires long-range dependency modeling without losing irregular spatial context. Existing Mamba-based approaches may linearize two-dimensional neighborhoods, whereas graph-based approaches can depend strongly on a single graph definition; simply combining the two archi...

Unknown authors · 0 citations
Conference Open access Oct 2026

Transformer-Based Hyperspectral Image Classification with Spectral Embedding and Permutation-Invariant Neighborhood-Set Attention

Direct adoption of generic vision-transformer tokenization is often inefficient for hyperspectral image (HSI) classification because the data are high-dimensional, spectrally redundant, spatially structured, and label-scarce. This paper presents a multi-benchmark empirical study built on two ideas: linear spectral embe...

Ilkim Aydar, Paul Cotae, Nian Zhang · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.