Regularized Fuzzy Discriminant Analysis for Hyperspectral Image Classification With Noisy Labels

Abstract

Numerous studies have been conducted for hyperspectral image (HSI) classification by assuming that the label information of training data is fully available and correct. However, such an assumption may not always be true in practical applications, which could impact feature extraction methods and eventually compromise the performance of hyperspectral image classification. To address this issue in hyperspectral image classification, we propose a Regularized Fuzzy Discriminant Analysis (RFDA) based feature extraction method to effectively utilize the spatial and spectral information of HSIs with noisy labels. Firstly, the physical properties of HSIs are explored to reconstruct the data. Secondly, the labeled training samples and their unlabeled spatial neighborhood samples are fuzzified using the Fuzzy K-Nearest Neighbor (FKNN) method. Finally, a regularization term using a Fuzzy Locality Preserving Scatter (FLPS) matrix is integrated into fuzzy discriminant analysis, and the spatial-spectral information of HSIs is effectively fused to construct the projection matrix. As a result, the proposed method not only corrects the mislabeled samples effectively, but also preserves the neighborhood relationship among the pixels in the spatial domain and the fundamental structure among the samples in the spectral-domain, which is beneficial for hyperspectral image classification. Experimental results on three synthetic datasets and three public hyperspectral datasets show that our proposed RFDA method outperforms several state-of-the-art feature extraction methods in terms of classification accuracy.