Response of Population Canopy Color Gradation Skewed Distribution Parameters of the RGB Model to Micrometeorology Environment in Begonia Fimbristipula Hance

Abstract

The high quality and efficient production of greenhouse vegetation depend on micrometeorology environmental adjusting such as system warming and illumination supplement. In order to improve the quantity, quality, and efficiency of greenhouse vegetation, it is necessary to figure out the relationship between the crop growth conditions and environmental meteorological factors, which could give constructive suggestions for precise control of the greenhouse environment and reduce the running costs. The parameters from the color information of the plant canopy reflect the internal physiological conditions, thus, the RGB model has been widely used in the color analysis of digital pictures of leaves. We take photographs of Begonia Fimbristipula Hance (BFH) growing in the greenhouse at a fixed time every day and measure the meteorological factors. The results showed that the color scale for the single leaf, single plant, and the populated canopy of the BFH photographs all have skewed cumulative distribution histograms. The color gradation skewness-distribution (CGSD) parameters of the RGB model were increased from 4 to 20 after the skewness analysis, which greatly expanded the canopy leaf color information and could simultaneously describe the depth and distribution characteristics of the canopy color. The 20 CGSD parameters were sensitive to the micrometeorology factors, especially to the radiation and temperature accumulation. The multiple regression models of mean, median, mode, and kurtosis parameters to microclimate factors were established, and the spatial models of skewness parameters were optimized. The models can well explain the response of canopy color to microclimate factors and can be used to monitor the variation of plant canopy color under different micrometeorology.

1. Introduction

Meteorological conditions are crucial factors that determine crop growth and leaf development [1,2]. Facility agriculture is an efficient crop production through the precise regulating of microclimate within the greenhouse [3,4,5]. Clarifying the relationships between plant physiological and ecological indicators and meteorological conditions are the basic method for quantitative assessment of agro-meteorological conditions, assessing the agro-meteorological disaster losses, predicting the yield, and applying the intelligent micrometeorology control of agricultural production. Thus, it is the right path to find the determining parameters that have a close relationship with meteorological factors and reflect plant growth status sensitively.

Digital image technology is well-developed and high-resolution camera equipment is widely used, with digital imaging technology having the obvious advantages of high resolution and low cost in the research of plant phenotypes [6,7]. A digital color image is not only convenient to obtain [8], but also contains abundant information about plant morphology, structure, and color gradation [9,10,11], which can reflect the internal status of plants. The RGB color model is the popularly used color representation for digital images [12]. The parameters of the traditional RGB model are mainly the mean values of each color channel and their combined values, which can approximately describe the color depth of the leaves and estimate the chlorophyll content of the leaves [13,14,15,16,17,18]. Moreover, the RGB model reflects the soil moisture and nutrient levels [19,20,21,22]. However, the mean values of image parameters are not comprehensive enough in the description of color distribution [23], as a result, the RGB model is generally used in the studies of water, fertility, and disease which may drastically change the leaf color [8,24,25,26]. The RGB model is also an effective method to reflect the relationship between the image information and the variation of environmental factors.

The quantitative analysis of mega-high quality digital photographs could explain the genotype-by-environment interactions in plant phenotypes [27]. The variation of meteorological factors affects the physiological and biochemical characters of plants slightly, such as influencing the openness of the stomata [28], changing the arrangement of the chloroplast [29], and increasing or decreasing the photosynthesis rate of the leaves [30,31]. All these affections change the plant’s color under a visible spectrum [32]. Therefore, the image color information parameters are used as representative references to reflect the meteorological environment of the growing plant. Skewed analysis expanded the RGB model parameters from the mean to the mean, median, mode, skewness, and kurtosis, with the number of parameters from 4 to 20 [33]. The color gradation skewness-distribution (CGSD) parameters describe leaf color information more accurately and comprehensively than the traditional mean parameters [33], including whether the color gradation distribution of a single plant varies the same as the leaf, and the field scale determines its sensitivity in response to the environmental meteorological factors. The CGSD for all scales varying consistently helps to construct an effective model between the external meteorological factors and morphological parameters of the plant. The quantitative evaluation of agro-meteorological conditions for a plant can be achieved by combining canopy color information indexes corresponding to different morphological parameters of plants, to provide information that supports precise management of plant production.

