Hyperspectral Unmixing Using Orthogonal Sparse Prior-Based Autoencoder With Hyper-Laplacian Loss and Data-Driven Outlier Detection

Blind Hyperspectral Unmixing Using Autoencoders: A Critical Comparison

“…We will come back to this question when discussing the experimental results. The methods in [41,73,[76][77][78][79][80] employ deep encoders. The method in [81] employs a deep 1D CNN encoder.…”

“…In unmixing, this can cause the abundance maps of certain endmembers to become zerovalued. The works [71,79,80,83,84] all use the ReLU activation. The LReLU activation is a good choice of an activation function as it is non saturating in both directions.…”

“…This implementation also takes care of the ANC constraint and, by choosing large γ achieves something similar to 1 -norm regularization on the abundance maps. The papers [79,80] and [83] use the softmax function without scaling to implement the ASC, while the methods in [73,77] use it with a scaling. The ASC constraint can also be implemented through regularization by adding a penalty term to the loss function that penalizes if the abundances do not sum-to-one.…”

“…This can, however, increase the instability of the training process and the time needed for convergence. The techniques described in [79,80] enforce the ASC through such regularization. A related implementation is to augment the data matrix and the decoder weight's matrix with constant vectors.…”

SNMF-Net: Learning a Deep Alternating Neural Network for Hyperspectral Unmixing

“…al [24] further introduced the Wasserstein distance as a regularization term to better consider the distribution similarity between the observation and the reconstruction. Under the observation that the abundance maps of different endmembers are nearly orthogonal, Dou et al embedded orthogonal sparse prior into the AE to better consider the relationship among abundance maps [25]. Qu et al introduced an untied denoising AE (uDAS) which decouples the decoder from the encoder weights for more accurate endmember extraction [26].…”