Deep Spectral-Spatial Ensemble Learning for Imbalanced Small-Sample Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification under imbalanced and small-sample conditions is often hindered by the neglect of spatial contextual dependencies and the high-dimensional spectral redundancy. Moreover, conventional deep learning methods are prone to majority-class bias under imbalanced data distributions, which degrades performance on the minority class. To alleviate these issues, this letter proposes a novel spectral-spatial-based ensemble-adapted SMOTE with focal loss (SS-EASF) for HSI classification. First, spectral and contextual spatial information is captured through a spectral-spatial weighted neighborhood within a fixed $13\times 13$ window using adaptive weights. Second, to obtain reliable nearest neighbor samples, a novel weighted distance space is introduced through the integration of eXtreme gradient boosting (XGBoost) feature importance, Fisher score, and Pearson correlation coefficient. Finally, an improved focal loss function with adaptive parameters is applied to base classifiers to prioritize hard-to-classify samples, thereby enhancing overall classification accuracy. Experiments conducted on three HSI datasets demonstrate that SS-EASF provides competitive results compared to the state-of-the-art method in imbalanced small-sample scenarios.