Field radiometric calibration of a multispectral on-the-go sensor dedicated to the characterization of vineyard foliage

Computers and Electronics in Agriculture 123 (2016) 184–194

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag

Original papers

Field radiometric calibration of a multispectral on-the-go sensor dedicated to the characterization of vineyard foliage Marie-Aure Bourgeon, Jean-Noël Paoli ⇑, Gawain Jones, Sylvain Villette, Christelle Gée Agrosup Dijon, UMR 1347 Agroécologie, GESTAD Dept., Precision Agriculture Team, 26 boulevard Dr Petitjean, BP 86510, F-21000 Dijon, France

a r t i c l e

i n f o

Article history: Received 1 March 2015 Received in revised form 4 December 2015 Accepted 26 February 2016 Available online 9 March 2016 Keywords: Agroecology Precision viticulture Multispectral imaging Proximal sensing Radiometric calibration NDVI

a b s t r a c t The accurate assessment of the vigor and disease impact is a major challenge in precision viticulture. It is essential for managing phytosanitary treatments. Up to now, some remote sensing techniques such as aerial imagery and handheld optical sensors have been applied to grapevine characterization. However each technique provides limited, specific information about foliage. To broaden the characterization of the foliage, we developed a proximal integrated, multispectral imaging sensor that operates in the visible and near-infrared bands. It is mounted on a track-laying tractor equipped with a Greenseeker-RT-100, coupled with a GPS-RTK. As the sensor is very sensitive to the ambient light, a radiometric calibration is required: it allows producing absolute reflectance images, using a color chart. If the chart is hidden by leaves, for instance, the images are corrected using the linear interpolation method. The adaptive radiometric method is evaluated as a function of the number of neutral patches selected on the color chart during the linear regression process and the efficiency of the spatial interpolation method is assessed using a leave-one-out-cross-validation (LOOCV) method. The radiometric calibration is validated by comparison of NDVI maps produced by imagery and by the Greenseeker, a commercial system. In the early stage of berry formation, we examined and quantified the spatial patterns and demonstrated a low-cost imagery method that is capable of analyzing correctly the vigor. This corroborates the efficiency of the calibration method encouraging the use of multi-spectral imagery for other vineyard applications, such as the characterization of physiological status. Ó 2016 Elsevier B.V. All rights reserved.

1. Introduction Precision agriculture can be defined as a set of methods that improves farm productivity by assigning specific attributes to individual production areas (Robert et al., 1994; Pierce and Nowak, 1999; McBratney et al., 2005). A four-stage process is traditionally depicted: observation of crops over time with sensors, interpretation of georeferenced data, evaluation for agronomic management, and in-field implementation (Bramley, 2001; Bramley and Lamb, 2003; Mathews, 2013). When adapted to grapevine management, precision agriculture is called precision viticulture (Bramley and Proffitt, 1999; Tisseyre et al., 2007). It allows characterizing the within-field variability (Bramley and Lamb, 2003) using appropriate proximal, airborne or spaceborne remote sensing techniques (Strever, 2007; Mulla, 2013). Multispectral or hyperspectral sensors may generate their own light sources (e.g., active sensors as the Greenseeker RT-100 or the Multiplex) or use natural light (e.g., passive sensors such as imaging systems). One can measure ⇑ Corresponding author. E-mail address: [email protected] (J.-N. Paoli). http://dx.doi.org/10.1016/j.compag.2016.02.019 0168-1699/Ó 2016 Elsevier B.V. All rights reserved.

the reflectance, transmittance and fluorescence of plants using both types of sensors and extract information about their internal biophysical and biochemical properties (Gitelson and Merzlyak, 1996; Moran et al., 1997; Johnson et al., 2003; Goutouly and Cerovic, 2008; Whalley and Shanmuganathan, 2013; Quemada et al., 2014). In particular, the reflectance signal allows the computation of vegetation indices such as the reflectance ratio (Mulla, 2013; Quemada et al., 2014) or the normalized difference vegetation index (NDVI) (Choudhury, 1987; Solari et al., 2008) which are related to plant status: nitrogen, chlorophyll content, green leaf biomass. As far as vineyard is concerned, Whalley and Shanmuganathan (2013) reviewed the main parameters that are used to manage within-field variability. Other authors explored variation yield (Bramley, 2001), disease detection (Sankaran et al., 2010), or vine vigor (Lamb et al., 2001; Hall et al., 2002; Johnson, 2003; Cerovic et al., 2009). To date, only a few aerial studies have focused on grapevine foliage architecture (Moran et al., 1997; Mathews and Jensen, 2013). This has motivated the development of proximal sensing devices for a better positioning of the acquisition system for foliage observation (Debuisson et al., 2010) and real time

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

management of within-field variability (Bramley, 2001; Mulla, 2013). The Greenseeker is one of the most popular proximal sensor used to appraise vine vigor (Mazzetto et al., 2009) especially at early growth stage. Several studies have shown its efficiency to monitor grapevine growth but with many precautions (Drissi et al., 2009; Mazzetto et al., 2010) and its reliability for in-field acquisition (Kim et al., 2012; Kipp et al., 2014). In the past, different visible imaging systems have been tested and results were promising (Tregoat et al., 2001; Lloret et al., 2011; Diago et al., 2012; Rodríguez-Pulido et al., 2012). A review of the wine literature seems to demonstrate that multispectral imaging systems, which are mainly used for remote sensing applications, have been considered for the first time in 2012 as mobile proximal sensors to detect grape leave diseases in field conditions (Tirelli et al., 2012). Two main problems may prevent the use of mounted proximal multispectral imaging systems: the difficulty to control ambient light variation during the measurements and the determination of a reliable method for foliage identification. Indeed, natural light variations make it difficult to process images due to the presence of shadows, saturated pixels. Beyond the development of a highly multi-purpose proximal imaging system to monitor grapevine growth (leaf, vine shoots and wood pruning . . .), this article presents a new calibration method for visible and near-infrared images to consider ambient light variation. The following sections describe the experimental field and the development of the sensors mounted on a tracklaying tractor. A particular attention has been paid to the radiometric calibration which relies on a color chart that is visible in the images. For images with unusable chart, an extended correction is proposed. In the last section the NDVI results are evaluated against the NDVI values provided by the Greenseeker RT-100 system. The assessment of the performance of the radiometric calibration method is based on a statistical method (ANOVA) from both NDVI results (imaging system vs Greenseeker). 2. Materials and methods We first present the experimental field (Section 2.1), the experimental set-up and the associated technical constraints (Section 2.2). The system requires geometric (Section 2.3) and radiometric corrections to produce reflectance images (Section 2.4). Second, we detail the image processing chain for the estimation of NDVI (Section 2.5) and NDVI maps (Section 2.6). 2.1. Site description and data collection The experiment took place in the experimental site (Plumecoq) of the CIVC institute (Comité Interprofessionnel du Vin de Champagne) located in Chouilly, France. The surface area is 0.72 ha (150 m long by 50 m wide) and all grapevine varieties were planted in 1996 with spacing between two rows of 1.10 m. According to the topography and the orientation, three main grapevine varieties have been planted in a Latin Square pattern to minimize the influence of heterogeneous agronomic effects (Vilain, 2012): Pinot Noir (PN), Chardonnay (CH) and Meunier (Mn). Fig. 1 shows the organization of the plot and the varieties are planted over three blocks called later in the article ‘‘top”, ‘‘middle” and ‘‘down” according to the slope level of the hill (Fig. 1). Thus these three blocks, composed of the three Champagne varieties (PN, CH and Mn), represents nine sub-plots of 0.08 ha (50 m long by 16 m wide). The field is made of 45 rows divided in three blocks of 15 rows each according to the grapevine variety (CH, PN and Mn) and the slope level (top, middle, down). Data collection was performed over a single, cloudless day during July 2013 starting at