Begonia Fimbristipula Hance (BFH) is a vegetation type widely used for medicine and food. BFH has abundant anthocyanins, polyphenols, and flavonoids and a strong antioxidant capacity [34,35,36]. BFH is cultivated in facilities that have a warm and humid environment to ensure annual cultivation [37]. After planting BFH in a glass greenhouse and collecting the meteorological factors and canopy photographs, the relationship between changes in canopy color and the meteorological factors in the facility environment was analyzed. The purpose of this research is: (1) to quantify the mathematical relationship between CGSD parameters and meteorological factors, and increase the accuracy of quantifying the changes in population canopy color with meteorological changes; (2) to construct population canopy color–meteorological response models by using the relationship of CGSD parameters and meteorological factors, including temperature, humidity, and solar radiation, and; (3) to provide new methods from the RGB model to monitor the patterns of canopy color variations due to meteorological environment changes. The results may construct the crop growth model based on the CGSD parameters of the RGB model, and provide a new approach for intelligent control of facility agricultural production.

2. Materials and Methods

2.1. Plant Material and Growth Conditions

The glass greenhouse for the experiment is in Guli Modern Agriculture Demonstration Park, Jiangning District, Nanjing, Jiangsu, China (24°57′ N, 116°30′ E). The BFH were potted by the rooting of cuttings and planted with a total nutrient matrix which was in the proportion of N: P₂O₅: K₂O in a 100 g matrix at 20:10:20, trace elements included Cu, Fe, Zn, Mn, B, and Mo ≥ 1%, and pH 5.85. Moreover, the matrix contained an active accelerator and sustained release agent of poly-fertilizer. The BFH experiment had 2 replicates, and each replicate contained 24 potted plants. The first growing period of BFH was from 1 December 2018 to 17 February 2019, and after cutting off the 10 cm top length on 18 February, BFH continued growing in the second growth period from 19 February 2019 to 11 April 2019.

2.2. Meteorological Data Acquisition

Meteorological data were obtained from the agricultural meteorological observation station (DZZ4, Jiangsu Provincial Radio Science Research Institute Co., LTD., Wuxi, China) in the facility. The observed meteorological elements included daily meteorological elements, such as the daily mean temperature (T_dm), daily mean relative humidity (RH_dm), daily mean ground temperature (TG_dm), daily mean soil temperature at 10 cm (TS_dm−10c), daily mean vapor pressure (VP_dm), daily mean dew point temperature (TD_dm), daily total global radiation (GR_dt), and daily total photosynthetically active radiation (PAR_dt). The accumulated temperature (AT), accumulated global radiation (AGR), and accumulated photosynthetically active radiation (APAR) were calculated from T_dm, GR_dt, and PAR_dt.

$$\mathrm{AT}={\displaystyle \sum {\mathrm{T}}_{\mathrm{dm}}}$$

(1)

$$\mathrm{AGR}={\displaystyle \sum {\mathrm{GR}}_{\mathrm{dt}}}$$

(2)

$$\mathrm{APAR}={\displaystyle \sum {\mathrm{PAR}}_{\mathrm{dt}}}$$

(3)

2.3. Canopy Image Collection

Canopy color images of BFH were collected in the facility by a monitoring camera (DH-SD-65F630U-HN-Q, Zhejiang Dahua Technology Co., Ltd., Hangzhou, China) with an image resolution of 1920 × 1080, and the installation height was 280 cm (Figure 1). The camera with a vertical lens was set with a fixed focal length shooting and automatic white balance and adopted a fixed time shooting mode to take a photograph at set times of 9:03 a.m. every day. The 117 images from 3 groups without direct sunlight were selected for analysis, including 102 images in the first growth period (1 December 2018 to 17 February 2019) of 2 replicates with 51 pictures each, and 15 images in the second growth period (19 February 2019 to 11 April 2019).

Figure 1. Schematic map of microclimate observation and digital camera in the glass greenhouse.

2.4. Cutting and Denoising of the Image

Adobe Photoshop V8.0 software (San Jose, CA, USA) and MATLAB2016R software (referred to as MATLAB [Math Works, Natick, MA, USA]) were primarily used to cut and denoise the BFH original image [38,39] (Figure 2a).

Figure 2. Flow map of Cutting and Denoising of the Image.

• Adobe Photoshop CS software (San Jose, CA, USA) was primarily used to intercept the range of 600 × 600 in the lower-left corner of the image, and the processed image was saved in a JPG image format (Figure 2b).
• The rgb2hsv function of MATLAB was used to convert RGB images into HSV images. Double cycle operation was used to set the H-value of the image background to 0 (i.e., black), while the H value of the plant remains unchanged. The hsv2rgb function was used to convert the processed HSV image into an RGB image (Figure 2c).
• The double cycle operation of MATLAB was used again to filter the color threshold value of the image processed in the previous step. The color opacity of the black part of the image was adjusted to 0 (that is, completely transparent), and the color image of the target leaf or canopy with the transparent background was saved as a PNG image mode (Figure 2d).

