Semi-supervised change detection of buildings based on consistent regularization and VMamba

Building change detection holds significant importance in land use management and illegal constructions identification. Currently, fully supervised change detection approaches typically demands a large volume of well-annotated samples. A novel semi-supervised approach is developed to mitigate this limitation, integrating consistency regularization with the VMamba backbone to enhance detection performance, termed the Consistency Regularization and VMamba Network (CReVNet). This network generates predictions using a visual state space module, a content-guided feature fusion module, and an attention-based decoder block, and employs deep supervision to progressively refine building change features across multiple layers. Our semi-supervised update mechanism leverages a consistency regularization method to ensure consistent prediction results for unlabeled data before and after perturbation. Our CReVNet achieves state-of-the-art results on LEVIR-CD and WHU-CD datasets in terms of F1-score and IoU, which are the primary metrics for change detection. Particularly on LEVIR-CD, it achieves the highest F1 (89.29%) and IoU (80.66%) with merely 5% labeled data, demonstrating its effectiveness, even as some methods achieve competitive Precision or Recall scores. On WHU-CD, CReVNet ranks highest in F1/IoU at 20% and 30% label ratios. Demonstrating its robust performance, although it is slightly surpassed by some competitors at lower label ratios (5% and 10%).