St-CGN: Multiscale Consistency and Dual-View Entropy Fusion for Source-Free Domain Adaptation in Remote Sensing Semantic Segmentation

Abstract

In unsupervised domain adaptation (UDA) for remote sensing (RS), source-free domain adaptation (SFDA) has emerged as a crucial paradigm to overcome strict data limitations. However, traditional self-training (ST) methods struggle with severe domain shifts in SFDA, which cause spatial and semantic inconsistencies that lead to early misclassifications into severe pseudolabel noise. To resolve these problems, we propose St-CGN, a novel two-stage framework designed to achieve robust class-level adaptation by establishing strict multiscale feature invariance and utilizing class-balanced recalibration to ensure pixel-level accuracy in the UDA mission of RS. First, to achieve better feature robustness and improve the model’s ability to capture features with semantic correlations, a feature-consistency learning (CL) strategy is proposed, which sculpts a semantic-invariant feature space for the network, ensuring that varied visual representations of the same semantic content are mapped in a consistent manner. Second, a prototype feature guidance ST procedure based on a pseudolabel refining strategy is proposed to ease the uncertainty predictions, leveraging the semantic feature spaces to mitigate the class bias caused by domain shift. Extensive experiments conducted on widely used datasets verify the effectiveness of the proposed method for RS applications.