Résumé
•High-frequency evapotranspiration (ET) phenotyping is key to drought-stress analyses.•Optimum ET sampling frequency identified by downscaling 15-min interval ET profiles.•Time series forecasting of ET was performed using ensemble machine-learning models.•ET forecasting and genotype classification performance was compared at each scale.•60-min interval of ET was found optimum with minimum redundancy and information loss.
Efficient selection of drought-tolerant crops requires identification and high-throughput phenotyping (HTP) of the complex functional (especially canopy-conductance) traits that elicit plant responses to continually fluctuating environmental conditions. However, phenotyping of such dynamic physiology-based traits has been immensely challenging especially due to the limited availability of adequate methods that can provide continuous measurements of plant-water relations. Therefore, gravimetric phenotyping of plants is being increasingly used to allow one-to-one monitoring of plant-water relations and generate continuous evapotranspiration (ET) profiles. The gravimetric sensors or load cells can provide ET estimates at very high frequencies, e.g. 15-min interval, as chosen by the user. There is however, no study on understanding the optimum frequency or the sampling time at which ET needs to be monitored, such that data-redundancy, noise and processing overhead could be reduced. Hence, this paper makes a novel attempt in identifying the optimum sampling time for phenotyping ET from load cells time series. The proposed procedure includes an ensemble Machine-Learning (ML) approach for optimizing the sampling time through time series forecasting of ET profiles and classification of genotypes using the forecasted ET values. High-frequency load cells data from the LeasyScan, HTP platform, ICRISAT were used to derive the ET profiles at frequencies or scales varying from 15-min to 180-min, followed by ET forecasting and classification at each frequency. For both forecasting and classification, an ensemble of three ML algorithms i.e. Support Vector Machines (SVM), Artificial Neural Network (ANN) and Random Forests (RF) were leveraged. Consequently, the performance metrics (of both the operations) obtained from the ensemble were used to compute the entropy-based optimum sampling time. The results reveal that 60-min interval HTP data could be credibly used for both, forecasting ET as well as correctly classifying the genotypes.