185

approximately 11:00 am, around solar noon to avoid shadows. The dataset is composed of 5356 images and it takes about one hour to cover the area. The phenological stage for all varieties is the early berry formation (BBCH stage N°71, Meier, 2001). An historical acquisition protocol was conducted: measurements were carried out from top to bottom of the field starting in the block of Chardonnay variety (CH Top) and ending in the block of Chardonnay variety (CH down). Moreover, in the past, the CIVC institute has developed a specific route protocol assuming that the exhaustive observation of 2 non-consecutive rows (continuous acquisition of images along the row) was representative enough of 5 row data. Consequently, 6 rows were acquired for each block of 15 rows. In the entire field, a total of 18 rows over 45 were then studied. 2.2. Experimental set-up Fig. 2 presents the acquisition system mounted on a tracklaying tractor assuming a constant tractor speed of 1 m/s. It consists of a visible-near-infrared multispectral camera (AD-130GE, JAI), a localisation system for data georeferencing (GPS-RTK, Trimble) at a frequency of 10 Hz and a computer to control the camera and to save the data. In order to validate the protocol developed to deduce the vegetation indices from the image acquisition, we also mounted a Greenseeker RT-100 (Trimble) which acquires data at a high frequency of 50 Hz. This active sensor is made of two LEDs (light emitting diodes) in the red (656 nm) and near-infrared (770 nm) spectral regions to estimate vine vigor using NDVI (Drissi et al., 2009). The Greenseeker field of view was 700 mm (height) by 20 mm (width). The multispectral camera is a prism-based 2-CCD progressive area scan camera, 8-bits encoding. This particular design captures simultaneously visible and near-infrared light spectrums using a single prism that splits the light gathered by the lens. Light is redirected to two CCD sensors (1296
966 pixels) to be recorded. It allows the synchronized acquisition of RGB (from 450 nm to 700 nm) and NIR (from 750 nm to 850 nm) images. Each CCD (4.86 mm
3.62 mm) is designed by Sony’s ICX447 CCD series. The image acquisition frequency is 3 Hz and it corresponds to around 3 images per meter. Aperture and exposure time are fixed by the operator before the dataset acquisition. Aperture is chosen according to the ambient light characteristics. Exposure time is set close to its minimal value (approximately 1/625 s) in order to avoid blurred image. Sensor’s gain (ISO speed) is controlled in real time by the laptop according to the luminance of the central region of the image (1/3 of the height and 1/3 of the width of the image) and it is adjusted in case of underexposure or overexposure. The short distance between vine rows (<1.10 m) implies that the camera is positioned at a short distance from vine foliage (<0.60 m). For that reason we used a wide angle lens (focal length of 2.8 mm) that produced 902 mm (height) by 670 mm (width) field of view. Fig. 3 shows an example of a snapshot image (RGB and NIR) of vine foliage: most of the leaves that are interesting in terms of foliage development are visible. A ColorChecker (‘‘MacBeth”) Chart was placed in the background to perform radiometric calibration (see Section 2.4). We used an umbrella (1) to avoid overexposure of parts of the image, and (2) to minor the presence of projected shadows on an image, which would degrade image quality and complicate the reliability of information extracted from it. In the same way, a black panel (Figs. 2 and 3) was installed in the background to focus measurement on grapevine and facilitate the image processing. All the height of images were resized according to the size of the black background panel (770 mm (h)
560 mm (w)) observed in the images. Indeed as the black background panel is fixed on the tractor, it is always visible at the same place on all the images of

186

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

Fig. 1. Presentation of Plumecoq vineyard composed of three grapevine varieties (CH: Chardonnay, PN: Pinot Noir and Mn: Meunier) delineated in the experimental vineyard.

separate visible from NIR ray light, geometrical distortions for both types of images are different. NIR and visible images could not overlap without this step of correction. Consequently to characterize vegetation areas we need to correct the images from geometric distortions produced by the wide angle lens and the prism of the imaging system. We use a classical algorithm developed by Bouguet (2007) for MATLAB Software and it is based on the work of Zhang (2000). It consists in determining intrinsic and extrinsic distortion parameters of the optical system. A checkerboard which is composed of 196 squares of 56 cm by 56 cm is taken as a reference. To correct the images, the algorithm uses the checkerboard images to calculate these distortion parameters. They are different for RGB and NIR images due to the presence of two CCD into the camera. 2.4. Protocol for radiometric calibration under natural daylight

Fig. 2. Multispectral camera, Greenseeker and GPS mounted on the track-laying tractor.

a dataset. The consequence is that the height of resized images was comparable to that of Greenseeker measurements. The recording of each data stream (Greenseeker, images and GPS) was tagged with date and time of the internal clock of the PC. For each image, a post-processing adjusted the GPS value and associated the Greenseeker value (average of 32 = 16
2 scans). 2.3. Protocol for geometrical corrections The optical device creates geometrical distortions on images: lengths, shapes and surfaces are modified, which will prevent the future calculation of morphometric indicators such as leaf area. Moreover, because the light flux is directed through a prism to

The radiometric calibration allows to compute vegetation indexes and to extract agronomic information on foliage. In field, in ambient light image acquisition also induces non-reproducing experimental conditions. To analyze and compare images we developed a reflectance imagery based on the use of a Macbeth color chart as a radiometric reference. In the case of remote and proximal sensing, literature reported very coarse radiometric calibration process: prior the field acquisition (Hruska et al., 2012; Del Pozo et al., 2014) a measurement of a calibrated reference panels (color, white or neutral grays) was acquired by the spectral imaging system. Then, images were radiometrically calibrated to confirm that the white panel used as ground control points represented 100% of the reflectance value with a digital number (DN) of 255. It can be considered as a rough approximation. Sometimes, during each image acquisition, a flux meter can record the natural light variations to optimize the calibration (Verger et al., 2014). Here, we present a very fine and adaptive radiometric calibration of multispectral imagery in in-field conditions allowing a specific correction for each image of the dataset. For images (images ‘‘without reference”) where the color chart is not visible an extended calibration will be needed. Let I be the set of all the images (5356): I = {IWR , IWOR } where IWR are the images ‘‘with reference” (40% of the dataset) and IWOR are the images ‘‘without reference” (60% of the dataset). The Fig. 4 presents the sampling points for both types of images. For this dataset of images, we can notice that the density of images in each plot is very different. In particular, we have noticed that the Pinot Noir seems to have a

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

187

Fig. 3. RGB (left) and NIR (right) raw images with the color chart provided by the multispectral camera.

