1 Introduction

Green leaves are important for the functioning of terrestrial plants for harvesting solar light for food production [1], by virtue of containing pigment systems as well as molecular machinery that help plants garner net primary productivity, allows gas exchange for carbon capture, and evapotranspiration. The chlorophyll, carotenoid and anthocyanin are the three most important pigments in leaves [2], with chlorophyll mainly responsible for green color and energy capture [3]. Among the pigment systems, chlorophyll is arguably the most important biological molecule and seat of the most indispensable biological process viz., photosynthesis. In fact, the chlorophyll content of leaves is often used to predict the overall physiological condition of the leaves, as influenced by various natural and anthropogenic factors [4, 5].

Leaf chlorophyll status is a measure of the photosynthetic rate and biomass production potential in crop plants and is invariably used as an effective surrogate for elucidating plant response to stress [6]. In order to harness the potential of this trait in genotypic selection especially under stress conditions, the chlorophyll estimation should be reliable and high throughput especially in situations when the germplasm set to be evaluated is large sample and scale of experiment is big [7]. The estimation of leaf chlorophyll content is based on the destructive solvent extraction of chlorophyll from leaves followed by spectrophotometric determination using absorbances [8, 9]. Equipments such as spectrophotometer, a fluorometer, or a high-performance liquid chromatography (HPLC) are often used to measure light absorptions at a various range of wavelength [10, 11], which are then used to determine leaf chlorophyll. The methods used for chlorophyll extraction in plants are almost always based on destructive extraction from leaf tissues using organic solvents such as acetone, ethanol, dimethyl sulfoxide or N-dimethyl formamide [10,11,12,13]. However, these methods are time-consuming and require sophisticated equipment for chemical digestion and other analytical procedures [14]. Such in vitro determinations are destructive, expensive, and time-consuming, and may therefore not be applicable for all purposes. The lab-based approach is invariably costly, labor intensive and time consuming, and does not provide for tracking the temporal dynamics of Leaf chlorophyll of the same leaves [15].

Spectroscopic and colorimetric methods are used in both field, greenhouse or growth chamber based experiments and offer rapid and non-destructive estimation of leaf chlorophyll in living leaves in situ and effectively reduce the need of a large number of sample collection in field, are high throughput and cost-effective and provide for the characterisation of a large number of samples [16]. These methods are based on the optical properties of leaves and estimate chlorophyll on the basis of relative reflectance and/or absorbance of radiation by chlorophyll. In 1980’s, the Soil Plant Analysis Development (SPAD, Minolta Camera Co., Osaka, Japan) developed a hand-held absorbance-based dual wavelength chlorophyll meter (SPAD models 501 and, later, 502) that has been extensively used in plant physiology experiments as a rapid and non-destructive approach for measurement of chlorophyll content in the field. The SPAD-502 m measures the transmittance of red (650 nm) and infrared (940 nm) radiation through the leaf, and estimates relative chlorophyll index by a value that corresponds to the quantity of chlorophyll in that leaf sample [17].

Using SPAD meter is a relatively efficient method to estimate the leaf chlorophyll content and, in most studies, SPAD values have shown a strong positive correlation (0.78 to 0.93) with the leaf chlorophyll content [18,19,20,21], and is a non-destructive method, but the cost of the device is very high (2000–3000 US$). However, as the research is moving fast towards artificial intelligence (AI) and machine learning (ML) based approaches [22], there is greater integration of AI and ML based strategies for tackling complex problem of agricultural research using high-end computing. The computer automation has significantly helped in enhancing reliability and effectiveness of complex research problems [23].

Very recently, the advances in Android and IOS based platforms have enabled researchers to use processed smartphone images using open-source apps and softwares for indirect estimation of water requirements of crops as well as and photosynthetic efficiency with fair accuracy [24]. The smart phone based colorimetric analysis, in addition to physiological parameters, have also been used to determine several mineral and biochemical parameters including soil phosphorus [25], and leaf chlorophyll content [26]. The imaging technology has, more specifically, emerged as a viable strategy for real-time estimation of chlorophyll content [27]. In case of rice, Wang et al. [14] showed that digital color image analysis could be a simple method of assessing N status under natural light conditions for different cultivars and different developmental stages. Recent developments in programming, AI and ML have also opened up avenues for assessing crop biophysical and biochemical parameters using Machine Learning Regression Algorithms from image analysis and spectral data [28].

