Meta Feature Disentanglement under continuous-valued domain modeling for generalizable remote sensing image segmentation on unseen domains

Abstract

As a long-standing challenge, the generalization ability of segmentation models has invoked enormous research on domain-agnostic learning, but current methods tend to be invalid in remote sensing. On the one hand, their common assumption that domains can be represented as discrete labels holds with difficulty in remote sensing, where domain shifts arise from dynamic and continuous changes. On the other hand, they struggle to perform well on unseen domains in remote sensing image segmentation tasks, where gaining diversity-sufficient and semantic-effective training distributions remains a significant challenge. To address these obstacles, this paper develops a novel domain generalization (DG) method, termed Meta Feature Disentanglement (MetaFD), for remote sensing image segmentation. To circumvent the inherent issue of discrete-valued domain modeling, MetaFD outlines domains in remote sensing with the continuous-valued space modeled by a variational autoencoder (VAE) and performs domain-label-free feature disentanglement aided by vector decomposition and semantic guidance. And to enhance generalization on unseen domains, MetaFD expands training distributions under the meta-learning framework by using the VAE to directionally randomize domain-specific variations, which can generate novel domains with vast diversity but no severe semantic distortions, and employs the generated data to maintain disentanglement consistency and design more realistic meta-episodes. Multiple public datasets are organized to construct three DG benchmark datasets for experimental studies. Extensive experimental results demonstrate that MetaFD significantly outperforms other state-of-the-art methods in remote sensing image segmentation tasks. The code is available at https://github.com/LCB1970/MetaFD.