Nonnegative collaborative representation for hyperspectral anomaly detection

Recently, collaborative representation detector (CRD) is widely studied since it is easy to implement and offers higher detection accuracy for hyperspectral anomaly detection. However, the original CRD introduces a regularization parameter that is hard to tune in practice; and the coefficient vector under the sum-to-one constraint can be negative, which has no physical meaning, thus leading to unstable results. To solve these problems, we impose strict nonnegative and l(1)-norm constraints on the coefficient vector to obtain a more accurate and meaningful coefficient vector. To facilitate the optimization procedure, we apply the accelerated projected gradient algorithm to solve this model. The experimental results on four real datasets demonstrate the superior performance of the proposed method compared with CRD.