FACANet: A Feature-Adaptive Complementary Alignment Network for Multisource Domain-Adaptive Remote Sensing Segmentation

Abstract

Existing feature-level multisource domain adaptation (MSDA) methods mainly utilize universal alignment. However, this habit ignores the different distribution shifts in each source domain to the target domain and inevitably leads to the suppression of a single source. Furthermore, a simple fusion operation fails to activate the semantic complementarity of different sources and destroys the originally aligned features, causing a new distribution shift to happen. To overcome these issues, a feature-adaptive complementary alignment network (FACANet) is proposed for remote sensing semantic segmentation. Specifically, the network explicitly calculates the difference in the source-to-target distribution, obtaining dynamic domain alignment. A cross-source attention mechanism is used to activate semantic complementarity. Simultaneously, dual-level consistency constraints are enforced to strictly balance this feature complementarity with domain alignment. Extensive experiments on two transfer tasks from the International Society for Photogrammetry and Remote Sensing (ISPRS) dataset demonstrate that FACANet consistently outperforms existing state-of-the-art multisource feature-level domain adaptation (DA) methods. The source code and models are available at https://github.com/JianyiZhong2001/FACANet#