Fig. 4. The left figure shows the spatial location of the three grapevine varieties (CH: Chardonnay, PN: Pinot Noir and Mn: Meunier); the middle figure presents the sampling points (2041 points) for images ‘‘with reference” (IWR ); the right figure displays the sampling points (3315 points) for images ‘‘without reference” (IWOR ).

set of IWR lower than the other grapevine varieties. It probably comes from a higher canopy density and the chart may be hidden by vine leaves.

conditions of illumination and observation, at a given wavelength (Meyzonnette and Lépine, 2003):

q¼ 2.4.1. Images with reference (IWR ) The mini ColorChecker Chart is used as a radiometric reference (Mansouri et al., 2005) to calibrate the images in reflectance. The reflectance (q) is defined as the ratio of the flux reflected (/reflected ) by an object to the incident one (/incident ) deduced from a perfectly white diffuse surface (i.e. patch) under the same

/reflected /incident

ð1Þ

The reflectance ranges between 0 and 1 and is a dimensionless quantity. Concerning the color chart, we first selected 6 patches (one black, four neutral grays, and one white) in order to mimic the illuminant. The high number of patches should prevent error measurement.

188

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

multispectral camera, and the corresponding RGB pixel values (Pascale, 2006). It is usually assumed that there is a linear relation between the measured (Xp with X = R, G, B channel) and the calibrated RGB values (Ycp with Y = R, G, B channel) of the pixels (Healey and Kondepudy, 1994; Barnard and Funt, 2002; Hadjit et al., 2013). We will make this assumption here. Let p be a pixel of IWR , its luminance values are Rp , Gp , and Bp in the red, green, and blue channels. If Rcp , Gcp and Bcp are the results of the radiometric calibration, we can calculate a slope aIWR and an intercept bIWR for each image with reference and each channel by linear regression between measured and theoretical values (Table 1) of the patches: R

Rcp ¼ aRIWR  Rp þ bIWR G

Gcp ¼ aGIWR  Gp þ bIWR Bcp

¼

aBIWR

 Bp þ

ð4Þ

B bIWR

We also hypothesized a linear regression for the NIR image calibration: NIR

NIRcp ¼ aNIR IWR  NIR p þ bIWR Fig. 5. The reflectance spectra obtained for the 6 patches using a FieldSpec 3 spectroradiometer. Vertical lines are associated to the central wavebands of the multispectral camera (450 nm, 550 nm, 600 nm and 800 nm).

The manufacturer only provides luminance values (sRGB: standard RGB color space) from 400 nm to 700 nm for all the patches. However, since the color chart is used both in the visible and nearinfrared regions, we need to determinate the reflectance of the 6 patches in the near-infrared domain. In laboratory, a hand-held spectroradiometer (FieldSpec3, ASD, Boulder, CO, USA) was used to measure the patch reflectance (Fig. 5). The value of reflectance spectra is quite constant on the wavebands of each of the camera channel (Fig. 5). So that the colorchart can be used as a radiometric reference both in the RGB and NIR spectral regions. The RGB values are graduated from 0 to 255. For each spectral channel (R: Red, G: Green, B: Blue and NIR: Near InfraRed) the pixel values are related to the reflectance (qR ; qG ; qB and qNIR ) by:

Rtheoretical ¼ 255  qR Gtheoretical ¼ 255  qG theoretical

B

ð2Þ

¼ 255  qB

Eq. (2) can be extended to the NIR as follows:

NIRtheoretical ¼ 255  qNIR

ð5Þ

where aRIWR ; aGIWR ; aBIWR and aNIR IWR are the slope coefficients of the linear regression for each spectral channel (R, G, B and NIR) of each image ‘‘with reference” (IWR ), and R

G

B

NIR

bIWR ; bIWR ; bIWR and bIWR are the intercept coefficients of the linear regression for each spectral channel (R, G, B and NIR) of each image ‘‘with reference” (IWR ). Fig. 6 shows the linear regression calculated for each spectral channel on all images ‘‘with reference” (IWR ) (where color chart is visible) of the dataset. Each point is associated to a neutral patch (from 1 to 6) and the experimental value corresponds to the average pixel value with its standard deviation read on the patch area for each spectral channel (R, G, B, NIR) of all images. For each channel, the white patch produces a too large dispersion of pixel values (see Section 3.1) that will affect the quality of the calibration of images ‘‘without reference” (IWR ). Therefore, the white patch will not be considered in the linear regression. For each channel, the coefficient of determination (R2) is higher than 0.98: it shows a strong correlation between theoretical and experimental values and the assessment of a linear radiometric calibration. In conclusion, for each image with reference (IWR ) we deduced the coefficients (aIWR , bIWR ) from the linear calibration providing a calibrated image in reflectance where calibrated pixel values are computing from Eqs. (4) and (5).

ð3Þ

Table 1 presents the reflectance values of the 6 patches at the central wavebands (450 nm, 550 nm, 600 nm, and 800 nm) of the

2.4.2. Images without reference (IWOR ) As abovementioned, ideally all the images should display a color chart but most of the time, it is hidden by leaves so that

Table 1 Laboratory measurement of reflectance values and RGB/NIR values for each neutral patches corresponding to theoretical values.

