Imbalanced sample feature enhancement of hyperspectral imagery classification

Due to its easy application, low computation consumption and promising generalization performance, extreme learning machine is widely used for hyperspectral imagery classification. However, most extreme learning machine algorithms ignored the depression of the majorities to minorities. To tackle this task, a model is proposed to improve the performance of imbalanced sample classification for hyperspectral images. Firstly, the spatial and spectral features are combined, in order to enhance the feature of minorities, guided filtering and enhanced neighborhood features are adopted, which will enlarge the samples of minorities and provide diversity for classification task. Secondly, the separating boundary is supposed to be pushed toward the side of minority class under imbalanced situation, which in fact favors the performance of majority class. A random under-sampling bagging extreme learning machine is employed, which will intensively improve the depression effect. Experiments carried on two widely used datasets. The results showed that, the imbalanced sample feature enhancement model proposed in this manuscript takes into account both the feature of enhanced small sample and the suppression of large sample on the classification boundary of small sample by extreme learning machine, which weakens the suppression effect of large sample on small sample, and further improves the classification accuracy of hyperspectral images.