[Paper Review] Unmixing urban hyperspectral imagery with a Gaussian mixture model on endmember variability

Yuan Zhou, Erin B. Wetherley+1 more|arXiv (Cornell University)|Jan 25, 2018

Urban Heat Island Mitigation5 references3 citations

TL;DR

This paper proposes a Gaussian Mixture Model (GMM)-based spectral unmixing method that models endmember variability in urban hyperspectral imagery, enabling accurate abundance estimation without manual library reduction. The approach achieves accuracy comparable to MESMA—especially with large spectral libraries—while being more robust across datasets due to its scalable, distribution-based formulation.

ABSTRACT

In this paper, we model a pixel as a linear combination of endmembers sampled from probability distributions of Gaussian mixture models (GMM). The parameters of the GMM distributions are estimated using spectral libraries. Abundances are estimated based on the distribution parameters. The advantage of this algorithm is that the model size grows very slowly as a function of the library size. To validate this method, we used data collected by the AVIRIS sensor over the Santa Barbara region: two 16 m spatial resolution and two 4 m spatial resolution images. 64 validated regions of interest (ROI) (180 m by 180 m) were used to assess estimate accuracy. Ground truth was obtained using 1 m images leading to the following 6 classes: turfgrass, non-photosynthetic vegetation (NPV), paved, roof, soil, and tree. Spectral libraries were built by manually identifying and extracting pure spectra from both resolution images, resulting in 3,287 spectra at 16 m and 15,426 spectra at 4 m. We then unmixed ROIs of each resolution using the following unmixing algorithms: the set-based algorithms MESMA and AAM, and the distribution-based algorithms GMM, NCM, and BCM. The original libraries were used for the distribution-based algorithms whereas set-based methods required a sophisticated reduction method, resulting in reduced libraries of 61 spectra at 16 m and 95 spectra at 4 m. The results show that GMM performs best among the distribution-based methods, producing comparable accuracy to MESMA, and may be more robust across datasets.

Motivation & Objective

• Address the challenge of endmember variability in urban hyperspectral unmixing, where spectral signatures of the same material vary due to illumination, geometry, and scale.
• Overcome the computational burden and heuristic limitations of traditional set-based unmixing methods like MESMA, which require extensive library reduction for large spectral libraries.
• Develop a scalable, distribution-based unmixing framework that models endmembers as probability distributions using Gaussian Mixture Models (GMM), reducing dependency on manual library curation.
• Validate the method on AVIRIS data across 16 m and 4 m spatial resolutions, using high-resolution ground truth to assess accuracy across six land cover classes.
• Demonstrate that the GMM-based approach maintains high accuracy even when using the full spectral library, unlike set-based methods that require reduced, manually curated libraries.

Proposed method

• Model each pixel as a linear combination of endmembers sampled from Gaussian Mixture Models (GMMs), where endmember spectral variability is captured through learned GMM parameters.
• Estimate GMM parameters (means, covariances, mixing weights) from large spectral libraries extracted from both 16 m and 4 m resolution AVIRIS images.
• Use a cross-validation-based approach (T_CVIC) to determine the optimal number of GMM components per endmember class, balancing accuracy and computational cost.
• Apply the GMM framework to estimate abundances per pixel by computing the most probable combination of endmembers from the learned GMM distributions.
• Incorporate spatial priors in some experiments, though they showed inconsistent performance and were ultimately not used in the final evaluation due to minimal improvement.
• Compare the GMM method against set-based methods (MESMA, AAM) and other distribution-based methods (NCM, BCM) using identical spectral libraries and ground truth.

Experimental results

Research questions

• RQ1 Can a GMM-based unmixing framework achieve unmixing accuracy comparable to MESMA while avoiding the need for manual library reduction?
• RQ2 How does the GMM-based method perform across different spatial resolutions (16 m vs. 4 m) when using the full spectral library?
• RQ3 Is the GMM-based method more robust across datasets than set-based methods that rely on library reduction tailored to a specific dataset?
• RQ4 How does endmember variability, modeled via GMMs, affect abundance estimation accuracy compared to fixed-endmember models?
• RQ5 What is the impact of spatial priors on GMM-based unmixing performance in heterogeneous urban scenes?

Key findings

• The GMM-based unmixing method achieved accuracy comparable to MESMA, with an R² of 0.867 for the 16 m data and 0.923 for the 4 m data—matching results reported in prior work using the same dataset.
• For the 16 m data, GMM slightly outperformed MESMA in terms of R², while for the 4 m data, the performance was slightly inferior but still highly competitive.
• The GMM method maintained high accuracy without requiring any manual library reduction, unlike MESMA, which relied on a reduced library of 61 (16 m) and 95 (4 m) spectra derived through user-guided selection.
• The GMM method demonstrated greater robustness across datasets, as it did not depend on dataset-specific library reduction, which can limit generalization.
• Spatial priors did not significantly improve performance, likely due to the relatively large pixel size (16 m and 4 m) and scene heterogeneity, which reduced the effectiveness of spatial smoothness assumptions.
• The dataset's ground truth contained minor errors due to misregistration between AVIRIS and high-resolution images, which may have contributed to slightly higher errors in the 16 m data, though this was mitigated by careful image registration in future work.