EndNet: Sparse AutoEncoder Network for Endmember Extraction and Hyperspectral Unmixing (1708.01894v4)

Abstract: 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.

• The paper introduces EndNet, a sparse autoencoder network for endmember extraction and hyperspectral unmixing, demonstrating an effective unsupervised neural network approach for this task.
• EndNet employs novel strategies like replacing inner products with Spectral Angle Distance (SAD), using zero-biased filters, and a custom loss function to improve discrimination, adhere to simplex assumptions, and enhance model convergence.
• Numerical results show EndNet achieves superior spectral unmixing performance on real datasets compared to state-of-the-art methods, exhibiting robustness in challenging scenarios and scalability for large data applications.

Analysis of "EndNet: Sparse AutoEncoder Network for Endmember Extraction and Hyperspectral Unmixing"

The research paper introduces EndNet, a sparse autoencoder network designed for endmember extraction and hyperspectral unmixing. It presents an innovative approach addressing challenges in hyperspectral unmixing by leveraging neural network architectures in an unsupervised setup.

EndNet employs a two-staged autoencoder network with modifications in the conventional autoencoder structure to better capture the non-linearity and sparsity of hyperspectral data. The authors replace the traditional inner product with spectral angle distance (SAD) to improve the discrimination power among spectral signatures in hyperspectral images. This adaption ensures more accurate extraction of endmembers with signature features that are less sensitive to illumination and other environmental factors.

Strategies Implemented in EndNet

1. Zero-biased Filters: The proposed network eliminates bias terms from the architecture, which prevents them from behaving like constituent materials in pixel mixtures. This adaptation is essential for adhering to the simplex set assumption prevalent in existing geometrical methods.
2. Activation Functions and Layers: The authors integrate ReLU activation functions combined with batch normalization to enhance non-linearity and sparsity in latent representations. These adjustments mitigate parameters from becoming ill-posed during training while selecting only the top layer responses to preserve fractional abundance estimates and improve overall network sparseness.
3. Introduction of Spectral Angle Distance (SAD): The replacement of inner products with SAD enhances the discriminative capacity of the network, fostering better separation of pixel data. SAD is particularly advantageous for maintaining robustness against nonlinear interactions observed in hyperspectral datasets, surpassing limitations associated with Euclidean distance metrics.
4. Novel Loss Function: The objective function introduces new penalty terms including a Kullback-Leibler divergence combined with the SAD metric to maintain angular similarity between the original and reconstructed data samples. This modification reduces the risk of local minima and refines model convergence.
5. Data Adaptations and Scalability: The paper emphasizes the stochastic-gradient approach enabling scalability for large hyperspectral datasets, making EndNet suitable for real-time and extensive data applications. Additionally, pre-initialization with geometrical algorithms like VCA expedites convergence towards optimal solutions.

Numerical Results and Comparisons

Extensive evaluations across multiple real-world hyperspectral datasets reveal that EndNet delivers superior spectral unmixing performance, often outperforming state-of-the-art methods including VCA, SCM, and others that leverage spatial priors. The improvements are notable both in endmember extraction and fractional abundance estimation with significant reductions in SAD and RMSE metrics.

The results emphasize EndNet's robustness in scenarios with severe spectral variability, illumination inconsistencies, and closely correlated material mixtures. Despite using an unsupervised neural network setup, EndNet was able to maintain or exceed the accuracy of methods employing supervised data or spatial priors.

Implications and Future Directions

EndNet represents a significant advance in hyperspectral unmixing methods, opening avenues for further research into neural network-based approaches in handling intricate pixel mixtures within hyperspectral images. The method's scalability and adaptability make it a promising candidate for real-time processing applications including environmental monitoring, mineral identification, and military surveillance.

Future research could explore integrating more sophisticated neural architectures or combining unsupervised and supervised learning paradigms, providing enhancements in accuracy and computational efficiency. Additionally, expanding EndNet's applicability to heavily corrupted or incomplete datasets could augment its utility in practical remote sensing scenarios.

In conclusion, the paper demonstrates a successful and innovative adaptation of autoencoder neural networks to hyperspectral unmixing, marking a potential shift towards leveraging deep learning in this domain. The insights from the paper could lead to more robust, accurate, and scalable hyperspectral analysis techniques in the future.