The colour appearance of objects can be quantified using three-dimensional colour spaces (Fig. 1). The three dimensions of a colour space are typically lightness, colourfulness, and hue. Lightness can be described as the brightness of a surface, and colourfulness indicates the strength of the chromatic content (redness of a red object). The International Commission on Illumination (CIE) describes hue as the “attribute of a visual perception according to which an area appears to be similar to one of the colours: red, yellow, green, and blue”, which is categorically different from lightness and colourfulness for inexperienced observers. Humans are typically more sensitive to hue shifts than changes in saturation and lightness [29]. Although hue can be calculated using colour appearance models (CAMs) quantifying the perceptual attributes of coloured stimuli, it has not been so widely used to quantify object colours in research. Hue angle (Δh) and hue (ΔH) differences can be calculated in two of most widely used colour spaces: CIE L*a*b* and CAM02 [30].

Fig. 1
figure 1

CIE L*a*b* colour space depicting colour perception

Full size image

A large number of studies have substantiated the use of colorimetric analysis for non-destructive and high throughput chlorophyll estimation [16, 31, 32]. The present study was based on the premise that non-destructive methods of leaf chlorophyll quantification could be reliably standardised for high throughput estimation especially in large diversity panels and could save time and labour associated with traditional acetone/ethanol-based extraction and quantification methods. Our main objective was to test and validate the use of a fast and cheap non-destructive method for estimation of chlorophyll using a handheld colorimeter based on CIE L*a*b* and δE parameters in rice. Since non-destructive imaging technologies are being increasingly used in almost all phenotyping systems including both physical and biochemical parameters with sufficient experimental evidence of reliability and reproducibility, the present method could be conveniently mainstreamed in phenotyping large germplasm sets with ease, economy and reliability of prediction. Therefore, the present study was aimed at testing and validating the use of a fast and cheap non-destructive method for estimation of chlorophyll using a handheld colorimeter based on CIE L*a*b* parameters in rice.

2 Materials and methods

2.1 Plant material

A set of 264 temperate rice diversity panel genotypes were used for the present study. The panel comprised of rice lines for diversity of grain shape (long slender indica and short bold japonica types), size (medium to large), colour (brown, black, purple and red rice) and adaptability ranges (low to mid to high altitudes of Kashmir valley) and includes landraces, released varieties and germplasm accessions including check varieties Jehlum, Shalimar Rice-1, Shalimar Rice-2, Shalimar Rice-3, Shalimar Rice-4 and Shalimar Rice-5 from temperate agroecological niches. All these varieties have been released by SKUAST-Kashmir during 2000 to 2020 with broad adaptability ranges.

2.2 Experimental setup

A controlled environment greenhouse experiment was conducted in the greenhouse of the Department of Plant Breeding & Genetics, SKUAST-Kashmir using the standard procedure given by Shashidhar et al. [33]. All 264 genotypes were evaluated for their root traits in a using a column culture phenotyping system comprising of polyvinyl chloride (PVC) columns of 120 cm height and 12 cm internal diameter. The seeds were surface sterilized for 1 min with 70% alcohol and 2% sodium hypochlorite (NaClO) followed by rinsing with deionized water in 10 × 10 cm petri plates. The sterilized seeds were kept for germination and the pregerminated seeds were transplanted into the PVC columns when the radicle was about 5 mm long. The growing medium was made by mixing soil and sand mixture in the ratio of a 1:1 and fertilised with Osmocote (19:6:12 N:P2O5:K2O), a slow-release fertiliser @ 2 g per column before sowing and evenly mixed with the top 2 cm growing medium. Four seeds were initially planted in each column, but after the establishment of plants, only two plants were retained in each column. Plants were maintained under optimum temperature. For imposing the drought treatment, plants were maintained under 80% field capacity from sowing to the four-leaf stage; then stress was imposed by withholding water for 4 weeks development stage (48 days after sowing; DAS), with the total drought stress duration of 28 days. At the end of the experiment, the soil moisture content in the drought treatment was 30%, as quantified on weight basis [34].

