Abstract: Rapeseed is one of the most important raw materials of edible vegetable oil. An accurate and timely yield prediction is crucial to national food and oil security. Unmanned aerial vehicle (UAV) hyperspectral technology can be expected to effectively enhance the data acquisition in traditional satellite remote sensing. A large volume of continuous narrow-band spectral data can also be captured to accurately characterize the physiological and biochemical features of the crops. In this study, the UAV platforms were utilized to capture the hyperspectral images during the flowering stage of the rapeseed. A yield prediction model was constructed using fractional order differentiation (FOD) and multi-band spectral indices. A systematic prediction of the yield was also evaluated on these spectral indices. Firstly, the FOD processing was applied to the hyperspectral data of the rapeseed canopy, and then two-dimensional (2D) and three-dimensional (3D) spectral indices were calculated using different order differential data; Secondly, Pearson correlation coefficient was utilized to examine the correlation between the spectral indices and yield observation. The most sensitive spectral indices were selected for the yield prediction; Finally, the support vector regression was employed to construct the yield prediction model using FOD spectral indices. A systematic investigation was carried out to evaluate the impact of different differential orders and spectral indexes on the prediction accuracy. The results indicate that the FOD processing enhanced the spectral characteristics of the red edge and yellow-green bands during the flowering stage of rapeseed. The potential spectral information was effectively extracted to preserve the original structure of the vegetation spectral curve. The correlation analysis showed that there was a generally low correlation between FOD spectral data and yield at the lower orders. The increase was observed at the higher orders. The excessively high orders (e.g., 2.0) were selected to introduce the noise into the spectral data, which reduced the correlation. Three types of the 3D spectral indices exhibited correlation coefficients of 0.77 with the yield, which were significantly higher than those of the 2D ones. The 2D spectral index with the FOD shared the highest correlation at an order of 1.8, with a correlation coefficient of 0.868, whereas the 3D spectral index shared the highest correlation at an order of 1.6, with a correlation coefficient of 0.887. The estimation of the yield was also carried out with the different spectral indices. Furthermore, the indices derived from the blue, green, and near-infrared bands were the most sensitive to the prediction of the rapeseed yield. The third spectral dimension in the 3D spectral index greatly contributed to the full utilization of the rich information in hyperspectral data. The yield prediction model with the 3D spectral index also outperformed that with the 2D spectral index. The R² values of the 3D and 2D spectral index ranged from 0.880 to 0.897 and from 0.624 to 0.896, respectively. The high accuracy and robustness were achieved in the yield prediction model using FOD with the multi-dimensional spectral indices. The high-precision early estimation of the yield also provided valuable scientific support to agricultural production. Future research should further explore the impact of rapeseed varieties, growth stages, and environmental conditions on yield prediction with the FOD spectral index. The potential application can also be extended to other crops. Additionally, future studies should explore more to minimize the noise impact in the multi-order differentiation, and then balance the trade-off between spectral resolution, spectral intensity, and noise. The more robust models can provide the data support for rapid, accurate, and early yield prediction.