2.5. Information Collection of the RGB Image

Transformation of double-precision arrays of the image. MATLAB was used to extract and analyze BFH RGB images. After reading color images by using the imread function, image (:, :, 1), image (:, :, 2), image (:, :, 3), and the rgb2gray function were respectively used to read every pixel gradation of red, green and blue channels, as well as gray-level images. The full circulation algorithm was used to retrieve and record the non-black pixel index codes of these pixels, which were combined with an array of leaf color gradation. Then the double function was used to transform it into double precision arrays again.

Establishment of the canopy color gradation skewness-distribution (CGSD) parameters table. The mean, median, mode, std, skewness and kurtosis functions were used to acquire the mean, median, mode, standard Deviation, skewness, and kurtosis of the double-precision arrays of red, green, and blue channels, as well as gray-level images [33]. The CGSD parameters were obtained, including R_Mean, R_Median, R_Mode, R_Skewness, R_Kurtosis, G_Mean, G_Median, G_Mode, G_Skewness, G_Kurtosis, B_Mean, B_Median, B_Mode, B_Skewness, B_Kurtosis, Y_Mean, Y_Median, Y_Mode, Y_Skewness, and Y_Kurtosis (Table 1). Finally, the CGSD parameters tables of color gradation distribution of BFH RGB images were formed.

Table 1. Color gradation skewness-distribution (CGSD) parameters.

Channel	CGSD Parameters
	Mean	Median	Mode	Skewness	Kurtosis
Red Channel (R)	RMean	RMedian	RMode	RSkewness	RKurtosis
Green Channel (G)	GMean	GMedian	GMode	GSkewness	GKurtosis
Blue Channel (B)	BMean	BMedian	BMode	BSkewness	BKurtosis
Gray Image (Y)	YMean	YMedian	YMode	YSkewness	YKurtosis

Array distribution normality testing. The lillietest and jbtest functions were used to conduct the Lilliefors and Jarque-Bera tests of normal distribution for the color gradation distribution of red, green, and blue channels, as well as gray-level images of BFH RGB images.

2.6. Prediction Model Construction and Goodness-of-Fit Detection

Correlation Analysis of 20 CGSD Parameters to Microclimate Factors. The Cor package of R was used to analyze the relationship between 20 CGSD parameters of population canopy in RGB images of BFH and the corresponding microclimate factors, including daily meteorological factors (T_dm, RH_dm, TG_dm, TS_dm−10c, VP_dm, TD_dm, GR_dt, and PAR_dt) and photothermal accumulation factors (AT, AGR, and APAR), which were collected for significant examination with a double tail inspection.

Linear prediction models establishment. By using R, the linear prediction models (Y1-Y20) of 20 CGSD parameters of population canopy in RGB images of BFH were established by a regression approach based on the least-square method, with meteorological factors as the independent variables. The probability of F-to-enter of the models was set to 0.050 or less, while the probability of F-to-remove was set to 0.100 or more. Decision coefficient optimization, significance test of regression model and regression coefficient, and collinearity diagnosis of independent variables of the regression model were conducted for the alternative regression models in turn [40], and the optimal regression models were finally determined.

Spatial polynomial model establishment. Two photothermal accumulation factors with the correlation coefficient of highest absolute value with CGSD parameters were selected as the independent variable, and the spatial polynomial models (Z1–Z4) of skewness of red, green, and blue channels, as well as gray-level images, were established by using polynomial fitting in the Curve Fitting Tool of R.

3. Results

3.1. Skew Analysis of the Distribution of Leaf Color Gradation of the RGB Images

After analyzing the cumulative distribution of color gradation of the 117 selected images of BFH, we found that all the red, green, and blue channels, and the gray-level for single leaf, single plant, and population canopy images showed a left-skewed distribution (Figure 3). Lilliefors and Jarque-Bera normality tests also indicated that color gradation data did not distribute normally (Supplementary Table S1). Twenty color gradient skewness-distribution (CGSD) parameters were obtained including the mean, median, mode, skewness, and kurtosis for the red, green, and blue channels, as well as the gray-level image, for each canopy image, respectively. These parameters can describe not only the depth of canopy color but also its distribution.

Figure 3. RGB model color gradation distribution of single leaf, plant canopy, and population canopy images of Begonia Fimbristipula Hance (BFH). The cumulative frequency histogram of red, green, and blue channels, as well as gray-level images, were drawn using the imhist function of MATLAB. The X-axis is the color gradation value, and the Y-axis is the cumulative frequency.

The main distribution of color gradation is positive skewness for RGB channels and the gray-level image near the value of 0. This increment of positive skewness as the canopy develops reflected the leaf color transit from light green to dark green (Figure 4).

