Crop breeding is currently hindered by the resource-intensive phenotyping of agronomically relevant plant traits in large breeding populations. Leaf hyperspectral reflectance data are gathered rapidly, in vivo , in a non-destructive, high-throughput manner, and have been effectively used to predict traits that are otherwise challenging to measure. However, hyperspectral reflectance data have not been used for feature engineering that may further boost performance of machine learning models for various crop traits. Here we propose a biophysical approach that relies on a non-canonical use of Fourier transform, rooted in our observation of power-laws in the power spectral density of hyperspectral reflectance data both in field and greenhouse acquired data at a leaf level. Using legacy and newly generated data from plant and crop species, we demonstrated that the exponents of these power-laws: ( i ) show differences between species and exhibit natural variability within a species, ( ii ) demonstrate high correlations to leaf biochemical, structural, and physiological traits within and between environments, and ( iii ) exhibit moderate to high broad-sense heritability. Therefore, our study provides an alternative approach for using hyperspectral reflectance data in speeding up crop breeding. • Non-canonical use of Fourier transform identifies 1/f-noise for HSR data. • The power-law exponents of 1/f-noise describe HSR data in lower dimension. • The power-law exponents show differences between and within species. • The exponents of 1/f-noise of HSR data show moderate-to-high heritability. • These exponents are predictive of leaf traits using simple correlation analysis.
Rahimi‐Majd et al. (Sat,) studied this question.