3 Colorimetric estimation (L*, a* and b*) using hand held colorimeter

A low-cost colorimeter (VEYKOLOR PRO MINI COLRIMETER, < 100 US$) was used to measure the leaf parameters for developing a low-cost translation method of estimating leaf chlorophyll content (Fig. 2). The colorimeter was able to take measurements with fair precision (δ E = 0.05). It can take measurements in terms of CIE L*a*b* (Lightness, greenness-redness, blueness-yellowness), CIE RGB (Red, green and blue) and CIE Lch (which uses the polar coordinates C* viz., chroma or relative saturation and h° viz., hue angle instead of the Cartesian coordinates a* and b* in L*a*b* scheme. However, lightness L* of CIE Lab remains unchanged). It has a high-definition camera, small aperture of 4 mm, comes with automatic calibration and can be conveniently used even at young stages on small leaves as well as under stress conditions when leaves are invariably rolled and measurements cannot be conveniently taken by SPAD. The measurements were taken on fully expanded uppermost leaves at five positions (as chlorophyll is not evenly distributed along the leaf blade) and averaged for all genotypes replicated twice. It is very small in size with dimensions of 7.99 × 2.76 × 2.6 inches that makes it very handy. Moreover, the data can be synchronized the instrument to the cloud and by personal color library in the app (COLORSPEC). It has a full spectrum light source and can take 7000 scans in one charging, and can be easily used with android, IOS and Windows.

Fig. 2
figure 2

Experimental set up for drought response of rice diversity panel and measurement of L*a*b* readings and chlorophyll from rice leaves

Full size image

In the CIE L*a*b*, the L* coordinate in CIE L*a*b* closely matches human perception of lightness, a* and b* dimensions represent the visual perception of red-green and yellow-blue chroma, respectively [35]. Both a* and b* are independent with image lightness (L*), and can have both negative and positive values (+ a* reds,—a* greens, + b* yellows,—b* blues). The three coordinates of L*a*b* are computed from the tristimulus values X, Y and Z as following Eqs. (14):

$${\text{L}}^{*} = 116{\text{ f }}\left( {{\text{Y}}/{\text{Yn}}} \right){ }{-}{ }16{\text{ L}}^{*}$$
$${\text{a}}^{*} = 500{ }\left[ {{\text{f }}\left( {{\text{X}}/{\text{Xn}}} \right){ }{-}{\text{ f }}\left( {{\text{Y}}/{\text{Yn}}} \right)} \right]$$
$${\text{b}}^{*} = 200{ }\left[ {{\text{f }}\left( {{\text{Y}}/{\text{Yn}}} \right){ }{-}{\text{ f }}\left( {{\text{Z}}/{\text{Zn}}} \right)} \right]$$

where X n, Y n and Z n describe a specified white object-color stimulus.

Unlike the RGB and CMYK color models, L*a*b* is designed to approximate human vision. The L* component closely matches human perception of lightness, though it does not take the Helmholtz–Kohlrausch effect (a phenomenon wherein the intense saturation of a spectral hue is wrongly perceived as part of the luminance of the colour) into account. CIE L*a*b*, even though, less uniform in the color axes, is highly useful for predicting small differences in color. In L*a*b* approach of estimating color differences, the particular advantage is the estimate δE (a measure of the disparity between two colors used for visual discernibility), that quantifies the differences between colours in terms of sum of squared deviations of differences between L*, a* and b* values of two colours. In terms of chlorophyll content variation, a major goal of characterising large germplasm sets is to assess the natural variation between genotypes for chlorophyll so that the lines could be effectively classified. The δ E can be effectively used to identify differences in the individuals based on three axes L*a* and b*.

