Calibration Methodology of a Hyperspectral Imaging System for Greenhouse Plant Water Stress Estimation

Much progress has been made on optimizing plant water supply based on different methods for irrigation scheduling, in both open-field and greenhouse cultivations, such as real-time measurements of solar radiation and soil or substrate water content. However, only a limited number of methods use plant-based physiological indicators to detect plant water stress and adapt irrigation scheduling accordingly. In addition, even fewer indicators can be estimated by non-contact, remote sensors (RS) that do not affect plant development. Hyperspectral imaging could be an accurate remote way to detect plant water status, taking into account crop characteristics. In this work, a methodology of hyperspectral imaging calibration and acquisition is presented. The method uses the reflectance characteristics in hyperspectral bands from 400 to 1000 nm and incorporates the appropriate radiometric and geometric corrections. The basic statistical parameters of mean and standard deviation values are used to estimate spatial and spectral correlation of each band on the extracted areas/pixels of interest. Several statistical techniques are used for the selection of optimal features that will lead to the development of appropriate plant water stress indices that could be used for incipient water stress detection in optimal irrigation scheduling systems. INTRODUCTION Greenhouse irrigation management, especially in hydroponic systems, needs a series of short-time water applications (10-25 doses per day). Even though several methods are used to detect plant water deficit, only a few methods use plant-based physiological indicators. Katsoulas et al. (2006) stressed the need for the creation of a suitable method to control irrigation frequency and proposed a technique based on crop transpiration. However, that method implies the knowledge of a crop coefficient that varies for different planting periods. Crop reflectance (Knipling, 1970), fluorescence (Norikane and Kurata, 2001) and thermal radiation transmittance (Jones and Schofield, 2008) are affected by plant water status. Several studies have attempted to detect and quantify water stress -through appropriate indicesusing reflectance in the visible and the near infrared regions (Sellers, 1985; Penũelas et al., 1993; Schlemmer et al., 2005). The use of a hyperspectral camera to identify plant reflectance variations related to leaf water deficit levels is a non-destructive and fast measurement method. Hyperspectral imaging technology could be used to study leaf reflectance changes caused by different water stress levels in more than one leaf, enhancing the reliability and sensitivity of plant water detection (Graeff and Claupein, 2007; Zhou et al., 2011). The reflectance sensor collects such plant-based data from the spectrum, capturing the energy reflected and emitted by the plants, while this energy changes according to leaf chemical compounds and water content (Fig. 1). A typically healthy plant has a small pick in the green band, a small drop in the blue and red bands (due to chlorophyll absorption), a rising peak in the near infrared (NIR) band (due to scattering by air content in sponge cavities) and a falling peak in the middle infrared MIR band (due to water stored in thylakoids that absorbs more radiation at that spectrum). The plant reflectance signal varies according to water stress emergence and chemical compounds metabolics. Until recently, the reflectance identity of the crop was defined by laboratory protocols and spectroradiometers that measured radiance reflectance coming from only a single point of the target. Even though this kind of sensors present low levels of equipment noise, they cannot give representative reflectance data of the canopy, due to leaf structure variability. Hyperpsectral imaging technology is capable of measuring reflectance in more than one leaves (or plants) and gives more reliable data for the canopy’s reflectance identity. However, this optic technique includes a variety of equipment noises that should be taken into consideration before plant image acquisition. Different types of sensor noise include the detector dark current, the sensor temperature, the readout noise, the exposure shot noise and should eventually be removed with statistical techniques and special filters, using samples of known reflectance (Polder et al., 2003). Even though this remote sensing technique has been successfully used for years in open field cultivations and relevant reflectance calculation models have been developed, it has not been extensively tested in the case of greenhouse crops. It has to be noted that open field methods cannot directly be applied in greenhouses due to difficulties arising mainly from shadows resulting from the greenhouse frame and equipment. The problems related to the greenhouse structure shadows or to other obstacles (like old leaves and soil background) could be eliminated by forming vegetation indices using the combination of data from two or more spectral bands (Jackson and Huete, 1991). According to Zakaluk and Sri Ranjan (2008), the most common forms of reflectance indices are the following: (1) reflectance ratios corresponding to the ratio of two spectral bands, which are referred to as “simple ratio” (SR) vegetation indices and (2) normalized difference (ND) vegetation indices, which are defined as ratios of the difference in reflectance between two spectral bands over the sum of the reflectance at the same bands. Consequently, the aim of this work is to study the possibility of detecting plant water stress in greenhouses using a hyperspectral imaging methodology and furthermore, to study the effect of system settings on the reflectance measurements and the resulting plant water stress indices. MATERIALS AND METHODS The hyperspectral camera Imspec V10 (Spectral Imaging Ltd, Finland) was used, which operates in the visible and near infrared (VNIR) ranges of 400-1000 nm. It was used as a push broom line scan camera and provided full spectral information for each pixel. The hyperspectral camera was attached to a rotary scanning system, in which, scanning speed and angle were controlled. A spectral data acquisition software was used to set the operational parameters of the camera, to start data acquisition and to monitor on-going tasks. The camera’s specifications and settings were: spectrograph: V10, spectral range: 400-1000 nm, spectral resolution (30 mm): 2.08 nm, spectral resolution peak: 435.8 nm (2.86 FWHM/nm), 696.5 nm (3.34 FWHM/nm), 912.3 nm (3.33 FWHM/nm), slit width: 8 mm, pixels in full frame: 1312 x 1024, exp. time range: 0.1-500 ms. The camera system was placed on a moving cart, so that images of the vertical canopy axis could be obtained. The hyperspectral imaging system was calibrated inside a light-controlled growth chamber. Light intensity was controlled with high pressure sodium lamps, 600W each. The chamber included 24 lamps in total (6 lamps per lightintensity level) with a maximum light intensity of 240 W m -2 . For extra illumination of the target area (70 x 100 cm), four quartz-halogen illuminators (500 W each) were used to provide calibration wavelength from 400 to 900 nm. The optic system was placed at a distance of 1 m from the target (white panel or plant). A spectrally flat black surface was placed as a background, to ensure a constant field of view without any shadows. The calibration of the hyperspectral imaging system requires geometric and radiometric calibration (Lawrence et al., 2003). Geometric calibration eliminates optical errors, such as curvature distortion of the spectral lines. The system was already geometrically calibrated by the supply company. Radiometric calibration includes the elimination of a variety of noise sources, such as photon noise, thermal noise, read out noise and quantisation noise. The proper number of lens aperture (f/) and exposure time (ms) ranges of the camera for the specific light signal conditions were evaluated, in order to achieve the most suitable readout values. The MATLAB software package (by MathWorks ® ) was used for image analysis. The acquired images were improved based on the above factors, by using the radiometric equation: