Improving the prediction of water stress-related traits in open-field tomato using multivariate models and variable importance-based indices

Abstract

The assessment of water stress levels in plants should be essential part of precise irrigation management, and a good and quick method is useful in plant phenotyping. There are many options for this task, but the effectiveness varies between crop species and environments. Apparently open field applications face the most difficulties. This study aimed to test a large number of vegetation indices (VIs) and multivariate models based on hyperspectral reflectance data (325-1075 nm) regarding their correlation and prediction abilities to leaf stomatal conductance, relative water content (RWC), and detailed chlorophyll, and carotenoid components. Data of the abovementioned variables was collected during three consecutive growing seasons in processing tomato cultivated under different water supply regimes to provide data with varying water stress levels. Then the relation of the measured variables to 226 VIs was tested created according to the formulas collected in the Index DataBase (IDB Project, indexdatabas.de ). New VIs were also developed derived from the most important variables of the ML algorithms, customised to tomato water stress assessment. Standard normal variate and its combination with Savitzky-Golay first derivative were used for pre-processing the spectra and principal component regression (PCR), partial least squares regression (PLSR), elastic net (ENET), support vector regression (SVR), random forest (RF) and extreme gradient boosting (XGB) algorithms were tested. The newly developed indices outperformed the existing formulas, except in the case of β-carotene. The most reliable index was developed for RWC estimation; that was the difference of the reflectance on the 986 and 701 nm wavelengths. The ENET and SVR algorithms produced the best models depending on the pre-processing method. The blue, near-infrared (NIR) and green regions, respectively, were the most important regarding all models according to the variable importance analysis. The model with the best metrics was developed for chlorophyll-a (R 2 =0.82, nRMSE=11%, RPIQ=2.41), followed by RWC (R 2 =0.72, nRMSE=14%, RPIQ=2.53).

Plant phenotyping relevance

ハイパースペクトル反射データと多変量・機械学習モデルを用いて、トマトの水ストレス関連生理形質を推定し、新規指標も開発・評価しているため、表現型取得・推定手法が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.