189

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

Fig. 6. Linear relation between experimental and theoretical values of the 6 neutral patches of the Macbeth ColorCheckerÒ for all images ‘‘with reference”. Horizontal error bars are the standard deviation associated to the experimental values.

we cannot use it to calibrate them radiometrically. To do it, we ^I ) of the image ^IWOR , b estimate the values of the coefficients (a WOR (IWOR ). We deduced them from the two nearest images ‘‘with reference” (IWR ), considering that the images have been acquired during a short temporal range (<3 s) along a row of a block assuming constant lighting conditions. We chose a simple determinist spatial interpolation method rather than a kriging method which estimates value according to the behavior of regionalized variable (Atkinson and Lloyd, 2010). We tested the linear Inverse Distance Weighting (IDW) predictor, which is the most common function in the literature (Arnaud and Emery, 2000; Lu and Wong, 2008). ^IWOR , In our case, the calculation of the interpolated coefficients (a ^I ) for an uncalibrated point is reduced to a linear interpolation b WOR

defined by:

Vj ¼ V1 þ

ðV 2  V 1 Þ
d1 ðd1 þ d2 Þ

ð6Þ

^IWOR (respecwhere V j are the interpolated values of the coefficient a

^I ) at an uncalibrated point (j) associated to the images tively b WOR ‘‘without reference” (IWOR ), V 1 and V 2 are the values of the coefficient aIWR (respectively bIWR ) at calibrated points (1 and 2) associated to the images ‘‘with reference” (IWR ), d1 and d2 are the distance between the uncalibrated point and the nearby measured point. Indeed during the field experiments, images were acquired along

a row and consequently for the interpolation we have restricted the search neighborhood to a vine row (Fig. 7). The assessment of the efficiency of this interpolation method ^I ) of images ‘‘with^I , b for the estimation of both coefficients (a WOR

WR

WR

for these images (IWR ) and for each spectral channel (R, G, B, IR) ^I ) to real ^IWR , b too. Finally, we compared estimated values (a WR (aIWR , bIWR ) ones to assess this linear interpolation method. In conclusion according to the situation, we directly apply a linear calibration on the images ‘‘with reference” or we perform an adaptive calibration based on a spatial linear interpolation method on images ‘‘without reference” to create in all cases new calibrated reflectance images. 2.5. Image processing for the determination of the NDVI Once the image calibration process is achieved, each image is rectified and calibrated in reflectance so that it contains the information on leaves only (spectral signature, morphometric parameters). Then we are able to compute vegetation indices such as NDVI (Normalized Difference Vegetation Index) (Rouse et al., 1973; Tucker, 1979):

NDVI ¼

Fig. 7. Schematic representation of the calculation of the interpolated value (V j ) of ^I ) for uncalibrated point (j). ^IWOR , b coefficients (a WOR

WOR

out reference” (IWOR ) is presented and discussed in the next section. It is based on a cross-validation method with the common type of Leave-One Out Cross-Validation (LOOCV) (Picard and Cook, 1984). First, the estimation procedure is done on images ‘‘without reference” (IWOR ) based on those of images ‘‘with reference” (IWR ). Then, we test the efficiency of the interpolation method only on a dataset of images ‘‘with reference” (IWR ) where aIWR and bIWR are well^I ), ^I , b known. We estimated these same coefficients, noted (a

qðNIRÞ  qðRÞ qðNIRÞ þ qðRÞ

ð7Þ

In remote sensing, NDVI ranges from 0 to 1. It is related to plant biomass and health state (Gamon et al., 1995; Bravo et al., 2004; Calcante et al., 2012): a NDVI close to zero means no vegetation or no photosynthesis activity whereas a value close to 1 indicates the high density of healthy leaves. According to the Eq. (7), the R and NIR channels of the reflectance images are selected to compute this vegetation index for each pixel to create NDVI images. However, an NIR image segmentation

190

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

Fig. 8. Flowchart of image segmentation.

algorithm (Fig. 8), based on Otsu’s method (Otsu, 1979), is designed to eliminate the background noise from the foliage for more accurate and robust estimation of vegetation vigor. It calculates the optimum threshold separating the two classes of pixels (vegetation and background) following bi-modal histogram. A binary image mask is created from the near-infrared images because it is more reliable to distinguish the leaves (NIR non-absorbent) and background (NIR absorbent) in the near-infrared images rather than in the visible images or reflectance images (Noh et al., 2005). Then this mask is applied on NDVI image to select vegetation pixels. For each NDVI image, we compute the average value NDVI. All images were postprocessed using the Image Processing Toolbox from MATLAB R2014b Software (MathWorks Inc., USA).

Greenseeker and a random factor ‘Vine row’. The reason for using this ANOVA is above all due to a row effect (difference between the downcoming and upgoing rows) that has been previously observed in the radiometric measurements and corrected using the appropriate calibration method. It is not considered as a major factor but may partly contribute to the total variability of the observed data (see next section). The ANOVA was applied for each of the 9 grapevine variety blocks. Each block is composed of 6 rows for which we compute average NDVI values. For each block the ANOVA datasets are composed of 6 average NDVI values by sensors (Greenseeker and Imagery). All statistical procedures were performed using the R statistical software (R Development Core Team, 2013).

2.6. Mapping the NDVI values

3. Results and discussion

Since the images are georeferenced, each average NDVI value can be mapped, providing a spatial representation of the variation of this index. We used geostatistics to generate such a map (Journel and Huijbregts, 1978). There are different methods depending on the stochastic properties of the random fields. The dataset is considered as a realization of a regionalized variable. The spatial structure of this variable is described by a model called variogram. The interpolation from measured locations to unmeasured locations is called kriging. It is a linear interpolation configured according to the variogram model, in order to minimize standard deviation values associated with errors of prediction (Papritz and Stein, 2002). The most common are the simple and ordinary kriging methods which assume a second order stationary for the regionalized variable. The first one assumes that the expectation of the variable is known and constant over the entire field, whereas the second one only admits that this expectation is constant on the local neighborhood of each estimation point. So we applied an ordinary kriging method (Isaaks and Srivastava, 1989). NDVI Maps were generated from the gridded interpolated data in ArcMap v10 (ESRI, USA). Two maps of average NDVI values provided from sensors, multispectral camera and Greenseeker, are presented and compared in the next section. Data analysis is performed using statistical analysis with a two-factor ANOVA without replication. We considered a fixed factor ‘Sensor’ composed of two modalities: image and

3.1. Optimization of the linear radiometric calibration on the (IWR ) images The measurements are performed in sunlight conditions and, as mentioned before, variations of light intensity are clearly observed in the dataset. These variations are illustrated in Fig. 9 through the measurement of the reflected light on neutral patches of the Macbeth color chart. For all the pale patches which the pixel intensity is stronger, we notice a wide Gaussian distribution of pixels values: it indicates a strong sensitivity to light. Moreover for the white patch, we noticed a large class of saturated pixels (value higher than 250). Conversely, for dark patches we observed a narrow distribution of pixel values. To explain this phenomenon regarding the wide distribution of pale patches we suggested two explanations: – Stronger is signal (i.e. pale patches) higher is the noise. Then the paler the patch is the more widely is the distribution as it is observed in the figure. – Considering pale patches with a strong signal, environmental effects are thus significant and they are observed on the tail of the Gaussian distribution. These environmental effects are relatively small and are considered as noise. They are probably due to lighting variations such as indirect lighting or the shaded leaves . . .

