The accurate and timely estimation of above-ground biomass (AGB) is crucial for grassland monitoring and rational grazing. The unmanned aerial vehicle (UAV)-based hyperspectral remote sensing can obtain images with high spatial, spectral and radiometric resolutions in a short time and has been widely used in many fields such as precision agricultural and high-throughput plant phenotype. To explore the applicability of UAV-based hyperspectral image (UAV-HSI) in grassland AGB prediction, we collected UAV-HSI, grassland AGB, and relevant auxiliary data in a grassland sample area of Qinghai Province. However, it is not only inconvenient to widely collect, store and transmit, but also inefficient in data processing for UAV-HSI because of large data volumes, which may restrict practices of UAV-HIS widely. For resolving the above problems, a spectral reconstruction and optimization method considering both data simplification and spectral fidelity was proposed, attempting to ensure the performance of grassland AGB prediction and effective reduction of data volumes. First, using the residual quantization, we obtained several binary cubes (H_i) and corresponding coefficient matrices (β_i) with low volumes. H_i and β_i can replace the original data for storage and transmission. Second, preliminarily reconstructed spectra (PRS) can be produced by H_i and β_i. Third, through the Savitzky-Golay (SG) filter, optimized PRS (OPRS) can be achieved by enhancing the spectral fidelity. To demonstrate the effectiveness of OPRS, we carried out the spectral fidelity experiment. Taking a grassland canopy spectrum as an example, three fidelity indices, i. e., the spectral vector distance (SVD), spectral correlation coefficient (SCC), and spectral angle mapping (SAM), were analyzed. Results showed that, on the three fidelity indices, OPRS was superior to PRS. And then, the correlations between AGB and OPRS bands were discussed. Compared with raw spectra and PRS, OPRS achieved the relatively high and most stable potential in forage AGB prediction. Furthermore, the partial least squares regression (PLSR) was used to calibrate models of grassland AGB prediction. Results demonstrated that, among raw spectra, order -1 to -4 PRS, and order -1 to -4 OPRS, The prediction performances of order -4 and -3 OPRS reached the optimal and sub-optimal levels with RPD (the ratio of performance to deviation)=2.31 and 2.23, respectively. Their RPD values were 0.26 and 0.18 higher than that of the original spectra, respectively. Therefore, with a reduction of one order of magnitude, OPRS achieved a better performance than the original spectrum in the grassland AGB prediction. In other words, OPRS had advantages with both data simplification and accurate grassland AGB prediction. This study provides a new solution to estimate grassland AGB for UAV-HIS effectively.