Abstract: Assimilating crop growth models with multi-source unmanned aerial vehicle (UAV) remote sensing observations provides an effective approach for improving the crop growth monitoring scope and yield estimation accuracy, thereby supporting precision regulation of water and fertilizer management. However, state-variable updating and parameter optimization are often implemented separately, which may cause inconsistency between the corrected crop state and the parameters that control subsequent model process. To further enhance the accuracy and stability of rice growth simulation, this study employed the Sobol method to screen yield-sensitive crop parameters within the WOFOST (world food studies) model—implemented in the PCSE (python crop simulation environment). Leaf lifespan at 35℃ (SPAN), maximum relative leaf area growth rate (RGRLAI), conversion efficiency of assimilates to storage organs (CVO), leaf maintenance respiration rate (RML), and dry matter weight at transplanting (TDWI) were selected for optimization. A collaborative assimilation framework (ensemble Kalman filtering with parameter optimization by whale optimization algorithm, EnKF-WOA) was further developed for the simultaneous updating of state variables and crop parameters by coupling ensemble Kalman filtering (EnKF) with the whale optimization algorithm (WOA). Based on UAV remote sensing retrievals of leaf area index (LAI), total above ground production (TAGP), and crop transpiration (TRA), a multi-variable and multi-window assimilation strategy was designed, including three univariate, three bivariate, and one multivariate strategy, were evaluated under four sequential assimilation-window settings (corresponding to the tillering stage, jointing-booting stage, heading-flowering stage and milking stage). The results showed that the WOFOST model after parameter localization could generally reproduce the seasonal process of LAI, TAGP and TRA, but still underestimated final yield with a relative error of -10.4%. The developed EnKF-WOA assimilation framework was able to update model state variables and corresponding crop parameters according to the posterior analysis values. Compared with the single EnKF and WOA algorithms, the coupled framework achieved better performance in both simulating rice growth dynamics and estimating final yield. In terms of variable selection, LAI was identified as the most effective variable for univariate assimilation. Assimilating LAI only at the tillering stage could reduce the relative error of yield estimation to -3.3%, indicating that LAI has strong potential for early yield prediction. For bivariate assimilation, the combination of LAI and TRA showed the most stable performance. The relative errors under all assimilation windows were 0.4%～5.2%, and the coefficient of variation decreased markedly from 15.3% to 9.1% as the number of assimilation windows increased. For multivariate assimilation, the combined assimilation of LAI, TAGP, and TRA achieved the best overall accuracy when four assimilation windows were used, with the relative error of 0.5%, and the coefficient of variation of 4.3%, demonstrating high estimation accuracy and lower uncertainty. Considering the number of assimilation windows required to achieve high accuracy, univariate LAI assimilation and bivariate LAI+TRA assimilation can provide early yield estimation at tillering stage, which is suitable for guiding timely adjustments to field management. In contrast, multivariate assimilation requires observations up to the jointing–booting stage to obtain more accurate yield estimation and lower uncertainty, which could be more suitable for scenarios with strict requirements for uncertainty. This study provides algorithm support and variables selection criteria for assimilating remote sensing data with crop growth models, holding significant practical value for monitoring crop growth and estimating yields.