Building Segmentation of Hazy Aerial Images via Collaborative Decoupling Knowledge Learning

Building segmentation from aerial images is critical for applications such as urban planning and disaster management. However, haze severely degrades image quality, leading to radiance distortion, reduced contrast, and inaccurate segmentation. These degradations introduce two major challenges: (1) conventional methods typically treat dehazing and segmentation as separate tasks, which causes error accumulation and limits performance in haze-laden scenarios; (2) the diverse scales of buildings and haze-induced domain shifts hinder robust feature representation and accurate boundary delineation. To address these issues, we propose a collaborative decoupling knowledge learning (CDKL) framework that jointly performs dehazing and segmentation within a unified model. Specifically, a dual-branch decoupled network (D-Net) separates image restoration and segmentation, while a bidirectional feature co-interaction module (BFCIM) enables cross-task guidance: segmentation features refine dehazing for clearer structural details, and dehazing features enhance segmentation to resolve haze-induced ambiguities. Furthermore, a teacher network (T-Net) pre-trained on clear aerial images guides D-Net through a collaborative knowledge transfer (CKT) system, which aligns features across domains using multilevel strategies, thereby mitigating distribution shifts and improving robustness. Extensive experiments on synthetic and real hazy aerial datasets demonstrate that CDKL achieves state-of-the-art performance, delivering significant gains in segmentation accuracy and robustness across diverse haze conditions. The source code is available at https://github.com/Shan-rs/CDKL.git.