3.1 Chlorophyll estimation in laboratory

Two samples were drawn from the diversity panel viz., calibration sample and validation sample. The chlorophyll of calibration as well as validation sample was estimated by acetone-ethanol (2:1 v/v) method given by Arnon in 1960’s [36]. Both calibration and validation samples were drawn based on diverse values of a* (-ive values as -a* corresponds to greenness). One gram of fresh leaf was macerated in a mortar and pestle with 4 mL of 99% acetone and centrifuged. The samples were left to stand for 30 min in the freezer in the dark, and centrifuged for 10 min at 2000 rpm. Absorbance readings were performed at wavelengths of 663 nm and 645 nm. The control was acetone/ethanol (2:1 v/v). The total chlorophyll was measured as total of chlorophyll a and chlorophyll b the below formula [37].

$${\text{Chlorophyll a }}\left( {{\text{mg}}/{\text{g}}} \right){ } = { }\left( {12.7^{*} {\text{ A}}663} \right){ } - { }\left( {2.69^{*} {\text{ A}}645} \right)$$
$${\text{Chlorophyll b }}\left( {{\text{mg}}/{\text{g}}} \right){ } = { }\left( {22.9^{*} {\text{ A}}663} \right){ } - { }\left( {4.68^{*} {\text{ A}}645} \right)$$

where A663 and A645 are the absorbance measured from 663 and 645 nm, respectively. The spectrophotometer was adjusted to zero using the acetone/ethanol mixture.

3.2 Fitting the model

We used four parameters to fit the model namely L* (Lightness or darkness with value of 0–100 with 0 meaning black and 100 meaning white), a* (Greenness or redness with value of -128 to + 128, with -a* meaning greenness and + a* meaning redness), b* (Blue or yellowness with value of -128 to + 128, with -b* meaning blue and + b* meaning yellowness) and δ E (that is sum of squared deviations of the L*a*b* values of sample from a reference chlorophyll sample of known concentration). We regressed the estimated chlorophyll with L*, a*, b*, and δ E separately and also did multiple linear regression of estimated chlorophyll with all the variables together.

Based on the regression equation of estimated chlorophyll of calibration sample with L*, a*, b*, and δ E we also validated the model using the validation sample and also tested the validation using regression between predicted chlorophyll based on model and the estimated chlorophyll using acetone method. The model fitting parameters such as probability and root mean square error (RMSE) were used to test the validity of model.

4 Results and discussion

4.1 Color and chlorophyll content traits of rice diversity panel leaves

The L, a*, b* for the diversity panel ranged from 13.4 to 54.4, − 12.4 to 4.40 and 12.9 to 36.9 and CV of 10.08%, 22.09% and 35.15% respectively. In calibration sample, L, a*, b* ranged from 12.20 to 44.10, − 13.30 to − 4.40 and 1.20 to 33.70 and CV of 36.34, 47.37 and 45.68% respectively. Similarly for validation sample, L, a*, b* ranged from 30.00 to 41.40, − 11.50 to − 11.30 and 17.20 to 31.40 and CV of 8.55, 0.77 and 14.71% respectively. The leaf chlorophyll ranged from 28.17 to 54.07 mg/g (Mean 50.21 and CV 5.73%) under irrigated conditions and 25.96 to 49.55 mg/g (Mean 41.93 and CV 7.60%) (Table 1). Li et al. [38] has reported similar results in case of Sassafras tzumu, where almost similar results were observed in calibration and validation samples using spectroscopic parameters L, a*, b* for chlorophyll estimation.

Table 1 Descriptive statistics of L, a* and b* for diversity panel, calibration and validation samples of rice under water stress
Full size table

4.2 Model prediction

The leaf chlorophyll content and four different parameters of leaf color, i.e. L*, a*, b* and δ E, were considered as single prediction model separately as well as multiple prediction analysis. Among the single prediction models, the regression coefficients of estimated chlorophyll with L*, a*, b*, and δ E separately and R2 values of 0.671, 0.552, 0.498 and 0.436 for a*, b*, δ E and L* respectively (Table 2, Fig. 3). Despite the values being acceptable (> 0.4 is desirable to establish the validity of the model), we performed a multiple regression of all parameters was done with estimated chlorophyll in order to check whether the combined regression of all the parameters cam improve the prediction process (Fig. 4). The model was better fit than individual analysis as it resulted in R2 value of 0.78. Following regression equations were obtained for L*, a*, b*, and δ E. Leon et al. [39] used a Minolta chromameter and estimated L*, a*, b* and hue angle and found correlation of 0.68 between SPAD and L* readings and concluded that non-destructive chlorophyll estimation by chromameter parameters especially L* and hue angle (Ho) can be effective approaches for chlorophyll estimation as against acetone/ethanol method. The R2 value was slightly higher (Fig. 4) in calibration sample (0.78) as compared to validation sample (0.724). Yang et al. [40] also reported slightly different R2 in verification sample and calibration sample in chlorophyll estimation using L*a*b* color space parameters. This could also be due to the fact that the calibration and validation samples are often not fully representative of the larger group from which they are drawn, as well as unintended subjectivity in sampling the calibration and validation samples.

