An Elliptic Kernel Unsupervised Autoencoder—Graph Convolutional Network Ensemble Model for Hyperspectral Unmixing

Abstract

Spectral unmixing is an important technique in remote sensing for analyzing hyperspectral images to identify endmembers and estimate fractional abundance maps. Over the past few decades, significant progress has been made in deep learning methods for endmember extraction and abundance estimation. This article introduces the autoencoder graph ensemble model (AEGEM), a novel ensemble-based framework designed to enhance performance in both endmember extraction and abundance estimation. In the initial stage, endmember extraction and abundance map estimation are carried out using a convolutional autoencoder. An elliptical kernel is then applied to compute spectral distances and generate an adjacency matrix based on elliptical neighborhoods. This information is used to construct an elliptical graph, where centroids serve as senders and surrounding pixels as receivers. A graph convolutional network (GCN) processes stacked input-abundance maps, senders, and receivers to refine the abundance estimations. Finally, an ensemble decision-making strategy selects the optimal abundance maps based on the root-mean-square error metric. The effectiveness of AEGEM is evaluated on benchmark datasets, including Samson, Jasper, and Urban, with additional performance validation on the Cuprite dataset. Experimental results demonstrate that AEGEM outperforms baseline algorithms in both endmember extraction and abundance estimation, particularly in complex and spectrally mixed scenarios.