Averaging Techniques in Processing the High Time-resolution Photosynthesis Data of Cherry Tomato Plants for Model Development

Abstract

Sensor-based plant diagnostic technology is an essential feature of the speaking plant approach (Hashimoto, 1989; Takayama, 2013). A recent plant diagnosis technology, which has been started to be installed in practical agricultural production in greenhouses, is the real-time photosynthesis and transpiration monitoring system (Shimomoto et al., 2020). The system allows to remotely and continuously measure photosynthesis of whole plants under greenhouse conditions without any contact or intrusive. Also, data are sampled with high time-resolution, which are recorded in 5-minute intervals. With these features, the system is desirable for precise quantification of plant responses to stimuli at the whole-plant level. The numerous photosynthesis data produced by the photosynthesis monitoring system can be used for analysis and forecasting by way of modeling. Estimation models for the net photosynthetic rate (Pn) as a function of environmental factors for greenhouse tomato utilizing numerous data were limited to the whole greenhouse, such as the work of Nederhoff and Vegter (1994a). They used three variables of incident photosynthetically active radiation (PAR), CO2 concentration, with double rectangular hyperbolic relation and leaf area index (LAI) to calculate the net canopy photosynthesis rate of tomato greenhouse with an R of 0.892. In another work, they modified two established mechanical models of Acock (Acock et al., 1978) and Thornley (Thornley, 1976) to fit their data and gave Rs of 0.893 and 0.817, respectively (Nederhoff and Vegter, 1994b). On the other hand, the accuracy of sensor-based technology in modern greenhouses is jeopardized by disturbances, such as inaccuracy of the measurements by the sensor itself due to dynamic variations in the greenhouse climate (van Mourik et al., 2019), especially when measurements are performed in high time resolution. Therefore, it leads to an issue of how to address the high time-resolution data produced by the sensors for further use of model development. As most real-time data contain erroneous values and noise, such data required further processing for filtering noise and smoothing. Averaging techniques commonly used for smoothing data include the moving average (Čampulová, 2018) and the simple average (Yaffee and McGee, 2000). When the data are correctly prepared, the quality of the model can be reliable (Pyle, 1999). The objective of the present study was to process the