Integrating Deep Change Vector Analysis and SAM for Class-Specific Change Detection

Change detection is an important task in Earth observation with applications in environmental monitoring, urban development, and disaster management. Traditional supervised deep learning approaches rely on labeled bitemporal datasets, which are often scarce, making unsupervised methods such as Deep Change Vector Analysis (DCVA) a popular choice. However, unsupervised methods detect generic changes without distinguishing between different types, limiting their applicability in scenarios where class-specific change detection is required. Recently, foundation models like the Segment Anything Model (SAM) have demonstrated strong generalization capabilities, allowing for precise semantic segmentation with minimal supervision. Leveraging these advancements, we propose a novel framework that integrates DCVA for unsupervised change detection with SAM's ability to segment specific targets using only a few examples. This hybrid approach enables low-cost, class-specific change detection, reducing the need for extensive labeled datasets while improving the interpretability and relevance of detected changes. The proposed method holds significant potential for targeted monitoring applications in Earth observation. We demonstrate its capability for three classes related to buildings, trees, and landslide.