PaperlayerView exact revision on arXiv

Source-linked AI summary

EndNet: Sparse AutoEncoder Network for Endmember Extraction and Hyperspectral Unmixing

Savas Ozkan, Berk Kaya, Gozde Bozdagi Akar

arXiv:1708.01894v4cs.CV

TL;DR

Hyperspectral unmixing must recover material signatures and fractions from pixels containing mixed spectra, while existing autoencoder designs have important limitations. EndNet restructures the autoencoder with specialized layers, SAD-based projections, sparsity penalties, and stochastic-gradient optimization. Experiments on well-known datasets report improved performance over state-of-the-art methods, including methods using spatial priors.

  • Problem

    Limited sensor spatial resolution produces mixed hyperspectral pixels, while conventional autoencoders can suffer from redundant responses, weakly discriminative inner products, insufficient regularization, and overfitting.

  • Method

    EndNet is a two-staged unsupervised autoencoder that adds architectural layers, replaces encoder inner products with SAD, and uses penalty terms and stochastic-gradient optimization for joint unmixing.

  • Results

    EndNet reports significant performance improvements on well-known hyperspectral datasets and outperforms DgS-NMF despite DgS-NMF using spatial scene priors.

  • Takeaways & Limitations

    The study presents an unsupervised end-to-end neural-network approach that addresses nonlinearity, sparsity, and physical constraints in hyperspectral unmixing.

Abstract

from arXiv · show

Data acquired from multi-channel sensors is a highly valuable asset to interpret the environment for a variety of remote sensing applications. However, low spatial resolution is a critical limitation for previous sensors and the constituent materials of a scene can be mixed in different fractions due to their spatial interactions. Spectral unmixing is a technique that allows us to obtain the material spectral signatures and their fractions from hyperspectral data. In this paper, we propose a novel endmember extraction and hyperspectral unmixing scheme, so called \textit{EndNet}, that is based on a two-staged autoencoder network. This well-known structure is completely enhanced and restructured by introducing additional layers and a projection metric (i.e., spectral angle distance (SAD) instead of inner product) to achieve an optimum solution. Moreover, we present a novel loss function that is composed of a Kullback-Leibler divergence term with SAD similarity and additional penalty terms to improve the sparsity of the estimates. These modifications enable us to set the common properties of endmembers such as non-linearity and sparsity for autoencoder networks. Lastly, due to the stochastic-gradient based approach, the method is scalable for large-scale data and it can be accelerated on Graphical Processing Units (GPUs). To demonstrate the superiority of our proposed method, we conduct extensive experiments on several well-known datasets. The results confirm that the proposed method considerably improves the performance compared to the state-of-the-art techniques in literature.

I. INTRODUCTION

Hyperspectral pixels often mix constituent material spectra because sensor spatial resolution is limited, motivating unsupervised extraction of endmembers and abundances. EndNet addresses unresolved unsupervised autoencoder adaptation with a two-stage network and specialized constraints, metrics, penalties, and optimization.

  • Motivation: Limited spatial resolution causes hyperspectral pixels to contain mixtures of material spectra, so unmixing estimates endmembers and their pixel-wise fractions.The mixture model includes endmembers, abundances, and a noise term representing sources such as sensor readout noise or illumination variability.
  • Limitations of existing approaches: Linear models cannot adequately handle multiple scattering, microscopic mixtures, or water-absorbed environments, while nonlinear projections can reduce scalability through high computational workload.Kernel-based alternatives introduce distortion into endmember and abundance estimates.
  • Unsupervised setting: Supervised neural approaches depend strongly on costly, impractical labels, leaving endmembers and abundances to be estimated blindly in unsupervised settings.The paper identifies unsupervised autoencoder adaptation as incompletely solved in prior work.
  • EndNet contribution: EndNet uses a two-staged autoencoder and end-to-end learning scheme specialized to extract endmembers and fractional abundances without supervision.The approach introduces novel features beyond the traditional autoencoder pipeline.
  • EndNet contribution: The method removes bias terms, uses ReLU and normalization for sparsity and nonlinearity, replaces inner products with SAD, and adds penalty terms for angular similarity, reconstruction, and sparse representations.Non-negativity and sum-to-one constraints are preserved through the architecture and learning process.
  • Optimization: Stochastic-gradient optimization jointly estimates endmembers and abundances, while VCA-like initialization helps filters converge faster toward the global optimum.The optimization is presented as scalable to large-scale data.

II. SPECTRAL UNMIXING

A conventional autoencoder maps hyperspectral samples to hidden representations and reconstructs them, but its standard design can produce unstable unmixing solutions. The paper identifies problems involving activation redundancy, weak discrimination from inner products, insufficient regularization, and convergence or sparsity issues.