Figure 4. RGB model color gradation distribution of population canopy images of BFH on 3 different dates. The cumulative frequency histogram of red, green, and blue channels, as well as gray-level images, were drawn using the imhist function of MATLAB. The X-axis is the color gradation value, and the Y-axis is the cumulative frequency.

3.2. Correlation Analysis Microclimate Factors in Glasshouse and Population Canopy CGSD Parameters

The relationship between 20 CGSD parameters and the microclimate factors including daily meteorological factors (T_dm, RH_dm, TG_dm, TS_dm−10c, VP_dm, TD_dm, GR_dt, and PAR_dt) and accumulative factors (AT, AGR, and APAR) are shown in Table 2. Table 2 indicates that temperature, humidity, ground temperature, soil temperature at 10 cm, and solar radiation were significantly correlated with the CGSD parameters. The accumulated temperature and solar radiation (AT, AGR, and APAR) had an extremely significant correlation with the 20 CGSD parameters, with the correlation coefficients higher than 0.8. The parameters of mean, median, and mode, which represented color depth, showed a significant positive correlation with the meteorological factors, while the parameters of skewness and kurtosis representing color uniformity showed an obviously negative correlation.

Table 2. Spearman correlation coefficient of population canopy CGSD parameters of BFH and microclimate factors.

	Tdm	RHdm	TGdm	TSdm−10c	VPdm	TDdm	GRdt	PARdt	AT	AGR	APAR
RMean	−0.249	0.606 **	0.278 *	0.629 **	−0.025	−0.029	−0.24	−0.251	−0.865 **	−0.856 **	−0.860 **
RMedian	−0.253	0.605 **	0.280 *	0.641 **	−0.024	−0.033	−0.227	−0.235	−0.843 **	−0.836 **	−0.840 **
RMode	−0.257	0.584 **	0.315 *	0.690 **	−0.041	−0.045	−0.247	−0.251	−0.815 **	−0.813 **	−0.816 **
RSkewness	0.225	−0.586 **	−0.292 *	−0.641 **	0.003	0.01	0.214	0.224	0.853 **	0.846 **	0.850 **
RKurtosis	0.142	−0.503 **	−0.265	−0.519 **	−0.054	−0.047	0.113	0.131	0.792 **	0.774 **	0.779 **
GMean	−0.282 *	0.633 **	0.236	0.588 **	−0.067	−0.053	−0.344 *	−0.358 **	−0.950 **	−0.947 **	−0.949 **
GMedian	−0.289 *	0.641 **	0.246	0.613 **	−0.068	−0.057	−0.343 *	−0.355 *	−0.931 **	−0.929 **	−0.931 **
GMode	−0.295 *	0.607 **	0.25	0.623 **	−0.088	−0.079	−0.326 *	−0.334 *	−0.901 **	−0.901 **	−0.902 **
GSkewness	0.231	−0.598 **	−0.307 *	−0.668 **	0.008	0.01	0.257	0.268	0.866 **	0.864 **	0.867 **
GKurtosis	0.157	−0.507 **	−0.261	−0.526 **	−0.031	−0.033	0.176	0.197	0.855 **	0.844 **	0.848 **
BMean	−0.332 *	0.625 **	0.235	0.632 **	−0.105	−0.112	−0.284 *	−0.293 *	−0.794 **	−0.789 **	−0.793 **
BMedian	−0.317 *	0.619 **	0.252	0.654 **	−0.088	−0.097	−0.272	−0.278 *	−0.790 **	−0.787 **	−0.790 **
BMode	−0.263	0.577 **	0.321 *	0.716 **	−0.047	−0.054	−0.292 *	−0.298 *	−0.750 **	−0.751 **	−0.754 **
BSkewness	0.243	−0.588 **	−0.294 *	−0.656 **	0.023	0.028	0.242	0.252	0.864 **	0.860 **	0.863 **
BKurtosis	0.205	−0.517 **	−0.238	−0.527 **	0.012	0.017	0.164	0.181	0.821 **	0.808 **	0.812 **
YMean	−0.281 *	0.632 **	0.25	0.610 **	−0.06	−0.053	−0.313 *	−0.327 *	−0.922 **	−0.917 **	−0.920 **
YMedian	−0.283 *	0.630 **	0.258	0.630 **	−0.059	−0.056	−0.304 *	−0.314 *	−0.899 **	−0.896 **	−0.898 **
YMode	−0.271	0.608 **	0.285 *	0.665 **	−0.056	−0.051	−0.316 *	−0.323 *	−0.884 **	−0.882 **	−0.884 **
YSkewness	0.229	−0.593 **	−0.305 *	−0.661 **	0.005	0.01	0.239	0.25	0.858 **	0.854 **	0.857 **
YKurtosis	0.159	−0.508 **	−0.259	−0.524 **	−0.033	−0.031	0.154	0.173	0.832 **	0.818 **	0.822 **

