Class Imbalance in the Automatic Interpretation of Remote Sensing Images: A Review

Abstract

Class imbalance is a very challenging problem in data science, affecting the development of several application fields. This problem also plagues the automatic interpretation of remote sensing images. Especially in tasks such as classification mapping, object detection, change detection, and scene classification, the classes of training samples required by machine learning exhibit uneven distribution, which seriously affects the accuracy of model training. Our meta-analysis is based on 171 journal papers retrieved and screened from the Web of Science database, covering publication years, highly productive countries, highly cited authors, remote sensing data types, data augmentation methods, and the distribution of the main application fields. The solution to the proposed problem involves three aspects: model innovation and optimization, loss function improvement, and data augmentation. Experiments on benchmark datasets have demonstrated the effectiveness of these methods. In terms of remote sensing task applications, we provide a comprehensive review and analysis of recent research cases on deep learning aimed at addressing the class imbalance problem. Finally, we discuss the synergistic relationship between models, loss functions, and data augmentation, summarize the current challenges in this field, as well as propose several ideas for addressing the class imbalance problem.