High-resolution mapping of tropospheric ozone is essential for urban environmental assessment; however, satellite-derived ozone products are generally too coarse to capture neighborhood-scale variability in complex megacities such as Tehran. This study introduces an interpretable deep-learning framework that downscales coarse Sentinel-5P ozone observations to a 30-m spatial grid by integrating a U-Net convolutional architecture with SHapley Additive exPlanations (SHAP). A diverse suite of predictors—including land-surface indicators, meteorological parameters, terrain morphology, and chemical precursors—was harmonized and resampled to a unified spatial resolution. SHAP analysis was applied to quantify each predictor’s contribution, enabling the removal of redundant or low-impact variables before model training. Using spring 2020 as the evaluation period, the optimized U-Net successfully reconstructed fine-scale ozone gradients and reproduced Tehran’s characteristic north–south pattern driven by topography and emission density. Comparative analysis with preliminary outputs demonstrates that feature optimization enhances spatial coherence, reduces noise artifacts, and improves the representation of localized hotspots. Statistical evaluation further showed strong agreement between the downscaled ozone estimates and observational data at both station and district scales, demonstrating effective generalization across heterogeneous urban environments. Overall, the findings highlight the potential of combining deep learning with interpretability techniques to refine coarse satellite ozone observations and provide a scalable, high-resolution framework for urban air-quality monitoring and exposure assessment.