UrbanGenNet: Domain-Generalized Remote Sensing Semantic Segmentation via Decoupled Style Perturbation and Class-Wise Invariant Learning

Abstract

Coupled radiometric and geographic shifts challenge domain-generalized semantic segmentation for optical remote sensing. The former arises from spatially nonstationary imaging conditions, such as shadows, illumination gradients, and sensor-dependent appearance changes, while the latter is reflected in city-dependent scene layouts and context priors. Existing domain generalization (DG) methods can improve appearance robustness through style/statistics diversification, but they do not explicitly model the spatially heterogeneous radiometric artifacts common in remote sensing imagery. Meanwhile, existing invariance-oriented methods do not directly enforce geographically stable semantic prediction in dense scenes, where layout and contextual priors may vary substantially across regions. To address these two limitations jointly, we propose UrbanGenNet, a two-stream remote sensing-specific framework with a shared backbone. An augmented stream applies spatially aware style perturbation (SASP) to emulate realistic spatially varying radiometric disturbances. A clean stream preserves unperturbed features for optimizing an urban loss, which performs class-wise invariant regularization across CLIP-guided latent environments to encourage geographically stable semantic prediction. By decoupling radiometric diversification from environment-based structural invariance learning, UrbanGenNet enables the backbone to acquire both robustness to nonstationary radiometric variation and invariance to geographic/contextual shift without mutual interference. Experimental results show that UrbanGenNet outperforms recent representative methods on WHU-Mix and remains highly competitive on FLAIR #1, showing complementary advantages in small-object completeness and robustness under shadow-like perturbations.