TwoWin-SOVA: Two Windows Discrete Cosine Transform and Synthetic Minority Oversampling Technique One-Versus-All Ensemble Classifiers for Imbalanced Hyperspectral Image Explainable Classification

Patchwise methods have been widely used in hyperspectral image (HSI) classification. In HSI classification methods that use only spectral information, a large number of unlabeled pixels, known as background pixels, are removed before training, which loses spatial information. Patchwise methods obtain patches through a local window to retain the spectral–spatial information in background pixels. Larger patches have more background pixels, but more redundant information in the extracted high-dimensional features reduces classification accuracy. Due to the complex and uneven distribution of land cover classes, class imbalance is a common problem in HSI. In this article, a two windows discrete cosine transform and synthetic minority oversampling technique one-versus-all ensemble classifiers (TwoWin-SOVA) method is proposed for addressing the problems of high dimensionality caused by large patches and class imbalance, which consists of feature extraction and ensemble classifiers. A two windows discrete cosine transform (TwoWin-DCT) is utilized for feature extraction, compressing large patches to extract spectral–spatial features, which achieves dimensionality reduction. Index features, which represent the correlation between spectral bands, are fused with them as spectral–spatial-index features. Synthetic minority oversampling technique one-versus-all ensemble classifiers (SOVA) construct ensemble classifiers utilizing one-versus-all (OVA) strategy with synthetic minority oversampling technique (SMOTE) and the base classifiers of light gradient boosting machine (LightGBM) to address class imbalance problem. The experimental results on the five public HSI datasets show that TwoWin-SOVA effectively tackles the problems of high dimensionality caused by large patches and class imbalance, achieving classification performance as good as several state-of-the-art methods.