Combining multiple spectral preprocessing and wavelength optimization methods improves potato aboveground biomass estimation

Abstract

Aboveground biomass (AGB) reflects the accumulation of crop photosynthesis, and AGB data guide agricultural production and field management practices. AGB can be estimated using UAV hyperspectral data; however, external factors and high-dimensional data lead to uncertainties. To address these issues, a cascading spectral preprocessing and band-optimized AGB estimation framework are proposed. We collected canopy hyperspectral reflectance and potato AGB data across two varieties, three planting densities, four nitrogen levels, and two potassium treatments during three growth stages. Then, we systematically compared the performance of Savitzky-Golay (SG) smoothing, multiplicative scatter correction (MSC), first-order differentiation (FOD) and their cascaded combinations. We also rigorously evaluated the ability of competitive adaptive reweighted sampling (CARS), successive projection algorithm (SPA) and their cascaded combination (CARS-SPA) to identify sensitive bands. The results indicated that cascaded spectral preprocessing methods significantly enhance the accuracy of potato AGB estimation. Among these approaches, the SG-MSC-FOD cascade performed most effectively. The combination of CARS and SPA yielded the fewest model variables while achieving the highest estimation accuracy. Furthermore, the integration of SG-MSC-FOD and CARS-SPA with partial least squares regression achieved the highest accuracy in AGB estimation across multiple growth stages, with a coefficient of determination (R²) of 0.73, root mean square error (RMSE) of 256.09 kg/hm², and normalized root mean square error (NRMSE) of 21.51 %. We validated the proposed method under different varieties, planting densities, and nitrogen and potassium treatments. This approach effectively reduces noise, lowers dimensionality, and enhances AGB estimation accuracy, providing a reliable solution for monitoring potato crop growth using hyperspectral remote sensing.