Note: A Spearman correlation analysis of SPSS software was used on 20 CGSD parameters (R_Mean, R_Median, R_Mode, R_Skewness, R_Kurtosis, G_Mean, G_Median, G_Mode, G_Skewness, G_Kurtosis, B_Mean, B_Median, B_Mode, B_Skewness, B_Kurtosis, Y_Mean, Y_Median, Y_Mode, Y_Skewness, and Y_Kurtosis) and the microclimate factors including daily meteorological factors (T_dm, RH_dm, TG_dm, TS_dm−10c, VP_dm, TD_dm, GR_dt, and PAR_dt) and photothermal accumulation factors (AT, AGR, and APAR) (n = 51). ** indicates significant correlation according to a two-tailed test (p < 0.01), * indicates significant correlation according to a two-tailed test (p < 0.05), the same as below.

3.3. Multiple Linear Relationships of the Microclimate Factors and Population Canopy CGSD Parameters

Multiple linear relationships between population canopy CGSD parameters and microclimate factors were established by the stepwise regression method using the ordinary least square method (OLS) (Table 3). Table 3 showed that all the multivariate determination coefficients (R²) of the population canopy color-meteorological response models equation were greater than 0.7, indicating that the equation well explained the response of BFH population canopy CGSD parameters to microclimate factors. The R² of the red and green channels, as well as the gray-level image, was 0.692–0.954, which was better than those of the blue channel with 0.674–0.791. The model between the mean of green channel and accumulated temperature, daily mean relative humidity, and the daily total photosynthetically active radiation had the highest R² of 0.954. Moreover, all the models were significant (p < 0.01).

Table 3. Population canopy color-meteorological response models of BFH and their goodness of fit.

Model		R-Square	Adjusted R-Square	RMSE	F Value	Significance F
RMean	Y1 = 11.556 − 0.017x1 + 0.524x2 + 0.549x3	0.848	0.839	2.456	87.634	0.000
RMedian	Y2 = − 28.395 − 0.015x1 + 0.629x2 + 0.703x3 + 1.871x4	0.868	0.857	2.630	75.860	0.000
RMode	Y3 = 14.284 − 0.024x5 + 2.596x4	0.746	0.735	3.668	70.458	0.000
RSkewness	Y4 = 2.100 + 0.001x1 − 0.199x4 + 0.107x6	0.803	0.791	0.142	63.968	0.000
RKurtosis	Y5 = 5.371 − 0.010 x1 − 0.024x7 − 0.028 x2	0.706	0.688	0.382	37.703	0.000
GMean	Y6 = 27.424 − 0.022x1 + 0.464x2 + 0.349x3	0.954	0.951	1.609	327.337	0.000
GMedian	Y7 = 34.402 − 0.036x5 + 0.366 x2	0.906	0.902	2.548	230.359	0.000
GMode	Y8 = 30.294 − 0.037x5 + 0.342 x2	0.845	0.839	3.456	130.821	0.000
GSkewness	Y9 = 1.501 + 0.030x5 − 0.052x7 − 0.013 x2	0.831	0.820	0.138	76.900	0.000
GKurtosis	Y10 = 2.824 + 0.002x1	0.732	0.726	0.310	133.539	0.000
BMean	Y11 = 30.430 − 0.013x1 + 0.341x2	0.703	0.691	3.136	56.787	0.000
BMedian	Y12 = −6.307 − 0.012x1 + 0.380x2 + 2.094 x4	0.752	0.736	3.470	47.483	0.000
BMode	Y13 = 4.464 − 0.020 x5 + 3.267x4	0.695	0.683	4.028	54.781	0.000
BSkewness	Y14 = 1.716 + 0.001x1 − 0.083x4	0.791	0.782	0.145	90.164	0.000
BKurtosis	Y15 = 2.874 + 0.002x1	0.674	0.668	0.346	101.377	0.000
YMean	Y16 = 38.927 − 0.019x1 + 0.307x2	0.892	0.888	2.240	198.372	0.000
YMedian	Y17 = 30.819 − 0.021x1 + 0.368x2	0.855	0.848	2.956	140.973	0.000
YMode	Y18 = 26.665 − 0.032 x5 + 2.144 x4	0.825	0.818	3.440	113.213	0.000
YSkewness	Y19 = 1.818 + 0.001x1 − 0.092 x4	0.784	0.775	0.156	87.302	0.000
YKurtosis	Y20 = 2.901 + 0.002x1	0.692	0.686	0.355	110.029	0.000

Note: x₁ is the accumulated temperature, x₂ is the daily mean relative humidity, x₃ is the daily total photosynthetically active radiation, x₄ is the daily mean soil temperature at 10 cm, x₅ is the accumulated photosynthetically active radiation, x₆ is the daily minimum ground temperature, x₇ is the accumulated global radiation.

