A novel framework and benchmark dataset for fine-grained urban building change detection in ultra-high-resolution UAV imagery

The accelerating pace of urbanization necessitates precise and timely monitoring of building changes, which is crucial for urban planning, resource management, and disaster response. Ultra-High-Resolution (UHR) Unmanned Aerial Vehicle (UAV) imagery offers a potent data source for fine-grained urban Building Change Detection (BCD). However, complex urban backgrounds in UHR imagery pose significant challenges: variable illumination conditions and occlusions frequently induce spurious changes, while textural similarities between building rooftops and other urban elements often lead to widespread misclassification. To address these challenges, this paper proposes a novel end-to-end framework, the Difference Interaction and Progressive Fusion Network (DIPFNet), specifically tailored for UHR UAV-BCD. Within the decoding phase, we introduce a Multi-dimensional Difference Feature Interaction Module (MDF). By synergizing subtraction, summation, and concatenation branches, MDF explicitly decouples change evidence – reinforcing structural integrity via summation and preserving semantic context via concatenation – to robustly capture subtle changes while suppressing background noise. Concurrently, a Progressive Hierarchical Feature Fusion Module (PHF) is developed. Employing an incremental top-down fusion strategy, PHF effectively leverages hierarchical semantic cues to progressively restore spatial details, generating highly accurate change maps with sharp boundaries. To alleviate data scarcity, we constructed Guangzhou UAV Building Change Detection (GZ-UAVBCD), a new benchmark dataset comprising 3536 UHR image pairs that capture intricate urban dynamics such as roof renovations and temporary structures. Extensive comparative experiments demonstrate that DIPFNet significantly surpasses 17 state-of-the-art (SOTA) methods. Specifically, compared to the next-best approach, DIPFNet achieves substantial improvements across key comprehensive metrics: F1, IoU 1, and Kappa are enhanced by 2.19%, 2.80%, and 2.33%, respectively, reaching a new SOTA F1-score of 75.88%. Moreover, DIPFNet achieves this superior accuracy with a rapid inference speed of 65 FPS, offering an optimal accuracy-efficiency trade-off for practical UAV applications.