PCDA: Pseudo-Label Correction for Source-Free Universal Domain Adaptation in Remote Sensing Scene Classification

Abstract

The precise classification of remote sensing images is crucial for applications such as disaster monitoring, land use planning, and national security. Since effective classification models often rely on large, well-annotated datasets, the remote sensing field faces two persistent challenges: the limited availability of labeled data and the domain shifts that hinder model generalization. Existing approaches primarily include data generation-based domain adaptation (DA) schemes and large model-based zero-shot learning methods. The former synthesizes target domain samples via generative adversarial networks (GANs) to address the source-free DA (SF-DA) problem, but consequently introduces significant noise interference, and the latter excessively relies on the generalization capability of large models; this capability is constrained by the inherent long-tailed distribution characteristic in remote sensing scene classification tasks, failing to be fully exploited. To overcome these limitations, we propose pseudo-label correction for source-free universal DA (PCDA), a novel framework that addresses both data privacy and domain/category shifts more effectively without source data. We propose generating class-strategy-adaptive pseudo-prototypes instead of dataset generation to reduce training costs while introducing source domain knowledge. And design an implicit pseudo-label correction mechanism that dynamically optimizes pseudo-labels, effectively reducing the interference of pseudo-label uncertainty in model training, thereby enhancing model robustness. Experimental results demonstrate that the proposed method achieves remarkable average classification accuracies of 88.1%, 82.7%, 71.7%, and 75.6% in the R-U, R-A, R-N, and A-N dataset pairs, respectively. The code is available at https://github.com/TJU-L-Ning/PCDA