All the models were verified by between-group and outside-group samples (Table 4 and Figure 5, Figure 6 and Figure 7), which indicated that the population canopy color fitting models performed well in predicting the mean, median, mode, and kurtosis. The accuracy of prediction of the skewness was about 80%, while other CGSD parameters exceeded 90%, with the mean of the blue channel, as well as the gray-level image exceeding 95% (Table 4) in the same modeling group samples. Similarly, the accuracy of prediction of the mean, median, and mode of RGB channels and the gray-level image was 90%, the kurtosis was about 75%, and the skewness was 53–58% in the between-group prediction. The accuracy of prediction of other CGSD parameters was 80% with the exception of the skewness and kurtosis of the red channel, the mode, and skewness of the green channel, and the median of the gray-level image in the outside-group. Moreover, the accuracy of prediction of the kurtosis of the green and blue channels, and the gray-level image exceeded 85%.

Table 4. Analytical results of prediction accuracy of population canopy color-meteorological response models.

Model		Modeling Group			Between-Group Prediction			Outside-Group Prediction			Average
		Prediction Sample Data	Eliminate Abnormal Data	Predictive Accuracy	Prediction Sample Data	Eliminate Abnormal Data	Predictive Accuracy	Prediction Sample Data	Eliminate Abnormal Data	Predictive Accuracy
RMean	Y1	51	0	96.27%	51	0	92.89%	15	0	90.92%	94.11%
RMedian	Y2	51	0	96.02%	51	0	91.98%	15	0	84.90%	92.84%
RMode	Y3	51	0	92.89%	51	0	92.47%	15	0	87.87%	92.06%
RSkewness	Y4	51	0	80.98%	51	10	53.21%	15	3	70.30%	68.80%
RKurtosis	Y5	51	0	90.48%	51	0	72.53%	15	0	73.56%	80.49%
GMean	Y6	51	0	97.98%	51	0	93.55%	15	0	93.71%	95.50%
GMedian	Y7	51	0	96.29%	51	0	91.97%	15	0	80.08%	92.33%
GMode	Y8	51	0	94.14%	51	0	92.38%	15	0	77.33%	91.22%
GSkewness	Y9	51	1	76.29%	51	9	52.40%	15	7	35.12%	63.73%
GKurtosis	Y10	51	0	92.98%	51	0	76.64%	15	0	86.75%	85.06%
BMean	Y11	51	0	95.46%	51	0	92.05%	15	0	84.66%	92.59%
BMedian	Y12	51	0	94.53%	51	0	90.57%	15	0	85.42%	91.64%
BMode	Y13	51	0	92.66%	51	0	92.65%	15	0	84.46%	91.60%
BSkewness	Y14	51	1	82.21%	51	8	56.11%	15	3	82.39%	72.36%
BKurtosis	Y15	51	0	92.17%	51	0	74.21%	15	0	84.12%	83.31%
YMean	Y16	51	0	96.95%	51	0	93.33%	15	0	89.60%	94.43%
YMedian	Y17	51	0	90.26%	51	0	93.64%	15	0	67.68%	88.84%
YMode	Y18	51	0	94.11%	51	0	93.15%	15	0	87.27%	92.82%
YSkewness	Y19	51	2	80.43%	51	10	58.03%	15	3	84.64%	73.81%
YKurtosis	Y20	51	0	92.56%	51	0	76.04%	15	0	86.53%	84.59%

Note: Predictive accuracy = (1 − |predictive value-measured value|/measured) × 100%.

Figure 5. Prediction accuracy interval analysis of the population canopy color-meteorological response models of the modeling group. The Boxplot function of MATLAB was used to draw the accuracy percentile distribution diagram of the population canopy CGSD parameters prediction model. The blue box in the figure represents the prediction accuracy of the model distributed between 25% and 75%. The shorter the box, the better the degree of data concentration. The red line in the box represents the median distribution of prediction accuracy, and the dotted line at both ends outside the box represents the endpoint value (maximum and minimum) of the prediction data distribution. The shorter the distance between the two endpoint values, the better the degree of data concentration. The red cross outside the box represents abnormal data. The fewer abnormal data, the better the degree of data distribution concentration.

