Noisy Labels Detection in Hyperspectral Image via Class-Dependent Collaborative Representation

Recently, it has been proven that noisy label is a pivotal problem to be solved for hyperspectral image (HSI) classification. In this article, we propose a new noisy label detection method for HSI classification that is based on the class-dependent collaborative representation (CDCR) algorithm. First, a Tikhonov regularization-based weight matrix among the training samples is calculated, which can estimate the differences among the training samples. Then, the residual of each training sample can be obtained by the CDCR method. Specifically, the larger the residual is, the greater probability belonging to a mislabeled sample will be. Finally, the noisy labels are detected by the defined decision function. Experiments based on the different real hyperspectral datasets demonstrate the effectiveness of the proposed method in detecting noisy labels and improving classification performance of HSI.