Towards scalable urban monitoring: Semi-supervised building change detection and the ICCD dataset for Indian cities

• A novel semi-supervised change detection framework (SemiSiamCD) is developed for monitoring urban transformation using remote sensing data. • Introduced AdaptiveMatch and Siamese-CASA modules to improve pseudo-label quality and spatial feature learning for building change detection. • Developed a new Indian Cities Change Detection (ICCD) dataset from multi-source Google Earth imagery for diverse urban environments. • Evaluated model performance on Levir-CD and ICCD datasets, demonstrating improved mIoU and cross-domain generalization for data-scarce regions. Monitoring urban transformation is essential for sustainable environmental management, disaster preparedness, and infrastructure planning. Though deep learning models struggle with limited labeled data and often exhibit domain shifts, their applicability remains limited across varied geographic regions. To address these limitations, we propose SemiSiamCD, a novel building change detection framework that comprises two proposed sub-parts, particularly AdaptiveMatch, a semi-supervised framework with uncertainty-aware class-wise adaptive thresholding mechanism for efficient high-quality pseudo-label generation as well as effective class imbalance mitigation; and Siamese-CASA, which is an attention-augmented Siamese U-Net architecture with a Sequential Atrous Spatial Pyramid Pooling module. We further present the Indian Cities Change Detection (ICCD) Dataset, a new benchmark derived from multi-source Google Earth Imagery of ten diverse Indian cities. It captures new building occurrences, structural modifications, and building removals, thus providing a comprehensive foundation for realistic building change detection applications. Extensive experimentation was done on both the Levir-CD dataset and the proposed ICCD dataset, where SemiSiamCD achieves a 2.14 % improvement in mIoU over the baseline FixMatch on the Levir-CD dataset and delivers up to a 4.61 % mIoU gain on ICCD, demonstrating strong cross-city generalization. The model further enhances building boundary delineation and significantly reduces false positives, making it particularly effective for deployment in data-scarce, rapidly urbanizing environments.