Satellite-based change detection plays a critical role in monitoring land-use dynamics, especially in rapidly developing urban areas. This study develops an advanced deep learning framework for analyzing land cover transitions in Mashhad and Maragheh, Iran, using multi-temporal Sentinel-2 Level-2A imagery (2019-2023) at 10m spatial resolution. We propose an ensemble approach combining multiple U-Net++ architectures to classify six key transition categories, focusing on urban expansion patterns such as the conversion of agricultural land and wastelands to built-up areas. The methodology incorporates a comprehensive processing chain including image tiling (112×112 pixel patches), multi-model inference, and post-classification refinement using seasonal NDVI analysis and cloud-shadow masking derived from Sentinel-2 probability layers. Ground truth data were meticulously prepared through visual interpretation in QGIS supplemented by Google Earth verification, ensuring accurate reference labels for model training and validation. Quantitative evaluation yielded precision (0.60), recall (0.72), and F1-score (0.65) metrics, demonstrating effective detection of major land-use changes while revealing challenges in distinguishing spectrally similar classes and precise boundary delineation. The framework's operational capability was further validated through successful application across different urban landscapes and temporal scenarios. This research contributes an automated, scalable solution for urban change monitoring that addresses practical challenges in heterogeneous environments. The integration of ensemble deep learning with multi-temporal spectral analysis advances current change detection methodologies, offering valuable tools for urban planners and environmental managers. Future work will focus on enhancing spectral discrimination capabilities and incorporating multi-sensor data fusion to improve detection accuracy for complex transition patterns.