遥感图像语义分割在土地覆盖与利用分类、城市规划以及变化检测等领域具有重要作用。作为一种极具潜力的无监督学习方法,域自适应技术大大推动了遥感图像语义分割的发展。然而,现有模型尚基于单任务来学习,学习到的分割特征不够充分,导致在分割过程中难以准确识别遥感图像中的复杂区域。为了解决这一问题,本文提出了一种多任务学习域自适应语义分割网络(MTLDANet),该网络通过协同学习遥感图像中的语义信息与高程信息,来提升分割特征的学习能力。具体来说,该方法将任务特定的语义特征和高程特征输入跨任务特征关联学习模块,挖掘任务之间的潜在相关性,从而获得更强的任务特定特征表达。通过伪标签引导的混合一致性学习模块提升伪标签质量,实现全局域对齐。此外,熵引导的类别级对齐进一步增强了难分类类别的区分性。本文基于ISPRS 2D和US3D数据集进行了四组跨场景遥感图像语义分割实验。结果表明,所提方法在多种复杂跨域场景下均显著优于现有的域自适应方法。

Semantic segmentation of remote sensing images (RSIs) plays a crucial role in land cover and land use classification, urban planning, and change detection. As a highly promising unsupervised learning method, domain adaptation has significantly accelerated the advancement of RSI segmentation. However, current models often rely on the limited feature learning capability of single-task approaches, making it difficult to accurately distinguish hard-to-classify regions in RSIs. To address this issue, a multi-task learning domain adaptive network (MTLDANet) is proposed, which jointly learns semantic and elevation information in RSIs, improving segmentation performance. The method feeds task-specific semantic and elevation features into a cross-task feature correlation learning module to explore latent correlations between tasks, thereby enhancing task-specific feature representations. A hybrid consistency learning module, guided by pseudo-labels, is employed to improve pseudo-label quality and achieve global domain alignment. Additionally, entropy-guided category-level alignment enhances the separability of challenging categories. The proposed method is evaluated on four cross-scene RSI segmentation experiments using the ISPRS 2D and US3D datasets. Results show that the method outperforms existing domain adaptation approaches, demonstrating significant advantages in various complex cross-domain scenarios.