Multi-expert contrastive learning for remote sensing long-tailed image classification

Abstract

ABSTRACT Most existing remote sensing scene classification methods assume that the training data follows a balanced distribution. However, remote sensing data in the real world often exhibits imbalanced long-tailed distributions. The traditional deep learning model fits the head class well, but the tail class is not well fitted. In the face of complex background interference and the variability of similar land features, this manuscript proposes a multi-expert model based on contrastive learning and optimal transport to address these challenges. This method comprehensively considers the high inter-class similarity and large intra-class differences of long-tailed remote sensing data. By capturing key spectral discriminative features of land features in remote sensing scenes, the model’s generalization ability is enhanced to improve the performance of tail classes. Experiments with different balance ratios are conducted on three public datasets. They have verified that the proposed method achieves superior performance in remote sensing long-tailed scene classification tasks.