Accurate monitoring of above-ground biomass (AGB) is an important part of farm production management, so rapid and accurate estimation of AGB is important for the development of precision agriculture. Traditionally, AGB has been obtained using destructive sampling methods, which makes large-area, long-term measurements difficult. However, with the advancement of science and technology, UAV hyperspectral remote sensing has become the most effective technical means to estimate AGB of large crops because of its advantages of high mobility, high spectral resolution and map integration. In this study, the canopy hyperspectral images and actual AGB data of potato tuber formation, tuber growth and starch accumulation stages were obtained by carrying imaging spectrometer sensors on the UAV platform and drying and weighing method, respectively. Correlation analysis method (CAM), random frog method (RFM) and Gaussian process regression bands analysis tool (GPR-BAT) were used to screen canopy original spectra (COS) and first-order derivative spectra (FDS) for sensitive wavelengths, respectively, combined with partial least squares regression (PLSR) and Gaussian process regression (GPR) techniques to establish AGB estimation models for each fertility period and the estimation effects of different models were compared. The results showed that (1) the effect of combining the two regression techniques based on the characteristic wavelengths screened by the same method for COS and FDS to estimate AGB all changed from good to bad from the tuber formation stage to the starch accumulation stage. (2) Based on the characteristic wavelengths screened by the three methods of FDS respectively, the models constructed by homogeneous regression techniques are more effective than those based on COS accordingly. (3) The number of characteristic wavelengths screened based on COS and FDS using CAM, RFM and GPR-BAT methods were 28, 12, 6 and 12, 23, 10 at the tuber formation stage, 32, 8, 2 and 18, 28, 4 at the tuber growth stage, and 30, 15, 3 and 21, 33, 5 at the starch accumulation stage, respectively. (4) The effect of sensitive wavelengths for AGB estimation based on COS and FDS screened by three methods at each reproductive stage were GPR-BAT, RFM and CAM in descending order. (5) The models based on sensitive wavelengths screened by FDS through the GPR-BAT method at each fertility stage combined with PLSR were more accurate and stable with R² of 0. 67, 0. 73 and 0. 65, NRMSE of 16.63%, 15.84% and 20.81%, respectively. This study shows that AGB can be accurately estimated using UAV hyperspectral imaging technology, which provides scientific guidance and reference for achieving dynamic monitoring of potato crop growth.