GMMDA: Gaussian Mixture Model-based Generative Domain Adaptive Semantic Segmentation for Remote Sensing Images

Unsupervised Domain Adaptation (UDA) for semantic segmentation transfers knowledge from an annotated source domain to an unlabeled target domain, reducing the cost of pixel-level annotation. Most existing methods rely on dense discriminative classifiers that estimate pixel-wise class posterior probabilities using a single weight vector per class. However, in remote sensing imagery, intra-class variability often gives rise to multi-peaked feature distributions with multiple sub-clusters, which may not be adequately captured by a single weight vector, leading to intra-class confusion. Moreover, such discriminative classifiers primarily focus on learning class decision boundaries, while the consistency of same-class sub-clusters across domains is only weakly constrained. Under domain shift, these sub-clusters may drift inconsistently, thereby aggravating cross-domain structural mismatch. To this end, we propose GMMDA, a Gaussian Mixture Model (GMM)-based generative domain adaptation method that predicts pixel labels through Bayesian inference based on class-conditional feature likelihoods. Specifically, we fit a class-wise GMM using an online momentum Sinkhorn EM procedure to model class-conditional feature densities, where each Gaussian component corresponds to an intra-class sub-cluster, enabling a richer characterization of multi-peaked distributions. Furthermore, component-aware contrastive learning treats Gaussian components as fine-grained semantic references, pulling target-domain features toward their best-matched components using a variance-aware distance that accounts for component scales. In addition, we introduce component-wise distribution alignment to align each source component with the corresponding source-guided target local statistics by jointly constraining their means and principal directions, aiming to reduce cross-domain structural mismatch. Experimental results show that GMMDA achieves competitive mIoU scores on the ISPRS and LoveDA benchmarks compared with existing methods.