Unsupervised Domain Adaptation Remote Sensing Image Classification Based on Class Centroid Alignment

Abstract

Deep learning algorithms based on data-driven approaches have shown remarkable success in remote sensing image classification. Nevertheless, these methods often assume that the training data and test data adhere to the same distribution, which is not always the case in real-world scenarios. To address the constraints of unsupervised learning in land classification for remote sensing images and facilitate the automated annotation of land at a large scale across different sensors, this paper proposes a novel technique known as centroid alignment-based conditional domain confusion. This method aligns the class feature centroids of the source and target domains, leveraging category information in the conditional domain discriminator to enhance discriminative capabilities and achieve cross-domain distribution alignment at the class level. The proposed method is evaluated on the MRSSC2.0 dataset by conducting comprehensive experimental analysis. The results substantiate the efficacy of the proposed method in mitigating the domain shift in data distribution and enhancing the classification accuracy of cross-sensor high-resolution remote sensing images.