RB-SCD: A New Benchmark for Semantic Change Detection of Roads and Bridges in Traffic Scenes

Abstract: With the rapid modernization of urban transportation, accurately detecting changes such as road and bridge construction, renovation, and demolition is crucial for urban planning and traffic management. However, existing methods often struggle to extract fine-grained semantic changes in complex traffic scenes, largely due to the lack of high-quality annotated change detection (CD) datasets. To address this, we introduce the Road and Bridge Semantic Change Detection (RB-SCD) dataset, a comprehensive benchmark consisting of 260 pairs of high-resolution remote sensing images. RB-SCD spans diverse geographic areas and includes a wide variety of road and bridge types across over ten cities in multiple countries. It covers 11 distinct categories of semantic changes, enabling detailed structural and functional analysis. Based on this challenging dataset, we propose a novel framework called the Multimodal Frequency-Driven Change Detector (MFDCD). For the first time, MFDCD integrates multimodal feature characteristics in the frequency domain. It comprises two key components: the Dynamic Frequency Coupler (DFC) and the Textual Frequency Filter (TFF). DFC couples hierarchical visual features with wavelet-based frequency components, enhancing the perception of fine-grained and cross-temporal structural changes. TFF transforms textual features extracted by the CLIP model into the frequency domain via Fourier transform and applies graph-based filtering to extract salient frequency responses. These are then fused with visual features to enable effective multimodal representation learning. Extensive experiments show that MFDCD achieves strong performance on RB-SCD and three public benchmarks. The RB-SCD dataset, with its rich and diverse annotations, serves as a valuable resource for advancing research in road and bridge change detection under complex traffic conditions.