NMF-SAE: An Interpretable Sparse Autoencoder for Hyperspectral Unmixing

Abstract

Hyperspectral unmixing is an important tool to learn the material constitution and distribution of a scene. Modelbased unmixing methods depend on well-designed iterative optimization algorithms, which is usually time consuming. Learning-based methods perform unmixing in a data-driven manner but heavily rely on the quality and quantity of the training samples due to the lack of physical interpretability. In this paper, we combine the advantages of both modelbased and learning-based methods and propose a nonnegative matrix factorization (NMF) inspired sparse autoencoder (NMF-SAE) for hyperspectral unmixing. NMF-SAE consists of an encoder and a decoder, both of which are constructed by unrolling the iterative optimization rules of L 1 sparsityconstrained NMF for the linear spectral mixture model. All parameters in our method are obtained by end-to-end training in a data-driven manner. Our network is not only physically interpretable and flexible but also has higher learning capacity with fewer parameters. Experimental results on both synthetic and real-world data demonstrate that our method is capable of producing desirable unmixing results when compared against several alternative approaches.

Blind nonlinear unmixing using nonnegative matrix factorization based bi-objective autoencoder

“…In this section, the details of the datasets used and the results obtained are presented. For the performance comparison, the models TV-RSNMF [15], DAEN [19], and NMF-SAE [21] are considered. The performance analysis is done using RMSE and SAD defined in section 2.3.3…”

“…The comparison of RMSE values with the standard deviation obtained for the synthetic dataset using the different models is illustrated in Table 1. The models considered for comparison are TV-RSNMF [15], DAEN [19], and NMF-SAE [21]. The proposed model achieved better performance for Ammonijarosite, Almandine, and Axinite compared to other models.…”

“…These objective functions are solved using the autoencoder designed motivated on [21]. The autoencoder pairs E1, E2, D1, and D2 are used to update the endmember and the abundance factors using the proximal decent approach.…”

“…The loss functions help the encoder to learn the required features. The spectral angular distance (SAD) is calculated using (21) and the total loss is calculated using (22).…”

“…Su et al [19] suggested a stacked and variational autoencoders were used for the unsupervised unmixing. The various popular autoencoder models are, a probability metric-based autoencoder model [20], a sparse autoencoder model [21], and sparsity-based autoencoder with orthogonal sparse prior [22]. In this article, a dual branch autoencoder was proposed to implement this novel orthogonal sparsity-based blind unmixing.…”