Imbalanced Data Classification Using Transfer Learning and Oversampling

A common challenge when applying transfer learning to specific domains is imbalanced datasets, where some classes have significantly fewer samples than others. This can skew the model's predictions toward the majority class. In this paper, we aim to investigate the effectiveness of oversampling methods including random oversampling technique and image augmentation in hyperspectral image classification, exploring their potential to enhance performance.This paper proposes an oversampling method which are applied on the imbalanced datasets by randomly duplicating samples from the minority classes until the number of samples matches that of the majority classes. In addition, image augmentation techniques (rotation, flipping, zooming) are conducted to increase the diversity of the minority class. Moreover, the training process of the proposed ResNet-18 transfer learning network is fine-tuned with optimized parameters to address challenges like overfitting and underfitting. Furthermore, the proposed transfer learning model's adaptability and robustness are evaluated and compared with the other two state-of-the-art transfer learning models, GoogLeNet and AlexNet. The experimental results show that the GoogLeNet model achieves impressively higher classification accuracy than the other state-of-the-art models for imbalanced and balanced Flowers dataset (93.61% and 93.89%, respectively), while the proposed ResNet-18 model obtains the highest accuracy for imbalanced and balanced Food dataset (87.63% and 89.69%, respectively). In addition, all transfer learning models are able to obtain higher accuracy on the balanced data than on the imbalanced data.