  • Conventional autoencoder: An autoencoder transforms an input sample into a hidden representation and reconstructs the original sample from that representation.The encoder uses trainable parameters, while the decoder restores the input with the objective of high accuracy.
  • Observed failure modes: A conventional autoencoder can produce bias responses that behave like scene endmembers, while its estimated signatures may violate non-negativity.The figure compares decoder parameters with the bias response and shows the physical-constraint problem.
  • Conventional autoencoder: Conventional formulations commonly use element-wise nonlinear activations such as sigmoid or tanh and optimize standard Euclidean reconstruction error.The reconstruction loss compares original and reconstructed samples.
  • Current limitations: The paper identifies redundant logistic responses, insufficiently discriminative inner products, and inadequately domain-regularized Euclidean reconstruction as limitations for unmixing.These issues can reduce sparsity and produce erroneous fractional abundances.
  • Interpretation: In hyperspectral unmixing, the first hidden representation intuitively corresponds to per-pixel fractional abundances.This interpretation connects the autoencoder's latent responses to the unmixing variables.

A. Sparse Autoencoder for Hyperspectral Unmixing

EndNet adapts a two-staged autoencoder for hyperspectral unmixing by restructuring its architecture, replacing inner products with SAD, and adding sparsity-oriented optimization terms. The resulting design represents endmembers and abundances while addressing nonlinearity, sparsity, separability, and parameter stability.

  • EndNet adapts an autoencoder pipeline to extract endmembers and estimate pixel abundances while preserving latent correlations in hyperspectral data.
  • The architecture removes biases, uses disjoint encoder and decoder parameters, and omits the decoder activation so decoder weights represent endmembers and hidden abstractions represent abundances.
  • A normalization layer before encoder ReLU mitigates ill-posed parameter effects while scaling, reordering, and selectively emphasizing activation responses.
  • Dropout, top-2 activation selection, and l1 normalization improve generalization, selectivity, and the sum-to-one constraint for abundance estimates.The dropout probability is scene-dependent, with the default set to 1.0.
  • SAD replaces encoder inner products because it separates coherent from same-material signatures more effectively, producing more discriminative hidden abstractions and endmembers.SAD similarity is combined with Euclidean reconstruction, Kullback-Leibler divergence, and l1 sparsity penalties in the loss.

B. Fractional Abundance Estimation

EndNet estimates fractional abundances either from hidden representations or by solving an inverse problem with decoder-derived endmembers. The inverse-problem approach combined with SPU produces further empirical improvements, so reported abundances use both methods.

  • Fractional abundances can be obtained from hidden abstracts y or by solving an inverse problem with estimated decoder endmembers W(d).
  • Batch-normalization effects can alter test-time hidden representations, motivating SPU with an SAD kernel for abundance estimation.
  • The inverse-problem and SPU combination empirically improves performance, although hidden abstracts still provide compatible abundance results.

III. EXPERIMENTS

The experiments report quantitative and qualitative hyperspectral-unmixing results, comparing estimated signatures and abundances with ground truth. Urban abundance results are visualized by material columns.

  • The study reports quantitative and qualitative results for the hyperspectral-unmixing problem.Estimated spectral signatures and fractional abundances are compared with corresponding ground truth.
  • The experimental section summarizes publicly available datasets, baseline methods, evaluation metrics, and parameter settings.
  • Urban abundance visualizations organize Asphalt, Grass, Tree, and Roof by columns.

A. Hyperspectral Datasets

The experiments use synthetic and real hyperspectral datasets to assess endmember extraction and unmixing. The real-data examples include Urban, Samson, Jasper Ridge, and Cuprite scenes.

  • The evaluation uses synthetic data and extensively used real datasets for endmember extraction and unmixing.Synthetic data provides true abundances and endmembers, while real data represents cases such as extreme nonlinearity that are difficult to simulate.
  • Urban signatures are visualized for Asphalt, Grass, Tree, and Roof using ground-truth and EndNet-DMaxD curves.
  • The Samson dataset contains Soil, Tree, and Water materials, while Jasper Ridge contains Tree, Water, Soil, and Road.

B. Baselines, Metrics and Parameter Settings

EndNet is compared with open-source and reported hyperspectral-unmixing methods using SAD and RMSE, alongside qualitative signature and abundance analyses. Parameter settings and computational scaling are also described.

  • Baselines: Baselines include VCA, MVSA, SPICE, SCM, and DMaxD, alongside reported results for l1|2-NMF and DgS-NMF.
  • Metrics: SAD evaluates estimated endmember quality against ground truth, while RMSE assesses estimated abundance accuracy.
  • Parameter Settings: Abundance estimation for MVSA, VCA, and DMaxD uses MLM, which yielded the best performance among the tested abundance models.
  • Evaluation: On Cuprite, SAD is reported without quantitative abundance results because the dataset lacks fractional-abundance ground truth.
  • Parameter Settings: EndNet requires scene-dependent tuning of dropout p, sparsity weight λ2, and mask-noise percentage, with defaults p = 1.0, λ2 = 0.1, and 40% mask noise.
  • Computational Complexity: The stochastic-gradient solver makes computation scalable through batch-based learning, with time and memory practically independent of data size.

