Mixture Distribution Modeling Pseudo-Label Propagation of Domain Adaptation for Low-Data Resource Cross-Scene Classification

The low-data resource cross-scene classification (LDR-CSC) aims to transfer learnable knowledge from a limited labeled source domain to an unlabeled target domain. However, different imaging conditions, diverse spatial layouts, and complex land covers can lead to significant domain discrepancies between the source and target domains. Moreover, the low-data resource condition of the source domain would further aggravate domain discrepancies that hinder knowledge transfer, resulting in inferior cross-scene classification performance. Therefore, in this article, a novel mixture distribution modeling pseudo-label propagation of domain adaptation called MDP2-DA is proposed for LDR-CSC. First, to improve learnable knowledge transferability under low-data resource conditions, a domain-invariant and domain-specific collaborative learning is designed to jointly capture interdomain spatial cues of consistency from dual-domain complicated remote sensing scenes and in-domain-specific local regions. It can bridge domain discrepancies and ensure a more complete description of unlabeled data for high-quality pseudo-label generation. Second, to improve the quality of cross-domain pseudo-label propagation, a mixture distribution modeling strategy is designed for cross-domain optimization learning (CDOL) of pseudo-label propagation. This strategy can build reliable and unreliable sets using a Gaussian-uniform model and form a self-calibration mechanism to mine the effective information latent in unreliable sets, facilitating cross-scene classification ability in the unlabeled target domain. Finally, extensive experiments are conducted on various LDR-CSC scenarios. The results demonstrate that the proposed MDP2-DA achieves state-of-the-art performance for LDR-CSC.