Abstract
Urban environments undergo continuous changes due to natural processes and human activities, which necessitates robust methods for monitoring changes in land cover and infrastructure for sustainable developments. Change detection in remote sensing plays a pivotal role in analyzing these temporal variations and supports various applications, including environmental monitoring. Many deep learning-based methods have been widely investigated for change detection in the literature. Most of them are typically regarded as per-pixel labeling and show their dominance, but they still struggle in complex scenarios with multi-scale features, imprecise & blurring boundaries, and domain shifts between temporal shifts. To address these challenges, we propose a novel Dual-Stream Hybrid Architecture (DSHA) that combines the strengths of ResNet34 and Modified Pyramid Vision Transformer (PVT-v2) for robust change detection for smart cities. The decoder integrates a boundary-aware module, along with multiscale attention for accurate object boundary detection. For the experiments, we incorporated the LEVIR-MCI dataset, and the results demonstrate the superior performance of our approach by achieving an mIoU of 92.28% and an F1 score of 92.50%. Ablation studies highlight the contribution of each component by showing significant improvements in the evaluation metrics. In comparison with existing methods, DSHA outperformed the existing state-of-the-art methods on the benchmark dataset. These advancements demonstrate our proposed approach's potential for accurate and reliable urban change detection, making it highly suitable for smart city monitoring applications focused on sustainable urban development.