Remote sensing-based semantic change detection plays a crucial role in various fields including urban development, disaster monitoring, and environmental change analysis, but faces the limitation of high dataset construction costs. To address this, Ch...
Remote sensing-based semantic change detection plays a crucial role in various fields including urban development, disaster monitoring, and environmental change analysis, but faces the limitation of high dataset construction costs. To address this, ChangeDiff, a diffusion model-based generative augmentation model, was proposed; however, the quality and realism of generated data remained limited. This study proposes Advanced ChangeDiff (ACDiff), which adopts ChangeDiff's diffusion-based Text-to-Layout and Layout-to-Image generation mechanisms while improving output data quality through post-processing algorithms. ACDiff applies instance-aware spatial control to the Text-to-Layout output, enabling change generation based on semantic instances and systematic change patterns reflecting real-world land use transitions. In the Layout-to-Image stage, a change-mask-based selective noise sharing strategy maintains consistency in non-change regions while ensuring diversity in change regions. Additionally, a composite score-based quality selection strategy ensures high-quality dataset construction. Experiments on the SECOND dataset demonstrated that models trained with ACDiff augmentation achieved the highest detection performance in both augmentation method comparison and data scale-dependent augmentation effect experiments. This study presents a practical methodology for automatic construction of high-quality semantic change detection datasets and is expected to lower the entry barrier for change detection research where labeling costs are high.