Figure 6. Prediction accuracy interval analysis of the population canopy color-meteorological response models of the between-group. The Boxplot function of MATLAB was used to draw the accuracy percentile distribution diagram of the population canopy CGSD parameters prediction model. The blue box in the figure represents the prediction accuracy of the model distributed between 25% and 75%. The shorter the box, the better the degree of data concentration. The red line in the box represents the median distribution of prediction accuracy, and the dotted line at both ends outside the box represents the endpoint value (maximum and minimum) of the prediction data distribution. The shorter the distance between the two endpoint values, the better the degree of data concentration. The red cross outside the box represents abnormal data. The fewer abnormal data, the better the degree of data distribution concentration.

Figure 7. Prediction accuracy interval analysis of the population canopy color-meteorological response models of the outside-group. The Boxplot function of MATLAB was used to draw the accuracy percentile distribution diagram of the population canopy CGSD parameters prediction model. The blue box in the figure represents the prediction accuracy of the model distributed between 25% and 75%. The shorter the box, the better the degree of data concentration. The red line in the box represents the median distribution of prediction accuracy, and the dotted line at both ends outside the box represents the endpoint value (maximum and minimum) of the prediction data distribution. The shorter the distance between the two endpoint values, the better the degree of data concentration. The red cross outside the box represents abnormal data. The fewer abnormal data, the better the degree of data distribution concentration.

3.4. Optimization of Canopy Color Skewness-Meteorological Models

Although the skewness of the red and green channels, and the gray-level image performed well, the between-group and outside-group prediction accuracy were generally lower than other CGSD parameters. The skewness was non-linear to microclimate factors [33]. Therefore, we fitted the polynomial Z1-Z4 that incorporates spatial bidirectional patterns for the RGB channels, and the gray-level image, by taking two accumulation factors, i.e., the accumulated temperature (x₁) and the accumulated photosynthetically active radiation (x₅) as the independent variables, which had the highest absolute value of the correlation coefficient.

Z1 = 0.3629 + 0.6382 × 10⁻³ x₁ + 2.626 × 10⁻³ x₅ + 5.312 × 10⁻⁶ x₁² − 1.072 × 10⁻⁵ x₁ x₅

(4)

Z2 = 0.2242 + 0.5163 × 10⁻³ x₁ + 3.11 × 10⁻³ x₅ + 3.08 × 10⁻⁶ x₁² − 7.463 × 10⁻⁶ x₁ x₅

(5)

Z3 = 0.3215 + 0.6477 × 10⁻³ x₁ + 2.673 × 10⁻³ x₅ + 3.93 × 10⁻⁶ x₁² − 8.614 × 10⁻⁶ x₁ x₅

(6)

Z4 = 0.2786 + 0.5116 × 10⁻³ x₁ + 3.114 × 10⁻³ x₅ + 4.244 × 10⁻⁶ x₁² − 9.285 × 10⁻⁶ x₁ x₅

(7)

After comparing the multiple stepwise regression and spatial bidirectional models using the skewness, the (R²) of the spatial bidirectional models was higher than the multiple stepwise regression models with lower RMSE. This indicated that the nonlinear models had better fitting results, or the skewness of the population canopy image of BFH has a close relationship to the microclimate in the glass greenhouse (Table 5).

Table 5. Comparison of multiple stepwise regression models and spatial bidirectional models for the skewness.

Parameters	Model		R-Square	Adjusted R-Square	RMSE	Significant F
RSkewness	Y4		0.803	0.791	0.142	0.000
	Z1	Equation (4)	0.884	0.874	0.111	0.000
GSkewness	Y9		0.831	0.820	0.138	0.000
	Z2	Equation (5)	0.902	0.893	0.107	0.000
BSkewness	Y14		0.791	0.782	0.145	0.000
	Z3	Equation (6)	0.900	0.891	0.102	0.000
YSkewness	Y19		0.784	0.775	0.156	0.000
	Z4	Equation (7)	0.894	0.884	0.111	0.000

The spatial bidirectional models for the skewness have 5.7 percentage points higher accuracy in prediction, and the number of outliers significantly decreased. Compared to the regression model, the average accuracy of the spatial bidirectional models was improved by 7.63, 5.75, and 6.25 percentage points in the modeling group, the between-group, and the outside-group, respectively. The RGB channels and gray-level image were all improved by 6.90, 10.87, 3.47, and 1.72 percentage points, respectively (Table 6).

Table 6. Comparison of the prediction accuracy of the multiple stepwise regression models and spatial bidirectional models for the skewness.