191

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

Table 2 ^I Þ) higher ^IWR Þ (and respectively Dðb Number of calibrated points (%) with an error Dða WR than 10%. ^IWR Þ > 10% (%) Dða

^ Þ > 10% (%) Dðb IWR

12.5 14.9 10.7 26.8

12.6 12.4 14.9 8.4

Red channel Green channel Blue channel NIR channel

Fig. 9. Dispersion of pixel values of each neutral patch of the Macbeth color chart patches read on all images (IWR ) of the dataset. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

This effective dispersion disturbs the estimation of the coefficients (aIWR , bIWR ) used to the linear regression (Fig. 6) and so it impacts the quality of the linear radiometric calibration. In Fig. 6 where we considered all neutral patches of the chart in the construction of the linear regression, we notice that these ones were particularly affected for the Blue, Red and the NIR channels. That is why, in order to optimize the radiometric calibration on (IWR ) images we have to find the best compromise between the number of patch required and the data dispersion of a patch. In view of these results (Figs. 6 and 9) and for dispelling any doubt we preclude the white patch in the construction of the linear calibration.

^I Þ > 10% represent 12.6%, 12.4% and 14.9% for the where Dðb WR red, green and blue channels, respectively. However, in the NIR channel, the errors seem to be more important for the slope than ^IWR Þ > 10% represents 26.8% of the for the intercept. The case Dða total number of points for the slope while it represents 8.4% for the intercept. ^I Þ values are in the same range as the Dða ^IWR Þ Although the Dðb WR values for errors higher than 10%, the errors higher than 25% seem to be more important in the case of the slope. Further studies on the interpolation and calibration methods should be explored to better understand these singular points. The linear regression (Eqs. (4) and (5)) might be improved by developing a complex spline interpolation, like in other applications. Compared to image calibrations conducted in the laboratory (Lathuiliere et al., 2005; Mansouri et al., 2005), we actually do not take into account factors such as the electronics of the optical devices (Mansouri et al., 2004) that were considered as minimal compared to the noise induced by outdoor conditions. Another way of optimization for next image acquisitions is about the acquisition system and it concerns: (1) the use of a wide diffuser panel instead of an umbrella (2) the size and the location of the color chart on the black background panel and (3) the test of the calibration with a gray scale chart rather than MacBeth color chart. 3.3. Comparison of NDVI maps

3.2. Extended calibration to the (IWOR ) using spatial interpolation and cross-validation The images ‘‘without reference” (IWOR ) are radiometrically corrected with a simple determinist spatial interpolation method. As ^I ) have ^IWOR , b mentioned in Section 2.4, the coefficients the (a WOR been estimated on images (IWOR ) using a simple linear interpolation method (Eq. (6)). However, we need to test the efficiency of the estimation method to guarantee the quality of the extended radiometric calibration. That is why we applied a cross-validation method (LOOCV) only considering the dataset of images (IWR ). The LOOCV allows testing and validating the estimation. Indeed, ^I ) parameters of the images ‘‘with^I , b the estimation of the (a WOR

WOR

^I ) of the ^IWR , b out reference” are performed with the estimated (a WR ^ ^I Þ and DðbI Þ are assessed image ‘‘with reference”. The errors Dða WR

WR

^IWR , by calculating the difference between the estimated values (a ^I ) and the values (aI , bI ) deduced from the linear regression. b WR WR WR We consider that an estimated value is aberrant when it deviated ^IWOR , from the ‘‘real” value by 10%. When we will estimate the (a ^ bI ), the aberrant estimated value of the image ‘‘with reference” WOR

will be not considerate and replace by the real value. Concerning the slope coefficient (aIWR ) of the linear regression, the number of sampling points where the error is higher than 10% represents 12.5%, 14.9% and 10.7% of the total number of points (2041) for the red, green, and blue channels respectively (Table 2). To find an explanation regarding the cause of the error, we done a spatial analysis using kriging method that suggests the errors are closely associated to a lack of reference data (Fig. 4). The cases

The radiometric calibration is validated comparing the NDVI maps obtained from imagery (radiometrically uncalibrated and calibrated images) to the Greenseeker values. Uncalibrated images mean that all images of the dataset have been geometrically corrected but not radiometrically calibrated. The maps of Fig. 10a–c were obtained using kriging method as detailed in Section 2.6. The first observation concerning the three NDVI maps: for each grapevine variety (PN, CH and Mn) and each block, the NDVI maps display similar patterns: the green color in Pinot Noir and Chardonnay grapevines indicates high vine vigor while the red color in Meunier grapevine indicates a lower vigor. These results are consistent with literature ones, as mentioned by Debuisson et al. (2010). For Fig. 10a and b, we notice that the kriged NDVI estimates for each block are quite similar (between 0 and 0.9). However, the Fig. 10c describes NDVI map based on uncalibrated images and it reveals globally lower values (<0.6) of NDVI. This result indicates that the radiometric calibration increases the NDVI values making them comparable to Greenseeker ones. So the radiometric calibration is an essential step to compare Greenseeker NDVI map and imagery NDVI map. The NDVI values provided by the two techniques are very similar (Table 3) whatever the grapevine varieties (CH, PN, Mn): 0.64 vs 0.62 for Chardonnay, 0.66 vs 0.65 for Pinot Noir and 0.45 vs 0.43 for Meunier. For each of the three grapevine varieties, there is no significant difference. To quantify this comparison, a statistical analysis of the variance (two-factor ANOVA without repetition) is done and is detailed in Table 4. We compute 9 (one per block) statistical tests of analysis of variance (ANOVA, a < 0.05) in the aim to compare the similarity of variance of data

192

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

Fig. 10. (a) NDVI map provided by Greenseeker data (b) NDVI map deduced from reflectance imagery (c) NDVI map deduced from uncalibrated images. (Grapevine variety: CH is Chardonnay, PN is Pinot noir and Mn is Meunier).

Table 3 Average of NDVI values for each grapevine variety (CH is Chardonnay, PN is Pinot noir and Mn is Meunier) and each of the 9 blocks obtained with Greenseeker and multi-spectral imaging systems. Dataset: beginning of berry formation Multi-Spectral Imaging system Row

