Spatial Density Peak Clustering for Hyperspectral Image Classification With Noisy Labels

The "noisy label" problem is one of the major challenges in hyperspectral image (HSI) classification. In order to address this problem, a spatial density peak (SDP) clustering-based method is proposed to detect mislabeled samples in the training set. Specifically, the proposed methods consist of the following steps: first, the correlation coefficients among the training samples in each class are estimated. In this step, instead of measuring the correlation coefficients by considering individual samples, all neighbor samples or K representative neighbor samples in a local window surrounding each training sample are considered. By this way, the spatial contextual information could be used, and two versions of the proposed method, i.e., measuring the correlation coefficients using all neighbor samples or K representative samples, are referred as SDP and K-SDP, respectively. Second, with the correlation coefficients calculated above, the local density of each training sample can be obtained by the DP clustering algorithm. Finally, those mislabeled samples which usually have lower local densities in each class are able to be identified by a defined decision function. The effectiveness of the proposed detection method is evaluated using a series of spectral and spectral-spatial classification methods on several real hyperspectral data sets.