Classification of imbalanced hyperspectral images using ensembled kernel rotational forest

Hyperspectral image classification suffers from an imbalance in the samples belonging to its different classes. In this paper, we propose a two-fold novel approach named oversampler + kernel rotation forest (O + KRoF). First, Synthetic minority oversampling (SMOTE) and adaptive synthetic oversampling (ADASYN) techniques are employed on original data to balance it due to their adaptive nature in the majority and minority samples. Finally, the ensembled KRoF classifier is applied, a combination of unpruned classification and regression trees (CART) as its base algorithm and kernel PCA for feature reduction and most significant nonlinear spatial-spectral feature selection. Furthermore, we designed a comparison study with frequently used oversamplers and related state-of-art tree-based classifiers. However, it is found that our ensemble model is suitable and performs better as compared to earlier works as it attains 90.92%, 97.1%, and 93.39% overall accuracies when experimented on the benchmark datasets, Indian Pines, Salinas Valley, and Pavia University, respectively.