Greenseeker

Block

Variety

Row

Block

Variety

0.60 0.66 0.69

0.61 0.59 0.65

0.62

0.58 0.62 0.67

0.63 0.67 0.71

0.61 0.65 0.70

0.65

0.43 0.48 0.45

0.47 0.40 0.51

0.40 0.44 0.46

0.43

CH Top Middle Down

0.64 0.73 0.73

0.64 0.64 0.66

0.65 0.75 0.74

0.63 0.34 0.61

0.61 0.62 0.68

0.59 0.62 0.69

0.63 0.62 0.69

0.64

0.57 0.61 0.65

0.62 0.66 0.66

0.67 0.69 0.69

0.64 0.29 0.62

0.56 0.61 0.61

PN Top Middle Down

0.63 0.65 0.72

0.58 0.64 0.75

0.69 0.67 0.68

0.59 0.67 0.70

0.60 0.64 0.74

0.63 0.69 0.68

0.62 0.66 0.71

0.66

0.62 0.63 0.70

0.59 0.70 0.71

0.63 0.62 0.68

0.63 0.64 0.72

Mn Top Middle Down

0.39 0.44 0.51

0.39 0.48 0.37

0.35 0.48 0.58

0.35 0.42 0.47

0.45 0.54 0.48

0.47 0.40 0.51

0.40 0.46 0.49

0.45

0.39 0.39 0.45

0.43 0.46 0.42

0.35 0.46 0.48

0.32 0.42 0.46

Table 4 Results of ‘Two-factor ANOVA without replication for the Data in Table 3. Two-factor ANOVA without replication Factor ‘Sensor’ F

p-value

CH Top Middle

1.237 1.569

0.316 0.265

Down

3.810

0.108

PN Top Middle Down

0.240 0.747 0.747

0.644 0.426 0.426

Factor ‘Vine row’

Fcritical (a = 5%)

6.61

F

p-value

2.366

0.183

24.379 3.334

0.001 0.106

2.807 1 0.785

0.140 0.5 0.601

Mn Top

0.028

0.871

18.618

0.003

Middle

5.952

0.058

Down

1.404

0.289

11.412 3.257

0.009 0.110

Fcritical (a = 5%)

5.05

issued from the two sensors and look at the row effect on the variability of the observed data. We now draw some conclusions from Table 4. Concerning the factor ‘Sensor’, no significant differences in NDVI measurements between both sensors were determined because for the 9 blocks the null hypothesis is accepted. Since the p-value (sensor) > 0.05 (=a), we can’t reject the Factor ‘Sensor’ null hypothesis, and so conclude (with 95% confidence) that there are no significant differences between the two sensors. Moreover for each of the 9 ANOVA, the calculated F-value is smaller than Fcritical confirming that all 9 tests are significant at a = 5%. If we look at the results of ANOVA about the Factor ‘Vine row’, we notice that the null hypothesis is rejected for 3 blocks: ‘‘CH_middle”, ‘‘Mn_middle” and ‘‘Mn_down” and the null hypothesis is accepted for the other cases. A possible explanation about the rejection of the null hypothesis is that for these three blocks there is an important external effect (agronomic, light, . . .) related to the row and this effect disturbs in the same way the NDVI values provided by the two sensors but it does not affect the results of the radiometric calibration.

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

Thus, these global results indicate that the estimation of the NDVI by the two sensors is consistent at the stage of the beginning of the berry formation. It validates our radiometric calibration method with regard to vine vigor when foliage is young. The choice of this growth stage is very important to assess the efficiency of the imaging system because the literature has demonstrated that the Greenseeker was highly correlated to foliage density at this stage and signals were not saturated (Drissi et al., 2009; Mazzetto et al., 2010). However, for the advanced growth stages the performance of the Greenseeker regarding the characterization of vine vigor is affected by it technical limits and so the correlation is less clear. That is why the multispectral imaging system and its development should be a promising technology regarding the literature (Diago et al., 2012). It will be more reliable than others for later stages of plant growth. This original methodology of radiometric calibration and these results are a first step in the assessment of multispectral imaging system to characterize vineyard foliage through the example of NDVI. 4. Conclusion This study has presented new radiometric calibration for the characterization of vineyard foliage using multispectral imagery (visible and NIR) as a proximal sensor mounted in a track-laying tractor. The radiometric calibration method allowed standardizing the reflectance of all images acquired in the field, even when the images are calibrated by spatial interpolation. Thus, changes in illumination during the acquisition can be controlled by this post-image processing. This procedure has enhanced the capability of reflectance imagery on assessing vineyard vigor when foliage is young (beginning of berry formation) by the calculation of a vegetation index. This study is a first validation of multispectral imagery to vineyard applications. The next step will be dedicated to the temporal monitoring of the vineyard over a growing season to assess the evolution of vine foliage. Thus we hope to demonstrate the limits of the two sensors depending of the vegetation growth. Afterwards, we should go further in the image processing looking at morphometric leaf parameters or the leaf area index (LAI). The challenge therefore will be to develop a strategy dedicated to the vine disease characterization. Acknowledgements The authors thank Christine FANT, colleague of AgroSup Dijon, for her comments and help with statistics. This project was funded by the Regional Council of Burgundy, and two technical institutes of Burgundy (Bureau Interprofessionnel des Vins de Bourgogne, BIVB) and of Champagne (Comité Interprofessionnel du Vin de Champagne, CIVC). M.A. Bourgeon gratefully acknowledges this financial support for the preparation of her PhD. The authors would also like to thank the anonymous reviewers, whose suggestions have substantially improved the quality of this article. References Arnaud, M., Emery, X., 2000. Estimation et interpolation spatiale : méthodes déterministes et méthodes géostatistiques. Hermès Science Publications. Atkinson, P.M., Lloyd, C.D., 2010. geoENV VII – Geostatistics for environmental applications. Proc 7th European Conference on Geostatistics for Environmental Applications, vol. 16. Springer Science & Business Media. http://dx.doi.org/ 10.1007/978-90-481-2322-3. Barnard, K., Funt, B., 2002. Camera characterization for color research. Color Res. Appl. 27 (3), 152–163. http://dx.doi.org/10.1002/col.10050. Bouguet, J.Y., 2007. Camera Calibration Toolbox For Matlab. (accessed May 21).

