Hyperpsectral imaging is an emerging technology used in remote sensing for acquiring images on hundreds of spectral bands at the same time. Often hyperspectral cameras have a small spatial resolution compared to RGB cameras, which often creates mixed pixels. Hyperpsectral unmixing algorithms aim to estimate the pure spectral pixels and their abundances in hyperspectral imagery. Developed unmixing algorithms are used to solve three main steps of the spectral unmixing problem: 1) accurately estimating the possible endmembers in a hyperspectral image or dataset; 2) finding the spectral signatures in the extracted endmembers; 3) calculating the abundances of each endmember in all of the pixels of a hyperspectral image.
Many of the available spectral unmixing algorithms are not capable of performing all of the main unmixing steps or struggle to do so resulting in an inadequate unmixing accuracy. While using an algorithms capable of completing all three steps with high enough accuracy greatly reduces errors between steps, integration and development times. Many of the available algorithms fall into three main categories:
• Sparse regression models. E. g.: CLSUnSAL [1] an upgraded algorithm based on alternating direction method of multipliers that achieved a signal reconstruction error (SRE) score on synthetic data of 21.47 dB and 8.79 dB for 2 and 6 endmembers respectively (with signal-to-noise ration (SNR) of 40 dB).
• Non-negative matrix factorization. E. g.: TV-RSNMF [2] a blind hyperspectral unmixing algorithm implemented with sparse and total variation regularizers. On synthetic dataset algorithm got spectral angle distance (SAD) of 0.0452 (SNR 10 dB) and 0.0060 (SNR 40 dB).
• Autoencoder networks. E. g.: DeepGUn [3] a network made from generative adversarial network and variational autoencoders to use spatial and spectral information of a hyperspectral cube. Reconstruction root mean squared error (RMSE) metric was used to test the performance of this algorithm and a value of 0.0448 was achieved on a synthetic dataset.
Many of these algorithms not only use different base methods to solve the unmixing problem but also these authors create different methodologies and datasets (e.g. synthetic data created from USGS spectral library [5]) to test their methods against each other, because a standardised experiments and universal datasets are not available. This creates a problem when trying to compare the algorithms against each other and trying select the best quality algorithms for any of the three steps. In turn selecting and improving the best available unmixing algorithm is difficult.
In this paper we review the most popular and newest hyperspectral unmixing algorithms, and create a standard-ised algorithm testing methodology to benchmark hyperpsctral unmixing algorithms. To accurately compare them, the algorithms are tested using three different metrics and five different datasets to more accurately asses the performance of solving any of the three unmixing problems. Used datasets: 1) an artificially created dataset in a laboratory [4]; 2) synthetic dataset create from USGS spectral library [5]; 3) IEEE GRSS 2018 data fusion contest hyperspectral dataset; 4) USGS created Cuprite dataset; 5) USGS created Urban dataset. The metric that are used: RMSE; SAD; SRE. In this paper an experiment using this methodology is conducted to test the algorithm robustness to a change in a number of endmembers and the results of this experiment will be presented.