C. Experiments on Hyperspectral Unmixing Datasets

The experiments evaluate EndNet-DMaxD against baseline methods on several hyperspectral unmixing datasets using quantitative metrics and qualitative abundance maps. EndNet generally achieves the strongest reported performance, including on scenes with highly mixed or correlated materials.

  • Evaluation setup: Experiments compare methods on Urban, Samson, Jasper Ridge, and Cuprite datasets using repeated tests and SAD, RMSE, and qualitative abundance evaluations.Tests are repeated 20 times, with mean and standard deviation reported where quantitative ground truth is available.
  • Urban: EndNet-DMaxD achieves the best overall Urban performance, improving approximately 4.2% and 1.2% over the second-best result for SAD and RMSE.The second-best method is DgS-NMF, which exploits spatial priors.
  • Dataset comparisons: Approximately 2% and 3% improvements are introduced for SAD and RMSE metrics respectively in another evaluated dataset.The corresponding qualitative abundance differences are generally small, with errors concentrated near material boundaries and water-ground intersections.
  • EndNet variants: EndNet-DMaxD maintains exceptional performance where EndNet-VCA decreases sharply, while DMaxD favors more mutually uncorrelated estimated spectral signatures.The paper attributes this difference to DMaxD maximizing distances among estimated signatures.
  • Cuprite: EndNet achieves the best SAD performance on Cuprite, with a 3% improvement over the second-best result despite the absence of abundance ground truth.Qualitative abundance estimates are compared visually and are reported as consistent with prior visual results.
  • Abundance estimation: SPU further improves abundance estimates over EndNet-DMaxD on three datasets, while EndNet-DMaxD achieves similar or better performance than baseline methods.The comparison is reported in Table VII.

D. Experiments on Hyperspectral Classification Datasets

The classification-dataset experiments assess SAD and qualitative abundance maps on University of Pavia and Mississippi Gulfport. EndNet-DMaxD obtains the strongest reported overall results and better abundance maps than the baselines.

  • Evaluation setup: University of Pavia and Mississippi Gulfport are evaluated with repeated tests using SAD and qualitative fractional-abundance comparisons.The datasets provide pixel-wise class labels, while abundance RMSE ground truth is unavailable.
  • Evaluation setup: Average spectra of pixels for each material are used as ground truth in SAD comparisons because endmember ground truth is unavailable.This ground-truth construction follows the cited prior approach.
  • University of Pavia: The proposed method achieves the best SAD performance on University of Pavia, with a 3% improvement over the second-best result.The method parameters and baseline settings are tuned for the dataset’s correlated materials.
  • University of Pavia: EndNet-DMaxD produces meaningful abundance estimates on University of Pavia and reduces repeated baseline confusion between Trees and Grass/Meadow.The paper links this result to the importance of abundance estimation for distinguishing scene classes.
  • Mississippi Gulfport: On Mississippi Gulfport, EndNet-DMaxD yields the highest overall SAD score and nearly optimum abundance estimates for Tree, Building, and Shadow.Baseline methods show comparatively high misprediction rates for these materials.

IV. CONCLUSION

The conclusion presents EndNet as an autoencoder-based method for unsupervised endmember extraction and hyperspectral unmixing. Its architectural and loss-function changes address nonlinearity, sparsity, and physical constraints, while experiments report improved performance over baselines.

  • Method: EndNet restructures a conventional autoencoder with additional layers and a novel loss function for endmember extraction and hyperspectral unmixing.The method is trained end to end.
  • Method: The proposed modifications address nonlinearity, sparsity, and some physical constraints in hyperspectral unmixing.These properties are identified as common challenges of the task.
  • Optimization: Stochastic-gradient backpropagation scales the method to large-scale data and supports GPU acceleration.The paper contrasts this with more complex derivations and inferences in the literature.
  • Abundance estimation: Replacing the standard l2 norm kernel with a SAD kernel further improves fractional-abundance estimation from estimated endmembers.This adaptation is implemented through the SPU algorithm.
  • Results: Experiments report significant performance improvements on well-known hyperspectral datasets, including results outperforming DgS-NMF despite its use of spatial priors.The comparison is presented as evidence against methods using additional scene spatial information.
  • Conclusion: The paper characterizes EndNet as an end-to-end neural-network algorithm achieving superior unsupervised hyperspectral-unmixing performance.The authors position the findings as a basis for further neural-network studies in this domain.
Loading 1708.01894v4…