A Spectral–Spatial Feature Rotation-Based Ensemble Method for Imbalanced Hyperspectral Image Classification

Hyperspectral image classification has two key characteristics: spatial dependency and class imbalance. The synthetic oversampling methods for solving the between-class imbalance have the problem of spatial information loss to some extent. However, the feature space optimization methods for solving data within-class complex distribution problems have not been studied extensively. To remedy such an issue, we propose in this article an improved spectral–spatial feature rotation-based support vector machine (SVM) ensemble method (SSFRoS). More precisely, in the data preprocessing stage and from the perspective of sample space optimization, we devise a spectral–spatial neighborhood-based adaptive synthetic sampling algorithm (SSAS) that focuses adaptively on hard-to-learn samples, with artificial samples interpolated in the spectral–spatial neighborhood to further utilize the spatial information. In the feature rotation stage and from the perspective of feature space optimization, we innovatively discover that the class overlapping problem can be handled by combining feature extraction, to achieve the goal of reducing the difficulty of imbalanced classification. Besides, the SVM and voting procedure are used in the ensembled classification stage. We then conduct a series of experiments to compare the proposed SSFRoS with several baselines in terms of seven metrics. The results show that SSFRoS has significant advantages in dealing with imbalanced hyperspectral image classification problems. In addition, a detailed ablation study shows that the lowest class recall can be improved by 46.67% and 5.26% at Indian Pines and Pavia University, respectively, which demonstrates that using SSAS can enhance the recognition of hard-to-learn classes.