193

Bramley, R.G.V., 2001. Progress in the development of precision viticulture – variation in yield, quality and soil properties in contrasting Australian vineyards. In: Currie, L.D., Loganathan, P. (Eds.), Precision Tools for Improving Land Management. Fertilizer and Lime Research Centre, Massey University, Palmerston North: Occasional report N°14, pp. 25–43. Bramley, R.G.V., Lamb, D.W., 2003. Making sense of vineyard variability in Australia. In: Proc. International Symposium on Precision Viticulture. Ninth Latin American Congress on Viticulture and Oenology, pp. 35–54. Bramley, R.G.V., Proffitt, A.P.B., 1999. Managing variability in viticultural production. Grapegrower Winemaker 427 (1), 11–16. Bravo, C., Moshou, D., Oberti, R., West, J., McCartney, A., Bodria, L., Ramon, H., 2004. Foliar disease detection in the field using optical sensor fusion. In: Rosana and Moreira (Eds.), Agricultural Engineering International: The CIGR Journal of Scientific Research and Development. Texas A&M University: Manuscript FP 04 008. Vol VI. Calcante, A., Mena, A., Mazzetto, F., 2012. Evaluation of ‘ground sensing’ optical sensors for diagnosis of Plasmopara Viticola on vines. Spanish J. Agric. Res. 10 (3), 619. http://dx.doi.org/10.5424/sjar/2012103-619-11. Cerovic, Z.G., Goutouly, J.P., Hilbert, G., Destrac-Irvine, A., Martinon, V., Moise, N., 2009. Mapping winegrape quality attributes using portable fluorescence-based sensors. In: Best, S. (Ed.), Proc. 8th FRUTIC, Conception, Chile: Progap INIA, Chillian, Chile, pp. 301–310. Choudhury, B.J., 1987. Relationships between vegetation indices, radiation absorption, and net photosynthesis evaluated by a sensitivity analysis. Remote Sens. Environ. 22 (2), 209–233. http://dx.doi.org/10.1016/0034-4257 (87)90059-9. Debuisson, S., Germain, C., Garcia, O., Panigai, L., Moncomble, D., 2010. Using multiplex and greenseeker to manage spatial variation of vine vigour in Champagne. In: Proc. 10th International Conference on Precision Agriculture, Denver Co, USA, pp. 1–21. Diago, M.P., Correa, C., Millán, B., Barreiro, P., Valero, C., Tardaguila, J., 2012. Grapevine yield and leaf area estimation using supervised classification methodology on RGB images taken under field conditions. Sensors 12 (12), 16988–16996, http://www.mdpi.com/1424-8220/12/12/16988. Del Pozo, S., Rodríguez-Gonzálvez, P., Hernández-López, D., Felipe-García, B., 2014. Vicarious radiometric calibration of a multispectral camera on board an unmanned aerial system. Remote Sens. 6, 1918–1937. http://dx.doi.org/ 10.3390/rs6031918, 2014. Drissi, R., Goutouly, J.P., Forget, D., Gaudillere, J.P., 2009. Non-destructive measurement of grapevine leaf area by ground normalized difference vegetation index. Agron. J. 101 (1), 226–231. http://dx.doi.org/10.2134/ agronj2007.0167. Gamon, J.A., Field, C.B., Goulden, M.L., Griffin, K.L., Hartley, A.E., Joel, G., Penuelas, J., Valentinni, R., 1995. Relationships between NDVI, canopy structure, and photosynthesis in three californian vegetation types. Ecol. Appl. 5 (1), 28–41. http://dx.doi.org/10.2307/1942049. Gitelson, A.A., Merzlyak, M.N., 1996. Signature analysis of leaf reflectance spectra: algorithm development for remote sensing of chlorophyll. J. Plant Physiol. 148, 494–500. http://dx.doi.org/10.1016/s0176-1617(96)80284-7. Goutouly, J.P., Cerovic, Z.G., 2008. Capteurs optiques portables pour une viticulture de précision. Revue des Oenologues 129, 17–18. Hadjit, H., Oukebdane, A., Belbachir, A.H., 2013. Atmospheric correction of Earthobservation remote sensing images by Monte Carlo method. J. Earth Syst. Sci. 122 (5), 1219–1235. http://dx.doi.org/10.1007/s12040-013-0337-4. Hall, A., Lamb, D.W., Holzapfel, B., Louis, J., 2002. Optical remote sensing applications in viticulture – a review. Aust. J. Grape Wine Res. 8, 36–47. http://dx.doi.org/10.1111/j.1755-0238.2002.tb00209.x. Healey, G.E., Kondepudy, R., 1994. Radiometric CCD camera calibration and noise estimation. IEEE Trans. Pattern Anal. Mach. Intell. 16 (3), 267–276. http://dx.doi. org/10.1109/34.276126. Hruska, R., Mitchell, J., Anderson, M., Glenn, N.F., 2012. Radiometric and geometric analysis of hyperspectral imagery acquired from an unmanned aerial vehicle. Remote Sens. 4, 2736–2752. http://dx.doi.org/10.3390/rs4092736, 2012. Isaaks, E.H., Srivastava, R.M., 1989. An Introduction to Applied Geostatistics. Oxford University Press, New-York, http://dx.doi.org/10.1016/0098-3004(91) 90055-I. Johnson, L.F., Roczen, D.E., Youkhana, S.K., Nemani, R.R., Bosch, D.F., 2003. Mapping vineyard leaf area with multispectral satellite imagery. Comput. Electron. Agric. 38, 33–44. http://dx.doi.org/10.1016/S0168-1699(02)00106-0. Johnson, Lee F., 2003. Temporal stability of an NDVI-LAI relationship in a Napa Valley vineyard. Aust. J. Grape Wine Res. 9 (2), 96–101. http://dx.doi.org/ 10.1111/j.1755-0238.2003.tb00258.x. Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. Academic Press, London, 600pp. Kim, Y., Glenn, D.M., Park, J., Ngugi, H.K., Lehman, B.L., 2012. Characteristics of active spectral sensor for plant sensing. Trans. ASABE 55 (1), 293–301. http://dx.doi. org/10.13031/2013.41239. Kipp, S., Mistele, B., Schmidhalter, U., 2014. The performance of active spectral reflectance sensors as influenced by measuring distance, device temperature and light intensity. Comput. Electron. Agric. 100, 24–33. http://dx.doi.org/ 10.1016/j.compag.2013.10.007. Lamb, D., Hall, A., Louis, J., 2001. Airborne remote sensing of vines for canopy variability and productivity. Aust. New Zealand Grapegrower Winemaker 449, 89–92. Lathuiliere, A., Voisin, Y., Marzani, F.S., 2005. Une méthode pour le prétraitement et le calcul de matrices de crosstalk pour des images couleur issues d’une caméra

194

M.-A. Bourgeon et al. / Computers and Electronics in Agriculture 123 (2016) 184–194

