Multi-Scale Time Series Analysis Of Evapotranspiration For High-Throughput Phenotyping Frequency Optimization

Abstract

This work is undertaken considering the significance of functional phenotyping (primarily measured from continuous profiles of plant-water relations) for crop selection purposes. High-Throughput Plant Phenotyping (HTPP) platforms which largely employ state-of-the-art sensor technologies for acquisition of vast amount of field data, often fail to efficiently translate sensor information into knowledge due to the major challenges of data handling and processing. Hence, it is imperative to concurrently find a way for dissociating noise from useful data. Additionally, another important aspect is understanding how frequent should be the data collection, so that information is maximized. This paper presents a novel approach for identifying the optimal frequency for phenotyping evapotranspiration (ET) by assimilating results from both time series forecast as well as classification models. Thus, at the optimal frequency, plant-water relations can not only be desirably predicted but genotypes can also be classified based on the characteristics of their ET profiles. Consequently, this will aid better crop selection, besides minimizing noise, redundancy, cost and effort in HTPP data collection. High frequency (15 min) ET time series data of 48 chickpea varieties (with considerable genotypic diversity) collected at the LeasyScan HTPP platform, ICRISAT is used for this study. Time series forecast and classification is performed by varying frequency up to 180 min. Multiple performance measures of time series forecast and classification are combined, followed by implementation of entropy theory for sampling frequency optimization. The results demonstrate that ET time series with a frequency of 60 min per day potentially yield the optimum information.