Predicting grain yield using canopy hyperspectral reflectance in wheat breeding data | JoVE Visualize

Area of Science:

Agricultural Science
Plant Breeding
Remote Sensing

Background:

Modern agriculture utilizes hyperspectral cameras to capture reflectance data across numerous narrow spectral bands.
This data is used to create vegetation indices (VI) for predicting crop traits like biomass, but VIs often discard significant spectral information and lack cultivar robustness.
Conventional VIs are limited as they only use a subset of spectral bands, leading to information loss and reduced prediction accuracy for key agricultural traits.

Purpose of the Study:

To develop and compare novel models that leverage all available spectral bands for enhanced prediction accuracy of wheat grain yield.
To evaluate the performance of models using all spectral bands against traditional vegetation indices (VIs) in diverse environmental conditions.
To assess the impact of integrating spectral data from multiple time-points and specific high-heritability bands on prediction accuracy.

Main Methods:

Utilized a dataset of 1170 wheat genotypes from CIMMYT's global program, evaluated across five environments for grain yield.
Collected hyperspectral reflectance data in 250 narrow bands (392-851 nm) at multiple time-points.
Compared advanced models (Ordinal Least Square, Bayes B, functional B-spline, functional Fourier, functional partial least square) using all bands against conventional OLS with individual/combined VIs.

Main Results:

Models analyzing all spectral bands simultaneously significantly outperformed traditional VIs in predicting wheat grain yield.
Functional B-spline and functional Fourier models exhibited the highest prediction accuracy across nine time-points.
Incorporating data from multiple time-points offered a marginal increase in accuracy, while using high-heritability bands (heritability > 0.5) in specific environments (Drought) showed notable improvements.

Conclusions:

Simultaneous analysis of all hyperspectral bands offers a more robust and accurate approach for predicting crop traits like grain yield compared to using limited vegetation indices.
Functional B-spline and Fourier models represent advanced, highly accurate methods for hyperspectral data analysis in crop breeding.
Strategic selection of spectral bands based on heritability and temporal data integration can further enhance prediction models for agricultural applications.