Table 2 Regression equations for chlorophyll estimation using L*, a*, b*, and δ E
Full size table
Fig. 3
figure 3

Regression graphs of chlorophyl with L*, a*, b* and δ E in calibration sample

Full size image
Fig. 4
figure 4

Multiple regression graphs of chlorophyl with L*, a*, b* and δ E in calibration and validation samples

Full size image

The comparative analysis of mean, range and CV (%) for L*, a*, b* and δ E values under irrigated and drought conditions (Table 3) revealed that mean of L*, a*, b* and δ E were higher under irrigated as compared to drought conditions, however the absolute value of a* was higher under irrigated as compared to drought conditions (negative sign). The CV values that indicate the spread of values across mean of the population were higher for L* and δ E (De) under irrigated as compared to drought, whereas reverse trend was observed for a* and b*. Since the a* signifies greenness, there is greater spread as genotypes show differential response to leaf yellowing under drought stress. Similar trend is depicted by Fig. 5, that shows greater spread of a* under drought as compared to irrigated conditions. The deviated points under irrigated conditions represent genotypes that have darker leaves as a genetic trait. No reports are available on effect of water stress on CIE L*a*b* color space parameters. However, similar results have been reported in Foraminifera reefs under elevated carbon dioxide concentration [41].

Table 3 Descriptive statistics of L*, a*, b* and δ E under irrigated and drought conditions
Full size table
Fig. 5
figure 5

Raincloud plot showing spread of L*a*b* and δ E under irrigated conditions

Full size image

4.3 Testing the validity of model

We used the parameters F test, probability, R2, adjusted R2, root mean square error (RMSE), mean absolute error (MAE) and Akaike information criterion (AIC). The significance of F-test, probability of < 0.05, lower value of MAE, RMSE and AIC implies higher accuracy of a regression model. However, a higher value of R2 as well as adjusted R2 is considered desirable (Table 4).

Table 4 ANOVA for regression for model fitting L*a*b* against chlorophyll
Full size table

A significant F-test indicates that the observed R-squared is reliable and is not a spurious result of oddities in the data set. Thus, the F-test determines whether the observed relationship between the response variable and the set of predictors is statistically reliable. It is equally good in case of prediction as well as explanation. In our case the F-test was significant (p < 0.0001, Table 4), that indicates statistical reliability of our observed relationship between chlorophyll content and L*a*b* and δ E.

The R2 or the coefficient of determination represents the proportion of the variance in the dependent variable which is explained by the linear regression model. It is a scale-free score i.e. irrespective of the values being small or large, the value of R square will be less than one. In our case the R2 was slightly higher in calibration sample (0.778) as compared validation sample (0.714) (Table 5). The R2 values were higher in case of multiple regression model as compared to individual regressions with L*a*b* and δ E.

$${\text{R}}^{2} = 1 - \frac{{\sum \left( {yi - \hat{y}} \right)^{2} }}{{\sum \left( {yi - \ddot{y}} \right)^{2} }}$$
Table 5 Goodness of fit parameters for calibration and validation models
Full size table

