Crude protein (CP) content is an crucial parameter for assessing the nutritional value of pasture grasses, and the crucial parameter for assessing of pasture CP is of significant importance for the evaluation of grassland degradation, grassland classification, and the advancement of high-quality animal husbandry practices.. This study focused on temperate desert grassland and temperate grassland in Inner Mongolia, using hyperspectral, unmanned aerial vehicle (UAV) aerial data and ground-measured data. Lasso regression was employed to identify sensitive spectral bands based on a comparison of hyperspectral data preprocessing methods. The results indicated that the preprocessing methods, including First Derivative (FD), Standard Normal Variate (SNV), and Multiplicative Scatter Correction (MSC), enhanced the correlation between the hyperspectral data and CP content of pasture grasses; the strongest correlation between hyperspectral reflectance and CP content was observed in the blue spectral bands (429, 413, 433 and 494 nm,) and red edge regions (631 nm, 720 nm, 729 nm, and 738 nm). Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to significantly reduce the number of sensitive bands. method was applied to reduce the number of sensitive bands significantly. bands. For the development of the estimation model for CP content in pasture grasses, the results indicated that the SNV treatment exhibited superior performance in modeling the CP content of both the combined samples of the two grass types and the individual samples of the temperate desert grassland in the test set, for the temperate grassland samples, the FD treatment demonstrated greater efficacy. Among the three machine learning models-Partial Least Squares (PLS), Random Forest (RF), and Support Vector Machine (SVM) constructed for the combined samples of the two grassland types, SG-SNV-Lasso-PLSR was determined to be the optimal model, exhibiting an R2 of 0.88, RMSE of 1.23 mg center dot g-1, and RPD of 2.89. For the separate samples of the two grassland types, the optimal model was SG-SNV-Lasso-PLSR, with R2 = 0.88, RMSE = 1.23 mg center dot g-1, and RPD = 2.89, respectively. The optimal models were SG-SNV-Lasso-RF (temperate desert grassland) and SG-FD-Lasso-SVM (temperate grassland), both of which demonstrated good prediction performances. This study proposes a rapid monitoring technology for natural pasture nutrient composition, and the related results provide a significant reference for the development and management of smart pasture systems, as well as grassland protection and restoration efforts.