Archetypal Analysis for Sparse Representation-based Hyperspectral\n Sub-pixel Quantification

Lukas Drees, Ribana Roscher, Susanne WenzelarXiv 2018 · DOI · arXiv

Abstract

The estimation of land cover fractions from remote sensing images is a frequently used indicator of environmental quality. This paper focuses on the quantification of land cover fractions in an urban area of Berlin, Germany, using simulated hyperspectral EnMAP data with a spatial resolution of 30m x 30m. We use constrained sparse representation, where each pixel with unknown surface characteristics is expressed by a weighted linear combination of elementary spectra with known land cover class. We automatically determine the elementary spectra from image reference data using archetypal analysis by simplex volume maximization and combine it with the reversible jump Markov chain Monte Carlo (rjMCMC) method. In our experiments, the estimation of the automatically derived elementary spectra is compared to the estimation obtained by a manually designed spectral library by means of reconstruction error, mean absolute error of the fraction estimates, sum of fractions, $R^2$, and the number of used elementary spectra. The experiments show that a collection of archetypes can be an adequate and efficient alternative to the manually designed spectral library with respect to the mentioned criteria.

Introduction

In urban remote sensing, sub-pixel quantification is a critical task. Satellite sensors like EnMAP capture data at a spatial resolution (e.g., 30m x 30m) where individual pixels often represent a mixture of different materials, such as asphalt, vegetation, and soil. To accurately assess environmental quality, these "mixed" pixels must be decomposed into their constituent land cover fractions—a process known as spectral unmixing.

Traditionally, unmixing relies on a spectral library: a collection of "pure" spectral signatures, or endmembers, representing specific materials. However, building these libraries manually is labor-intensive, subjective, and often lacks the flexibility to adapt to different geographical regions or sensor characteristics. This study proposes an automated pipeline to extract representative elementary spectra directly from hyperspectral data using archetypal analysis.

Methodology

The proposed approach utilizes a three-stage pipeline to automate the quantification process:

1. Archetypal Analysis via SiVM

The researchers employ Simplex Volume Maximization (SiVM) to identify "archetypes"—data points that reside on the convex hull of the data distribution. These archetypes represent the most extreme spectral signatures in the dataset, allowing any other pixel to be represented as a weighted combination of these points.

2. Kernel Space Mapping

Because real-world spectral mixtures often exhibit non-linear characteristics, the archetypal analysis is performed in a kernel space using a Gaussian radial basis function. This allows the model to identify archetypes on a "concave hull," better capturing the inherent variability of urban materials.

3. rjMCMC Optimization

To determine the optimal number and composition of archetypes, the Reversible Jump Markov Chain Monte Carlo (rjMCMC) algorithm is used. This method allows the model to "jump" between different dimensions, testing various library sizes to find a sparse subset of archetypes that minimizes reconstruction error without redundancy.

Representative spectral signatures for urban land cover classes (impervious surface, vegetation, soil, and water) used in the unmixing process.

Experimental Setup and Datasets

The study utilized the Berlin-Urban-Gradient dataset, featuring high-resolution HyMap data (3.6m and 9m) to extract reference information and simulated 30m EnMAP data for testing. The performance of the automated archetypal library (AA-Opt) was benchmarked against a gold-standard, expert-curated library (LibCom) containing 75 spectra.

Results and Discussion

The automated AA-Opt library achieved a Mean Absolute Error (MAE) of 10.16%, performing remarkably close to the expert-designed library's 8.15%. Notably, the rjMCMC process discovered that many spectra in the manual library were redundant; an optimized manual library required only ~44 spectra to maintain performance.

Comparison of reconstruction error between the automated archetypal library and the manually designed library across varying library sizes (W).

As shown in the reconstruction error analysis, the archetypal library converges at a rate nearly identical to the manually designed library. This suggests that data-driven archetypes are as effective at representing the spectral space as those selected by human experts. The system showed particular strength in identifying impervious surfaces and vegetation, though it faced challenges with soil and water due to their lower spectral diversity in the urban study area.

Conclusion

This research demonstrates that automated archetypal analysis, combined with rjMCMC optimization, provides a viable and efficient alternative to manual spectral library design. By removing the need for subjective expert intervention, this approach facilitates more scalable and adaptable environmental monitoring. Future work will focus on incorporating spatial context and refining the classification of discovered archetypes to further improve sub-pixel accuracy.

Summary updated

Abstract

The estimation of land cover fractions from remote sensing images is a\nfrequently used indicator of the environmental quality. This paper focuses on\nthe quantification of land cover fractions in an urban area of Berlin, Germany,\nusing simulated hyperspectral EnMAP data with a spatial resolution of\n30m$\\times$30m. We use constrained sparse representation, where each pixel with\nunknown surface characteristics is expressed by a weighted linear combination\nof elementary spectra with known land cover class. We automatically determine\nthe elementary spectra from image reference data using archetypal analysis by\nsimplex volume maximization, and combine it with reversible jump Markov chain\nMonte Carlo method. In our experiments, the estimation of the automatically\nderived elementary spectra is compared to the estimation obtained by a manually\ndesigned spectral library by means of reconstruction error, mean absolute error\nof the fraction estimates, sum of fractions, $R^2$, and the number of used\nelementary spectra. The experiments show that a collection of archetypes can be\nan adequate and efficient alternative to the manually designed spectral library\nwith respect to the mentioned criteria.\n
BibTeX
@misc{drees2018archetypal,
  author = {Lukas Drees and Ribana Roscher and Susanne Wenzel},
  title = {{Archetypal Analysis for Sparse Representation-based Hyperspectral\textbackslash\{\}n Sub-pixel Quantification}},
  year = {2018},
  eprint = {1802.02813},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.02813},
}
Figure