Nondestructive Estimation of Circadian Time in Harvested Green Perilla Leaves Using Hyperspectral Data

Vegetables that are harvested early in the morning or late in the afternoon are valued in plant production based on the diurnal variation of metabolism (Clarkson et al., 2005), which is regulated by a circadian clock with a period of approximately 24 hours. The circadian clock plays an important role in the regulation of biological processes, such as growth, photosynthesis, and flower induction (Barak et al., 2000; Dodd et al., 2005). Therefore, information on circadian time, that is, the internal body time denoted by the circadian clock, is useful in improving the quality of plant production. Recently, a statistical oscillatory analysis of time-series RNA sequencing (RNA-Seq) data, referred to as a molecular timetable method (MTM) (Ueda et al., 2004), was used to estimate the circadian time of various types of plant leaves, including lettuce and tomato (Higashi et al., 2016; Takeoka et al., 2018). A few hundred genes identified as time-indicating genes (TiGs) were found to exhibit circadian rhythms in their expressions, and the overall profile of their phases was found to represent circadian time. Although this MTM successfully estimates circadian time precisely, it requires extraction of RNA, which involves destroying plant tissues, and it also requires considerable time for sequencing. In generally, optical features of plant tissues have been used in nondestructive real-time methods. Near-infrared spectroscopy can be used to estimate the soluble solids content of vegetables and fruits (Khuriyati and Matsuoka, 2004), and a combination of visible and near-infrared wavelengths can be used to estimate the amount of chlorophyll (Markwell et al., 1995). Multispectral imaging (MSI) with a few dozens of different spectral bands has been shown to have an excellent ability to determine the spatialspectral signature of plants, especially in characterizing a variety of chemical compositions and assessing the physiological status of plants. A recent study reported that MSI can be used nondestructively to detect circadian rhythms of chlorophyll concentration in soybean leaves (Pan et al., 2015). Therefore, it believed that circadian time can be estimated by means of multispectral analysis, although this has not been previously confirmed. In this study, we used a hyperspectral camera, which is a device with an exceptionally high resolution (more than 100 bands) from visible to near-infrared wavelengths, because optical indices for estimation of circadian time have not been identified. In addition, it is expected that rich information is required for precise estimation of circadian time like as many TiGs in MTM. We also used a machine learning method based on an artificial neural network (ANN) to address nonlinearity of the relationships among wavelengths. As a starting point, we focused on the circadian time at harvest, which is a critical time for production quality. Our experiments were carried out using green perilla (Perilla frutescence var. crispa f. viridis), a