DBCD: Dual Branch Contrastive Diffusion Network for Unsupervised Building Extraction

Read full text Summarise with VIRAT
Authors: Chengming Zhang, Shengkai Gao, Xin Lyu, Yiwei Fang, Jue Zhang, Fan Liu, Hongmin Gao, Xin Li
Cited by: 0
References: 83

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Generative artificial intelligence has reshaped perspective on data representation of remote sensing images (RSIs), offering powerful solutions for regions lacking annotated data. While these generative paradigms excel in learning complex data patterns, their direct application to unsupervised downstream tasks, such as building extraction, remains hindered by spectral ambiguity and boundary degradation. To address these limitations, we propose a novel Dual-Branch Contrastive Diffusion Network (DBCD). This physics-informed generative framework is built on the physical properties of RSIs, and achieves robust annotation-free building segmentation. Based on the insight that stable segmentation requires both reliable semantic anchors and geometric constraints, our framework integrates two synergistic branches. Specifically, a conditional branch leverages similarity-guided superpixel priors to extract robust semantic cues, resolving spectral confusion through feature disentanglement. Simultaneously, a Diffusion Branch performs progressive reconstruction guided by an edge-aware contrastive mechanism, which uses image-derived edges to emphasize boundary-aligned features during denoising. Furthermore, we introduce a dynamically coordinated curriculum learning strategy that adaptively shifts optimization focus from coarse semantic alignment to fine-grained detail refinement, ensuring training stability. Experimental results on the ISPRS Vaihingen and Potsdam datasets demonstrate that DBCD achieves state-of-the-art performance. Specifically, the proposed method yields 48.1% mIoU, 62.3% F1-score, and 34.7% BoundF on Vaihingen, and 46.2% mIoU, 61.3% F1-score, and 36.0% BoundF on Potsdam, proving the effectiveness of guided generative modeling in handling label-scarce remote sensing tasks. The code is available athttps://github.com/KlausZcm/DBCD.

Fields of Study & Subjects

Automated Road and Building Extraction Remote-Sensing Image Classification Advanced Neural Network Applications

Publication details

Journal / source
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Volume
19
Pages
13171–13190
Publication date
1 January 2026
Language
EN