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

TL;DR

This paper tackles hyperspectral unmixing by presenting AEGEM, an ensemble framework that couples a convolutional autoencoder with a graph convolutional network. A key novelty is the elliptical kernel that defines an elliptical neighborhood to build an elliptical graph whose spectral-angle-based relations feed a GCN that refines abundance maps; an ensemble RMSE-based decision selects the best maps. The approach demonstrates superior endmember extraction and abundance estimation on Samson, Jasper Ridge, and Urban datasets, outperforming baselines such as CNNAEU, UnDIP, and SGSNMF, with notable gains for water, roof, asphalt, and other materials. The work offers a scalable, spatially-aware SU solution with potential applications in large-area land-cover mapping and change detection, supported by public data and NASA-funded validation.

Abstract

Spectral Unmixing is an important technique in remote sensing used to analyze hyperspectral images to identify endmembers and estimate abundance maps. Over the past few decades, performance of techniques for endmember extraction and fractional abundance map estimation have significantly improved. This article presents an ensemble model workflow called Autoencoder Graph Ensemble Model (AEGEM) designed to extract endmembers and fractional abundance maps. An elliptical kernel is applied to measure spectral distances, generating the adjacency matrix within the elliptical neighborhood. This information is used to construct an elliptical graph, with centroids as senders and remaining pixels within the geometry as receivers. The next step involves stacking abundance maps, senders, and receivers as inputs to a Graph Convolutional Network, which processes this input to refine abundance maps. Finally, an ensemble decision-making process determines the best abundance maps based on root mean square error metric. The proposed AEGEM is assessed with benchmark datasets such as Samson, Jasper, and Urban, outperforming results obtained by baseline algorithms. For the Samson dataset, AEGEM excels in three abundance maps: water, tree and soil yielding values of 0.081, 0.158, and 0.182, respectively. For the Jasper dataset, results are improved for the tree and water endmembers with values of 0.035 and 0.060 in that order, as well as for the mean average of the spectral angle distance metric 0.109. For the Urban dataset, AEGEM outperforms previous results for the abundance maps of roof and asphalt, achieving values of 0.135 and 0.240, respectively. Additionally, for the endmembers of grass and roof, AEGEM achieves values of 0.063 and 0.094.

Figures (13)

• Figure 1: The spectral unmixing workflow designed for extracting endmembers and fractional abundance maps consists of eight stages. First, the HSI is loaded. Subsequently, a convolutional autoencoder is employed to extract endmembers and fractional abundances maps. An elliptical kernel is then applied to construct an elliptical graph, where centroids serve as senders and the remaining pixels within the geometry act as receivers to measure spectral distances, thereby generating the adjacency matrix within the elliptical neighborhood. The next step involves stacking abundance maps, senders, and receivers as inputs for the GCN. The GCN processes this input to refine abundance maps. Finally, an ensemble decision-making process determines the best abundance maps based on the RMSE metric.
• Figure 2: Architectural Illustration of the convolutional autoencoder with its respective operations and activation functions used to perform spectral unmixing analysis. In the encoder, abundance maps estimation is conducted, while in the decoder, endmembers extraction is performed.
• Figure 3: Illustration of the elliptical kernel with the locations of their respective centroids. The intersections in each ellipse are highlighted in brown. This elliptical kernel is used to conduct similarity measurements for the elliptical graph in the GCN.
• Figure 4: Architectural illustration of the GCN, showcasing the operations conducted and the activation functions used for abundance map extraction. The depiction includes the preprocessing stage with the HSI, the abundance maps, and the ground truth.
• Figure 5: Samson dataset with corresponding endmembers and abundances maps for tree, soil, and water.