This study focused on paddy fields with different rice varieties and nitrogen application levels. Canopy hyperspectral data were collected during the tillering, jointing-booting, heading-flowering, and milk ripening stages using an unmanned aerial vehicle (UAV)-borne hyperspectral imaging system. A total of 18 spectral parameters, including 10 vegetation indices and 8 differential indices, were extracted and subjected to correlation analysis with rice yield. The eight spectral parameters with the highest correlations were selected as input variables for model construction. Subsequently, two machine learning algorithms, partial least squares regression (PLSR) and random forest (RF), were employed to construct rice yield prediction models for different growth stages, respectively. Results indicated significant differences in rice yield estimation performance across growth stages, ranked as follows: jointing-booting stage>heading-flowering stage>milk ripening stage>tillering stage. Among these, the jointing-booting and heading-flowering stages were identified as the optimal growth stages for yield estimation. At the jointing-booting stage, the PLSR model outperformed the RF model in prediction accuracy; conversely, at the heading-flowering stage, the RF model exhibited superior prediction accuracy compared to the PLSR model. Overall, the PLSR model performed best at the jointing-booting stage, achieving a coefficient of determination (R²) of 0.856 and a root mean square error (RMSE) of 0.970 t/hm² on the validation set.