Abstract: Abstract: Crude protein (CP) is the key indicator for evaluation of the quality and feeding value of pasture grass. Timely, accurate and non-destructive assessment of pasture grass CP content is important for pasture grass growth monitoring and making-decisions for adjusting stocking rate and pasture management, eventually preventing grassland degradation. Hyperspectral remote sensing technology provides the potential for monitoring the nutrition in large areas of grassland. In order to obtain the distribution of pasture grass CP content in Jinyintan Grassland, which is a typical prairie in Haiyan County, Qinghai Province, a new type of hyperspectral imaging system based on high altitude airship (named ASQ-HAA380) was used to collect the high-resolution hyperspectral images, and the ground-based pasture grass CP samples datasets were collected at the same time and analyzed in Qinghai University. The aim of this study was to establish the regression model and seek the optimal model to estimate CP content and draw its distribution map. This study analyzed the possibility using several spectral variables and different modeling methods. On the basis of a comprehensive analysis of the hyperspectral data, the best spectral indices i.e. simple ratio spectral index (SR) and normalized difference spectral index (ND) were taken as independent variables to build univariate models. Besides, the multivariate stepwise linear regression method and multivariate nonlinear regression method were used to build estimate models of other spectral variables, including original reflectance spectrum (R), first derivative of reflectance (D(R)), logarithm transformation of reflectance (log(R)), normalized transformation of reflectance (N(R)), first derivative of log(R) (D(log(R))), logarithm transformation of N(R) (log(N(R))), first derivative of N(R) (D(N(R))), band depth (BD), and continuum removed derivative reflectance (CRDR). Afterwards, the accuracies of these models were evaluated through cross-validated coefficient of determination (R2) and cross-validated root mean square error (RMSE). The results showed: 1) Derivative spectral variables could effectively estimate CP content with high stable ability among all spectral variables, and R2 is more than 0.794. 2) Compared with these multiple regression models, the nonlinear regression model had higher precision than the corresponding linear regression model. 3) The accuracy of the multiple nonlinear regression model of D(log (R)) built in the study was the highest, R2 was 0.918 and RMSE was 0.054, and the model of D(log(R)) was the optimal model for prediction of CP content. The inversion nonlinear regression model of D(log(R)) was applied to the hyperspectral image to obtain the spatial distribution of CP content in the study area. The research provides reference and technical basis for the quantitative inversion of CP content in large area scales and the efficient implementation of precision livestock husbandry based on hyperspectral images, and also lays the foundation for the development of wisdom livestock husbandry in the future.