Abstract: Aboveground biomass (AGB) of vegetation is one of the most important indicators to evaluate the ecosystem health and sustainable utilization. Among them, the Loess Plateau in northern Shaanxi is characterized by the multi-dimensional attributes of an agricultural and pastoral intertwined zone, an ecological barrier area, and a carbon sink potential area. It is of great significance to accurately estimate the vegetation AGB for the ecological balance and biodiversity. The regional carbon sink potential can also be quantified to guide sustainable vegetation. Fortunately, the active and passive remote sensing data can be combined with machine learning for the broad application prospects in AGB estimation. This study aims to estimate the above-ground biomass of vegetation in the Loess Plateau of northern Shaanxi using improved gray wolf optimization of XGBoost. The variable parameters were also extracted from the Landsat 8 multispectral remote sensing data, Sentinel-1 radar remote sensing image data, and digital elevation data. The importance of feature variables was calculated using the random forest. The top 15 high-importance feature variables were selected as the modeling factors, because the higher IncNodePurity indicated the stronger variable importance. According to the optimal features and field survey data, four empirical models were established, including the multiple linear regression (MLR), back propagation neural network (BPNN), random forest (RF), and extreme gradient boosting (XGBoost). The AGB values of forest and shrub-grass land were estimated to compare the accuracy of different models. The improved gray wolf optimization (IGWO) was introduced to determine the optimal parameter combinations. The XGBoost model was enhanced for the best performance of the improved model. Finally, the optimal model was obtained to invert the vegetation AGB. The results showed that the texture features shared the outstanding advantages in the AGB modeling. The near infrared band (B5) and cross-polarization (VH) features were of high importance in the forest AGB, while the short-wave infrared bands (B6 and B7) and co-polarization (VV) features were more important in the shrub-grass AGB. The XGBoost model was adjusted for the minimal fitting errors after the residual correction and regularized optimization. Thereby, the higher predictive accuracy was achieved after optimization. The IGWO-XGBoost model that was constructed by the IGWO algorithm performed the best to invert both forest and shrub-grass AGB, where the determination coefficients (R²) were 0.823 and 0.735, respectively, with the increase by 0.070 and 0.053, respectively. And the root mean square error (RMSE) decreased by 12.1% and 9.9%, respectively. The average forest AGB was 47.26 t/hm², which was distributed mainly between 30-60 t/hm². While the average shrub-grass AGB was 1.96 t/hm², which was distributed mainly between 1.00-2.25 t/hm². The overall spatial distribution of AGB shared a pattern of "high in the south - broken in the middle - low in the north". There were some patterns of differentiation with the variation in the topography (elevation and altitude). The average AGB of forests showed a gradual increase in the distribution with elevation, which was consistent with the anthropogenic disturbances. The average AGB of shrubs and grasses showed the opposite pattern, due mainly to the decrease of the soil nutrients at high altitude and the air temperature, as well as the harsh conditions for the vegetation growth. There were some differences in the biomass on the different slopes, with the highest distribution of AGB in the forest on half-shady slopes and the highest distribution of AGB in the shrub-grass on the sunny slopes. There were the highest average AGB values of the forest and shrub-grass on the shady slopes, with the better standing conditions. This finding can provide the technical support for the accurate estimation of the AGB on the Loess Plateau in northern Shaanxi. A scientific basis can also offer for subsequent studies, such as ecosystem function evaluation.