Parameters	Model	Modeling Group			Between-Group			Outside-Group			Average
		Prediction Sample Data	Eliminate Abnormal Data	Predictive Accuracy	Prediction Sample Data	Eliminate Abnormal Data	Predictive Accuracy	Prediction Sample Data	Eliminate Abnormal Data	Predictive Accuracy
RSkewness	Y4	51	0	80.98%	51	10	53.21%	15	3	70.30%	68.80%
	Z1	51	0	90.02%	51	1	60.56%	15	3	75.58%	77.76%
GSkewness	Y9	51	1	76.29%	51	9	52.40%	15	7	35.12%	63.73%
	Z2	51	0	88.31%	51	1	60.82%	15	3	73.71%	74.60%
BSkewness	Y14	51	1	82.21%	51	8	56.11%	15	3	82.39%	72.36%
	Z3	51	0	90.16%	51	1	61.27%	15	3	75.61%	75.83%
YSkewness	Y19	51	2	80.43%	51	10	58.03%	15	3	84.64%	73.81%
	Z4	51	0	89.05%	51	1	60.10%	15	3	82.34%	75.53%

4. Discussion

The high quality and efficient production of greenhouse plants depend on the environmental regulation focusing on the plants’ growth in the facility. The changing physiological parameters result from the variation of temperature, radiation, and humidity rapidly and instantaneously [41,42,43,44]. Ineffective control of environmental factors will restrict the improvement of glass greenhouse yield and crop quality [45]. Therefore, coupling the relationship between the plant growth state indicated by color parameters and controlling the meteorological factors will help to promote the high quality and efficient production in facility plants, realize the precise control of the facility environment, and decrease the costs.

Digital color images’ information can reflect the growing state of plants [11], which provides a convenient way of plant growth monitoring. Research on conducting qualitative and quantitative descriptions of phenotypic traits for plant appearance by using digital imaging technology has increased in smart agriculture [6]. The RGB color model has been widely used to process digital color images to study chlorophyll content and its related plant nutritional status and stress response [24,46,47]. The RGB model and leaf color parameters can describe changes in tobacco leaf color depth and homogeneity [33]. Now, we found that images of the RGB channels and the gray-level images all showed a skewed distribution at the single leaf, plant, and population canopy images scale in BFH.

Environmental changes lead to changes in the color of plants’ leaves [21,26,48,49,50]. Our work indicated that temperature, humidity, ground temperature, soil temperature at 10 cm, and solar radiation in glass greenhouses were significantly correlated with multiple CGSD parameters of canopy images in BFH. Moreover, the accumulative value of temperature and solar radiation (AT, AGR, and APAR) had an extremely significant correlation with 20 CGSD parameters, with the correlation coefficients beyond 0.8. The mean, median, and mode parameters representing color depth have a positive relationship with the accumulative parameters, while the skewness and kurtosis parameters representing the color distribution showed a significantly negative correlation with the accumulative factors. The deep canopy color of BFH images has higher accumulative factors depending on the analysis of the peak changes of the cumulative distribution diagram of canopy color and provides a new approach to quantitative plant growing conditions.

The multi-dimensionality of CGSD parameters enables them to have a basis for building relationships with complex meteorological factors. Canopy color–meteorological responding models have high prediction accuracy after calculating the image mean, median, mode, and kurtosis parameters and skewness distribution using the spatial bidirectional models. The change of accumulative factors is generally linear, while skewness is calculated from a higher-order function. Although the skewness was significantly related to microclimate factors, it cannot be simply described and fitted by linear fitting [33]. Compared to the multiple stepwise regression models, the spatial bidirectional models had better fitness and higher accuracy.

The digital color image has been applied to automatic observation in agrometeorology [51,52], and is proposed as a new technical approach that can be combined with crop growth models [53], which agriculture meteorological services will automatically and intelligently apply in future services [54,55]. The relationship between the CGSD parameters of the population canopy and the environment meteorological factors makes it possible to use the canopy color information to construct a plant growth model from digital images, which can provide new methods for quantitatively evaluating agricultural meteorological conditions, assessing agricultural meteorological disaster loss, predicting yield and quality, and supporting precise management of crop production.

5. Conclusions

Analyzing the canopy images’ color histograms skewed distribution of BFH in the glasshouse, we found that 20 population canopy CGSD parameters were sensitive to the microclimate factors, especially the radiation and temperature accumulation factors. The canopy color–meteorological responding models showed that the CGSD parameters have a close relationship with microclimate factors, which expands the potential application of the RGB model in monitoring the variations of plant or leaf color. The CGSD parameters can describe the canopy color information comprehensively and accurately and are considered a new indicator for quantitative evaluation of plant growth conditions. These parameters help construct a plant growth condition model and provide a new method for agrometeorological assessment and prediction, and, ultimately, support the precise management of crop production.