Use of Hyperspectral Image Data Outperforms Vegetation Indices in Prediction of Maize Yield

Abstract

Hyperspectral cameras can provide reflectance data at hundreds of wavelengths. This information can be used to derive vegetation indices (VIs) that are correlated with agronomic and physiological traits. However, the data generated by hyperspectral cameras are richer than what can be summarized in a VI. Therefore, in this study, we examined whether prediction equations using hyperspectral image data can lead to better predictive performance for grain yield than what can be achieved using VIs. For hyperspectral prediction equations, we considered three estimation methods: ordinary least squares, partial least squares (a dimension reduction method), and a Bayesian shrinkage and variable selection procedure. We also examined the benefits of combining reflectance data collected at different time points. Data were generated by CIMMYT in 11 maize (Zea mays L.) yield trials conducted in 2014 under heat and drought stress. Our results indicate that using data from 62 bands leads to higher prediction accuracy than what can be achieved using individual VIs. Overall, the shrinkage and variable selection method was the best‐performing one. Among the models using data from a single time point, the one using reflectance collected at 28 d after flowering gave the highest prediction accuracy. Combining image data collected at multiple time points led to an increase in prediction accuracy compared with using single‐time‐point data.

Plant phenotyping relevance

ハイパースペクトル画像からトウモロコシ収量を推定する予測手法を複数比較・評価しており、植物形質の取得・推定法が研究の中心である。

Code and data availability

The paper's hyperspectral reflectance and maize grain yield phenotyping data (11 CIMMYT trials, 2014) are stated to be publicly available at the CIMMYT data repository. No specific dataset identifier is given, and no author analysis code URL is provided (scripts only appear in Supplemental Methods without a public URL,