Abstract: Leaf Nitrogen Content (LNC) is an important indicator to evaluate the quality and yield of fruits, where nitrogen is an essential nutrient element for the growth of citrus. Satellite remote sensing has been widely used to rapidly and nondestructively capture nitrogen content data for the cultivation and production of citrus in recent years. In this study, a two-layer stacking ensemble learning framework was constructed using Landsat8 OLI satellite remote sensing images and ground sample data, thereby accurately estimating the nitrogen content of citrus leaves in critical growth periods. K-Nearest Neighbor (KNN), Random Forest (RF), and Adoptive boosting (Adaboost) were utilized as base models, whereas, Linear Regression (LR) was employed as the meta-model. The LNC values were sorted from high to low and then divided into 6 groups at equal intervals. A systematic analysis was also made to compare the spectral characteristics under different LNC. There were significant differences in the spectral reflectance in the visible light range (400-760 nm) and near-infrared band (760-1 250 nm), due mainly to the absorption of chlorophyll and the multiple reflections of the canopy. The trees with higher LNC commonly presented lower spectral reflectance. The correlation coefficient between vegetation indices (VIs) and LNC was calculated to optimize the spectral features. Grid search and 5-fold cross validation were utilized to train the model, where the LNC distribution map was generated for the study area. The results showed that the Stacking presented the best performance in the testing dataset, with coefficient of determination (R2) of 0.761, Mean Absolute Error (MAE) of 1.046 g/kg, Root Mean Squared Error (RMSE) of 1.366 g/kg and Mean Absolute Percent Error (MAPE) of 3.494%.. Compared with Adaboost, the best performance was achieved using individual models, where the R2 increased by 0.025, whereas RMSE, MAE, and MAPE decreased by 0.07, 0.109 g/kg and 0.325 percentage point, respectively. It revealed that the Stacking was fully integrated into the base models for a higher estimation accuracy. However, there was an obviously underestimated phenomenon in the measured value from each model, particularly in the LNC estimated values of >32 g/kg. Meanwhile, by comparing the Akaike Information Criterion (AIC) of each model, the AIC value of Stacking was significantly lower than other individual model, indicating that Stacking was the best LNC estimation model in the observation period of this study. In addition, soil background and model performance were discussed. The spectral information was interfered by soil background. Based on the concept of soil line, many researchers proposed some VIs to reduce the influence, which were also adopted. In this study, only spectral features were used to build the model, which limited the capability of the model. It would be considered to increase the observation periods and add texture features to construct a more comprehensive estimation model. In summary, Stacking could accurately and effectively estimate citrus LNC, providing the potential to estimate the nitrogen content in citrus leaves using satellite remote sensing.