Managing rapid urbanization requires reliable tools to forecast how metropolitan areas expand, especially in semi-arid regions where seasonal vegetation changes and reflective barren soils often confuse traditional models. We address this by developing a hybrid simulation framework that couples deep learning with spatial modeling to project urban growth in the Mashhad metropolitan area, Iran, up to 2030. Utilizing seasonal Sentinel-2 time series alongside Google's Dynamic World land-cover dataset, we trained a U-shaped Spatio-Temporal Attention (U-TAE) network to capture multi-seasonal spectral features. This step generated high-fidelity built-up probability maps, where a decision threshold of 0.20 was chosen via sensitivity analysis to optimize spatial accuracy. These maps then served as the transition potential layer within a Cellular Automata (CA) model. During training, the U-TAE model converged smoothly, reaching a validation Intersection over Union (IoU) of 0.9836 at epoch 11. Crucially, the spatiotemporal attention mechanism successfully distinguished highly reflective peripheral soils from permanent structures, eliminating seasonal false positives that typically impair static baselines. The integrated U-TAE-CA model effectively simulated low-density sprawl and edge expansion along dynamic corridors. These results demonstrate that combining temporal deep learning attention with cellular simulation provides a robust, realistic approach to monitoring dynamic land-use patterns, offering planners actionable predictive insights for sustainable infrastructure development.