Accurate in-season yield prediction is critical for guiding nitrogen management in rice production, particularly under the increasingly variable weather conditions in cold regions. However, the prediction accuracy of conventional crop models is often limited by uncertainties in model inputs and the inability to capture real-time crop status. This study proposed a yield prediction framework by coupling the CERES-Rice model with re-mote sensing observations and machine learning techniques. Field experiments were conducted in Northeast China from 2014 to 2018. Spectral data collected by the Rap-idSCAN active canopy sensor were used to retrieve key agronomic parameters, including aboveground biomass (AGB), leaf area index (LAI), plant nitrogen concentration (PNC) and plant nitrogen uptake (PNU), using a Random Forest algorithm. These variables were then assimilated into the CERES-Rice model to calibrate its sensitive parameters and im-prove yield predictions. The analysis showed that updating agronomic variables using remote sensing resulted in acceptable modeling accuracy (R² > 0.7). Using PNU and AGB outperformed LAI in enhancing model performance, while using PNC was unreliable. Without assimilation, CERES-Rice gave moderate performance (R² = 0.58, RMSE = 1.3 t ha⁻¹), while coupling with PNU improved prediction accuracy(R² = 0.67, RMSE = 0.79 t ha⁻¹). This study highlights the potential of integrating active canopy sensing and ma-chine learning with crop models for dynamic yield forecasting and provides a valuable reference for optimizing nitrogen management in rice systems.