tri-CCD. In: Proc. 20e Colloque Sur Le Traitement Du Signal et Des Images, GRETSI, Groupe d’Etudes du Traitement du Signal et des Images, pp. 1040–1043. Lloret, J., Bosch, I., Sendra, S., Serrano, A., 2011. A wireless sensor network for vineyard monitoring that uses image processing. Sensors 11 (6), 6165–6196, http://www.mdpi.com/1424-8220/11/6/6165. Lu, G.Y., Wong, D.W., 2008. An adaptive inverse-distance weighting spatial interpolation technique. Comput. Geosci. 34 (9), 1044–1055. http://dx.doi.org/ 10.1016/j.cageo.2007.07.010. Mansouri, A., Marzani, F.S., Gouton, P., 2005. Development of a protocol for CCD calibration: application to a multispectral imaging system. Int. J. Robot. Autom. 20 (2), 94–100. Mansouri, A., Marzani, F.S., Gouton, P., 2004. Systematic noise characterization of a CCD camera: application to a multipsectral imaging system. In: Proc. Complex Systems Intelligence and Modern Technological Applications (CSIMTA), Cherbourg, France, pp. 660–644. Mathews, A.J., 2013. Applying geospatial tools and techniques to viticulture. Geogr. Compass 7 (1), 22–34. http://dx.doi.org/10.1111/gec3.12018. Mathews, A.J., Jensen, J.L.R., 2013. Visualizing and quantifying vineyard canopy LAI using an Unmanned Aerial Vehicle (UAV) collected high density structure from motion point cloud. Remote Sens. 2013 (5), 2164–2183. http://dx.doi.org/ 10.3390/rs5052164. Mazzetto, F., Calcante, A., Mena, A., 2009. Comparing commercial optical sensors for crop monitoring tasks in precision viticulture. J. Agric. Eng. 40 (1), 11–18. http:// dx.doi.org/10.4081/jae.2009.1.11. Mazzetto, F., Calcante, A., Mena, A., Vercesi, A., 2010. Integration of optical and analogue sensors for monitoring canopy health and vigor in precision viticulture. Prec. Agric. J. 11 (6), 636–649. http://dx.doi.org/10.1007/s11119010-9186-1. McBratney, A., Whelan, B., Ancev, T., Bouma, J., 2005. Future directions of precision agriculture. Prec. Agric. J. 6 (1), 7–23. http://dx.doi.org/10.1007/s11119-0050681-8. Meier, U., 2001. Growth stages of mono-and dicotyledonous plants. BBCH Monographie. (G. F. B. R. C. for A. and Forestry, Ed.), 2nd ed. U. Meier. Meyzonnette, J.L., Lépine, T., 2003. Base de radiometrie optique, 2e édition. Cépaduès-Editions Paris, 256 pp. Moran, M.S., Inoue, Y., Barnes, E.M., 1997. Opportunities and limitations for imagebased remote sensing in precision crop management. Remote Sens. Environ. 61 (3), 319–346. http://dx.doi.org/10.1016/S0034-4257(97)00045-X. Mulla, D.J., 2013. Twenty five years of remote sensing in precision agriculture: key advances and remaining knowledge gaps. Biosyst. Eng. 114 (4), 358–371. http:// dx.doi.org/10.1016/j.biosystemseng.2012.08.009. Noh, H., Zhang, Q., Shin, B., Han, S., Reum, D., 2005. Dynamic calibration and image segmentation methods for multi-spectral imaging crop nitrogen deficiency sensors. Trans. ASAE 48 (1), 393–401. http://dx.doi.org/10.13031/2013.17933. Otsu, N., 1979. A threshold selection method from grey-levels histograms. IEEE Trans. Syst., Man, Cybernet. SMC-9 (1), 62–66. http://dx.doi.org/10.1109/ TSMC.1979.4310076 (1979). Papritz, A., Stein, A., 2002. Spatial prediction by linear kriging. In: Stein, A., Van der Meer, F., Gorte, B. (Eds.), Spatial Statistics for Remote Sensing, vol. 1. Springer, pp. 83–116. http://dx.doi.org/10.1007/0-306-47647-9_6. Pascale, D., 2006. RGB coordinates of the macbeth colorchecker. Technical report, The Babel Color Company, June 2006. .

Picard, R.R., Cook, R.D., 1984. Cross-validation of regression models. J. Am. Stat. Assoc. 79 (387), 575–583. http://dx.doi.org/10.1080/01621459.1984.10478083. Pierce, F.J., Nowak, P., 1999. Aspects of precision agriculture. Adv. Agron. 67, 1–85. http://dx.doi.org/10.1016/S0065-2113(08)60513-1. Quemada, M., Gabriel, J., Zarco-Tejada, P., 2014. Airborne hyperspectral images and ground-level optical sensors as assessment tools for maize nitrogen fertilization. Remote Sens. 6 (4), 2940–2962, http://www.mdpi.com/20724292/6/4/2940. Robert, P., Rust, R., Larson, W., 1994. Site-specific management for agricultural systems. In: Proc. 2nd International Conference on Precision Agriculture, Madison, WI. ASA/CSSA/SSSA. Rodríguez-Pulido, F.J., Gómez-Robledo, L., Melgosa, B., Gordillo, M., González-Miret, M.L., Heredia, F.J., 2012. Ripeness estimation of grape berries and seeds by image analysis. Comput. Electron. Agric. 82, 128–133. http://dx.doi.org/ 10.1016/j.compag.2012.01.004. Rouse, J.W., Hass, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems in the great plains with ERTS. In: Proc. 3rd Earth Resources Technology Satellite (ERTS) Symposium, Washington, DC, USA: NASA SP-351, NASA, Vol 1, pp. 309–17. Sankaran, S., Mishra, A., Ehsani, R., Davis, C., 2010. A review of advanced techniques for detecting plant diseases. Comput. Electron. Agric. 72 (1), 1–13. http://dx.doi. org/10.1016/j.compag.2010.02.007. Solari, F., Shanahan, J., Ferguson, R.B., Schepers, J.S., Gitelson, A.A., 2008. Active sensor reflectance measurements of corn nitrogen status and yield potential. Agron. J. 100, 571–579. http://dx.doi.org/10.2134/agronj2007.0244. Strever, A.E., 2007. Remote sensing as a tool for viticulture research in South-Africa with specific reference to terroir studies. Acta Horticult. 754, 393–399, http:// www.actahort.org/books/754/754_52.htm. Tirelli, P., Marchi, M., Calcante, A., Vitalini, S., Iriti, M., Borghese, N.A., Oberti, R., 2012. Multispectral image analysis for grapevine diseases automatic detection in field conditions. In: Proc. International Conference of Agricultural Engineering CIGR-AgEng, pp. 1–6. Tisseyre, B., Ojeda, H., Taylor, J., 2007. New technologies and methodologies for sitespecific viticulture. Journal International des Sciences de la Vigne et du Vin 41 (2), 63–76. Tregoat, O., Grenier, G., Ollat, N., 2001. Etude comparative de la précision et de la rapidité de mise en oeuvre de différentes méthodes d’estimation de la surface foliaire de la vigne. Journal International des Sciences de la Vigne et du Vin 35 (1), 31–39. Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8 (2), 127–150. http://dx.doi. org/10.1016/0034-4257(79)90013-0. Verger, A., Vigneau, N., Cheron, C., Baret, F., 2014. Green area index from an unmanned aerial system over wheat and rapeseed crops. Remote Sens. Environ. http://dx.doi.org/10.1016/j.rse.2014.06.006. Vilain, M., 2012. Méthodes expérimentales en agronomie: pratique et analyse. Tec & Doc. Lavoisier, p. 424. Whalley, J., Shanmuganathan, S., 2013. Applications of image processing in viticulture: a review. In: Piantadosi, J., Anderssen, R.S., Boland, J. (Eds.), Proc. 20th International Congress on Modelling and Simulation, December 2013, pp. 531–537. . Zhang, Z., 2000. A flexible new technique for camera calibration. IEEE Trans. Pattern Anal. Mach. Intell. 22 (11), 1330–1334. http://dx.doi.org/10.1109/34.888718.