Accurate potato yield prediction is vital for optimizing agricultural management, food security, and economic planning. This study evaluates the potential of integrating Sentinel-2 (S2) spectral bands and vegetation indices (VIs) with machine learning models to enhance potato yield forecasting in Egypt. Fifty-two VIs were tested alongside spectral bands across different growth stages, comparing two predictive models: Multi-Linear Regression (MLR) and Random Forest Regression Model (RFRM). Results revealed that the RFRM outperformed the MLR, achieving the highest accuracy with an R2 of 0.734 and RMSE of 1.71 kg m-2 using the Potato Productivity Index (PPI) and B12 band as key features. In contrast, the best MLR model produced a lower R2 of 0.38. The integration of spectral bands and VIs, particularly from mid-growing stages (December and January), significantly improved yield prediction accuracy. Feature importance analysis showed that PPI and B12 were highly effective in capturing critical growth stages and assessing water stress, especially in semi-arid regions. The study produced detailed spatial yield maps that revealed considerable variability across the field, highlighting the importance of precision farming techniques in optimizing resource allocation and improving overall productivity. This research underscores the value of combining remote sensing data with machine learning to improve crop yield predictions, providing a practical approach for precision agriculture and sustainable farming practices in Egypt.