Hierarchical Structure-Based Noisy Labels Detection for Hyperspectral Image Classification

Abstract

In hyperspectral image (HSI) classification, the performance of supervised learning tends to be affected by priori knowledge, i.e., quantity and quality of samples. However, it is inevitable to limit the performance of supervised classification due to the presence of noisy labels in the training samples. In this article, we first propose a hierarchical constrained energy minimum (HCEM) method to detect mislabeled samples (noisy labels) of original training set trained with supervised task and boost the performance of classifiers and spectral-spatial classification methods in HSI applications. The basic idea behind this work is that the filter output energy of noisy label spectrum is hierarchically suppressed based on the cascade of CEM detectors at different layers. The proposed HCEM method consists of four key steps: First, the distance information among samples is obtained to calculate the spectral similarity between samples per class in original training set. Then, the confidence spectra per class is constructed according to the maximum similarity domain. Next, the spectra of mislabeled samples is suppressed and the spectra of true samples is preserved by multilayers CEM detector. Finally, the noisy labels are detected and removed based on the output evaluated by the proposed HCEM method. Experimental results on four real hyperspectral datasets are verified by a series of spectral classifiers and spectral-spatial classification methods, it demonstrated that the proposed HCEM method, can accurately remove noisy labels of original training set and effectively improve the performance of supervised classification task.