Accurate crop yield prediction is a key challenge for sustainable water and soil management in arid and semi-arid regions. This study estimated and classified irrigated wheat and barley yield in Parsabad, Ardabil province, Iran, during the 2014 growing season using Landsat-8 imagery and machine learning. Twenty-nine spectral, temporal, and topographic features were extracted from two phenological windows, tillering and anthesis. After removing yield outliers, Pearson correlation identified 24 significant features (p < 0.05) for wheat and 20 for barley. Three algorithms, Random Forest (RF), Gradient Boosted Trees (GBT), and Support Vector Machine (SVM), were trained using the full feature set and a reduced set selected via a 30% cumulative importance threshold, then compared. Random Forest achieved the best performance for both crops, with R² = 0.433 and RMSE = 862 kg/ha for wheat, and R² = 0.457 and RMSE = 869 kg/ha for barley. Reducing the feature space to six variables for wheat and five for barley had a minor effect on wheat accuracy but markedly improved the GBT model for barley. In binary yield classification, RF reached 74.7% overall accuracy (Kappa = 0.499) for wheat and 84.6% (Kappa = 0.690) for barley, outperforming GBT and SVM on all metrics. The observational yield gap was 30.3% for wheat and 40.6% for barley, while satellite-based estimates were 21.6% and 26.1%, respectively. Convergence of mean actual yield between the two approaches, with a difference under 6%, supports the validity of the remote sensing-based approach for crop yield estimation.