Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 131 of the 132 references it cites were found in a registry or an index.

Blind Hyperspectral Unmixing Using Autoencoders: A Critical Comparison

Open access
Palsson, Burkni, Sveinsson, Johannes R., Ulfarsson, Magnus O. · 2022 · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

Abstract

Deep learning (DL) has heavily impacted the data-intensive field of remote sensing. Autoencoders are a type of DL methods that have been found to be powerful for blind hyperspectral unmixing (HU). HU is the process of resolving the measured spectrum of a pixel into a combination of a set of spectral signatures called endmembers and simultaneously determining their fractional abundances in the pixel. This article details the various autoencoder architectures used in HU and provides a critical comparison of some of the existing published blind unmixing methods based on autoencoders. Eleven different autoencoder methods and one traditional method will be compared in blind unmixing experiments using four real datasets and four synthetic datasets with different spectral variability. Additionally, extensive ablation experiments with a simple spectral unmixing autoencoder will be performed. The results are interpreted in terms of the various implementation details, and the question of why autoencoder methods are so powerful compared to traditional methods is unraveled. The source codes for all methods implemented in this article can be found athttps://github.com/burknipalsson/hu_autoencoders.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1109/jstars.2021.3140154

Cites

131 of 131 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Schematic of a simple autoencoder with a single hidden layer bottleneck.
Fig. 1 Schematic of a simple autoencoder with a single hidden layer bottleneck.CC-BY · From the PDF
Fig. 2 Schematic of a sparse autoencoder with a single hidden layer. Dark hidden units have higher activation strength than the lighter ones.
Fig. 2 Schematic of a sparse autoencoder with a single hidden layer. Dark hidden units have higher activation strength than the lighter ones.CC-BY · From the PDF
Fig. 3 Probabilistic graphical model of a V AE.
Fig. 3 Probabilistic graphical model of a V AE.CC-BY · From the PDF
Fig. 4 Schematic of an adversarial autoencoder.
Fig. 4 Schematic of an adversarial autoencoder.CC-BY · From the PDF
Fig. 5 Schematic of a general unmixing autoencoder. The figure shows an example having two hidden layers, but the actual number could vary from one to several. The decoder has a nonnegative weight constraint and linear activation. The ASC constraint is enforced with a normalizing utility layer or softmax activation.
Fig. 5 Schematic of a general unmixing autoencoder. The figure shows an example having two hidden layers, but the actual number could vary from one to several. The decoder has a nonnegative weight constraint and linear activation. The ASC constraint is enforced with a normalizing utility layer or softmax activation.CC-BY · From the PDF
Fig. 6 RGB images of the datasets used in the experiments. The Samson and Urban images were colored using the technique in [130]. The Apex and Houston images are actual RGB images. (a) Samson-simulated. (b) Urban-simulated. (c) Houston-actual. (d) Apex-actual.
Fig. 6 RGB images of the datasets used in the experiments. The Samson and Urban images were colored using the technique in [130]. The Apex and Houston images are actual RGB images. (a) Samson-simulated. (b) Urban-simulated. (c) Houston-actual. (d) Apex-actual.CC-BY · From the PDF
Fig. 7 Architecture of the simple autoencoder used in the ablation experiments.
Fig. 7 Architecture of the simple autoencoder used in the ablation experiments.CC-BY · From the PDF
Image 8
CC-BY · From the PDF
Image 9
CC-BY · From the PDF
Fig. 10 Comparison between having a linear or nonlinear decoder. The encoder uses the LReLU activation. Four different batch sizes are tested. The graphs in the top row show the average SAD for extracted endmembers, while the bottom row shows the RMSE of the abundances. The black lines show the standard deviation. All experiments consisted of 25 runs. (a) Urban dataset-SAD. (b) Samson dataset-SAD. (c) Houston dataset-SAD. (d) Urban dataset-RMSE. (e) Samson dataset-RMSE. (f) Houston dataset-RMSE.
Fig. 10 Comparison between having a linear or nonlinear decoder. The encoder uses the LReLU activation. Four different batch sizes are tested. The graphs in the top row show the average SAD for extracted endmembers, while the bottom row shows the RMSE of the abundances. The black lines show the standard deviation. All experiments consisted of 25 runs. (a) Urban dataset-SAD. (b) Samson dataset-SAD. (c) Houston dataset-SAD. (d) Urban dataset-RMSE. (e) Samson dataset-RMSE. (f) Houston dataset-RMSE.CC-BY · From the PDF
Fig. 11 Sampled endmembers for the synthetic experiment based on the Urban image for two different sampling radii. (a) Sampled endmembers using radius = 0.05. (b) Sampled endmembers using radius = 0.4.
Fig. 11 Sampled endmembers for the synthetic experiment based on the Urban image for two different sampling radii. (a) Sampled endmembers using radius = 0.05. (b) Sampled endmembers using radius = 0.4.CC-BY · From the PDF
Fig. 12 Bar plot showing the average SAD for the four synthetic datasets. The spectral variability increases with increasing sampling radius. The six rightmost methods all use the MSE objective function. The black vertical lines show the standard deviation. All experiments consisted of 25 runs.
Fig. 12 Bar plot showing the average SAD for the four synthetic datasets. The spectral variability increases with increasing sampling radius. The six rightmost methods all use the MSE objective function. The black vertical lines show the standard deviation. All experiments consisted of 25 runs.CC-BY · From the PDF
Fig. 13 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Urban dataset with four reference endmembers. The number of runs is 25.
Fig. 13 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Urban dataset with four reference endmembers. The number of runs is 25.CC-BY · From the PDF
Fig. 14 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Urban dataset with five reference endmembers. The number of runs is 25.
Fig. 14 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Urban dataset with five reference endmembers. The number of runs is 25.CC-BY · From the PDF
Fig. 15 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Urban dataset with six reference endmembers. The number of runs is 25.
Fig. 15 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Urban dataset with six reference endmembers. The number of runs is 25.CC-BY · From the PDF
Fig. 16 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Houston dataset with four reference endmembers. The number of runs is 25.
Fig. 16 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Houston dataset with four reference endmembers. The number of runs is 25.CC-BY · From the PDF
Fig. 17 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Apex dataset with four reference endmembers. The number of runs is 25.
Fig. 17 Plots of all extracted endmembers by all methods (blue curves) and the reference endmembers (red) for the Apex dataset with four reference endmembers. The number of runs is 25.CC-BY · From the PDF
Fig. 18 Abundance maps for the run with the best mSAD score for all methods for the Urban dataset. The reference abundance maps are in the top row.
Fig. 18 Abundance maps for the run with the best mSAD score for all methods for the Urban dataset. The reference abundance maps are in the top row.CC-BY · From the PDF
Fig. 19 Box plot of the average mSAD of methods grouped by whether their loss function is scale invariant, semi-invariant, or not at all.
Fig. 19 Box plot of the average mSAD of methods grouped by whether their loss function is scale invariant, semi-invariant, or not at all.CC-BY · From the PDF
Fig. 20 Bar plot of the mSAD score for all methods for the Samson dataset with four different levels of SNR and the original uncorrupted dataset. The black vertical lines show the standard deviation. All experiments consisted of 25 runs.
Fig. 20 Bar plot of the mSAD score for all methods for the Samson dataset with four different levels of SNR and the original uncorrupted dataset. The black vertical lines show the standard deviation. All experiments consisted of 25 runs.CC-BY · From the PDF
Fig. 21 Bar plot of the mSAD score for all methods for the synthetic datasets having (a) no spectral variability and (b) moderate spectral variability ( r = 0 . 2 ) and with three different levels of SNR (20, 30, and 40 dB). (a) Synthetic dataset having no spectral variability with three different levels of SNR. (b) Synthetic dataset having moderate spectral variability with three different levels of SNR.
Fig. 21 Bar plot of the mSAD score for all methods for the synthetic datasets having (a) no spectral variability and (b) moderate spectral variability ( r = 0 . 2 ) and with three different levels of SNR (20, 30, and 40 dB). (a) Synthetic dataset having no spectral variability with three different levels of SNR. (b) Synthetic dataset having moderate spectral variability with three different levels of SNR.CC-BY · From the PDF
Image 22
CC-BY · From the PDF
Image 23
CC-BY · From the PDF
Image 24
CC-BY · From the PDF
Image 25
CC-BY · From the PDF

