A Comparative Study of Unsupervised Unmixing Algorithms to Detecting Anomalies in Hyperspectral Images

In this paper, we present a comparative study of several unsupervised unmixing algorithms to anomaly detection in hyperspectral images. The algorithms are called minimum volume constrained non-negative matrix factorization (MVCNMF) [1], gradient descent maximum entropy (GDME) [2] and unsupervised fully constrained least squares (UFCLS) [3] . Several variants of the above algorithms were also implemented and evaluated. Actual hyperspectral image containing 4 panels and 2 small targets from the AF were used in our studies. In our experiments, MVCNMF gets the best detection results when we use the down-sampled image with full bands. Result obtained using UFCLS is close to MVCNMF and better than GDME. In addition, the speed of MVCNMF is the fastest among three methods. Updated algorithms do not provide better result than our previous work. Finally, the dimension reduction using principal component analysis (PCA) operation does affect the final results.