Self-training guided disentangled adaptation for cross-domain remote sensing image semantic segmentation

Qi Zhao, Shuchang Lyu, Hongbo Zhao, Binghao Liu, Lijiang Chen, Guangliang Cheng

Abstract: Highlights•We propose a novel architecture (ST-DASegNet) for UDA RS semantic segmentation.•We propose an EMA-based cross-domain separated self-training paradigm on this task.•We provide an insight on exploiting feature disentangling to bridge the domain gap.•Feature-level adversarial learning and disentangling are compatible on this task.•ST-DASegNet boosts performance of UDA semantic segmentation on remote sensing data.