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 132 references

Research works printed but not linked: 3

  1. Goodfellow, 2016, Deep Learning
  2. Chen, 2015, Marginalized stacked denoising autoencoders, 16, 3849
  3. Maas, 2013, Rectifier nonlinearities improve neural network acoustic models, 30, 3

Cited by

See all 114

Metadata

Authors
Palsson, B., Sveinsson, J. R., & Ulfarsson, M. O.
Affiliations
University of Iceland (IS)
Venue
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Published
2022
Volume
15
Pages
1340-1372
ISSN
  • 1939-1404 (ISSN-L)
  • 2151-1535
OpenAlex
W4206620949
Cited by
107
References
132
Citations by year
2026: 16 · 2025: 27 · 2024: 32 · 2023: 24 · 2022: 8
FWCI
8.88
Type
journal article
Language
en
Open access
Locations
  • IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Opin vísindi (Opin vísindi) (submittedVersion, other-oa)
  • Opin vísindi (Opin vísindi) (submittedVersion, cc-by-nc-sa)
Funders
Icelandic Research Fund (174075-05, 207233-052)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Remote-Sensing Image Classification
  • Physical Sciences › Earth and Planetary Sciences › Geochemistry and Petrology › Geochemistry and Geologic Mapping
  • Physical Sciences › Chemistry › Analytical Chemistry › Spectroscopy and Chemometric Analyses
Keywords
autoencoder, endmember extraction, fractional abundance estimation, spectral signatures, ablation study, deep learning for remote sensing

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Blind Hyperspectral Unmixing Using Autoencoders: A Critical Comparison. https://app.cochat.ai/papers/10.1109/jstars.2021.3140154
The paper
Palsson, B., Sveinsson, J. R., & Ulfarsson, M. O. (2022). Blind Hyperspectral Unmixing Using Autoencoders: A Critical Comparison. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. https://doi.org/10.1109/jstars.2021.3140154

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.