The root mean squared error (RMSE) is the squared root of the variance of the residuals and is an indicator of the absolute fit of the model. RMSE shows the correspondence between observed and model’s predicted values. Compared to R-square, which is a relative measure of model fitness, RMSE is an absolute measure of model fitness. RMSE is actually the standard deviation of the unexplained variance. The utility of RMSE is augmented by the fact that it has the same units as that of the response variable. Lower RMSE values correspond to better model fitness, and indicates the degree of accuracy the model in terms of its ability to predict the response. It’s the most important criterion for fit if the main purpose of the model is prediction. RMSE is widely used to elucidate the performance of the regression model with other random models as it has the same units as the dependent variable (Y-axis). In the present study, the RMSE value for calibration sample was 2.46 and for validation sample was 0.662.

$$RMSE = \sqrt {MSE} = \sqrt{\frac{1}{N}} \sum \left( {yi - \hat{y}} \right)^{2}$$

Adjusted R-squared invariably decreases as the predictors are added if the increase in model fit does not make up for the loss of degrees of freedom. Likewise, it will increase as predictors are added, only if the increase in model fit is worthwhile. In fact, more than R2, the adjusted R2 should always be used with models with more than one predictor variable. It is interpreted as the proportion of total variance that is explained by the model. However, R2 is relevant only when main aim of regression analysis is prediction. When the major focus is the relationship between variables, R2 is usually irrelevant. In our study adjusted R2 for calibration sample was 0.719 and for validation sample was 0.689 indicating comparatively lesser decrease by incorporating all the components viz., L*a*b* and δ E. in regression model (Table 5).

Mean absolute error (MAE) is a measure of the average magnitude of the errors in a set of predictions, regardless of their direction. It is estimated from the average of the absolute differences between model prediction and actual observations, where all individual differences have equal weight. Similar to RMSE, MAE is an indicator of the prediction error. Mathematically, it is calculated as the average of absolute difference between observed values and model predicted values. However, compared to RMSE, MAE is relatively less sensitive to outliers.

$$MAE = \frac{1}{N} \sum {\text{I}}yi - \hat{y}{\text{I}}2$$

where,

ŷ = predicted value of y.

y = mean value of y.

As a measure of model prediction error, both MAE and RMSE can take the values from 0 to ∞ and are indifferent to the direction of errors. Both RMSE and MAE are negatively-oriented scores, meaning that lower values are always better as the indicate lower prediction error. In the present study, MAE value for calibration sample was 3.696 and for validation sample was 0.955 (Table 5).

Another indicator of model fitness viz., Akaike information criterion (AIC) is a measure of fitness of different regression models. AIC is calculated as:

$${\text{AIC }} = 2{\text{K }} - 2ln\left( {\text{L}} \right),$$

where:

K: The number of model parameters.

ln(L): The log-likelihood of the model (this value tells US how likely the model fits the data).

AIC measures how likely the model is, accurate for a given data set. The best fit model should have a lower value of AIC. There are no benchmark vales of AIC that could be designated as “good” or “bad” as it is simply used to compare the regression models. However, the best fit model is the one with the lowest AIC values. In the present study, AIC values for calibration sample was 40.256 and for validation sample was − 7.732 (Table 5).

4.4 Effect of water stress on chlorophyll and L*a*b* δ E parameters

Water stress significantly reduced the chlorophyll content in rice diversity panel. Under irrigated conditions, chlorophyll had a mean vale of 50.21 mg/g (28.56 to 54.07 mg/g) as compared to 41.93 mg/g (25.96–49.55 mg/g) fresh weight (Table 1). In terms of the differential response of L*a*b* and δ E parameters under irrigated and drought conditions, there was an increase in L*, a*, b* and δ E under drought conditions. Since under water stress, leaves have decreased chlorophyll meaning much lighter color and as such higher value for L*. Higher negative value of a* indicated more greenness and as such, under water stress, there was a decrease in absolute negative value of a*. In case of b*, since leaves turn increasingly yellowish due to decrease in leaf chlorophyll, there was an increase in a* under water stress. Similarly, an increase in deviation component δ E, that signifies deviation from L*, a*and b* attributes of a standard chlorophyll solution of known concentration, also validates the assumption that chlorophyll content decreases under water stress. Wang et al. [22] also reported usefulness of L*, a*and b* based model as the indices derived from the CIE L* a* b* color model had relatively higher correlation coefficients with SPAD readings in case of rice.

5 Conclusion

Crop phenotyping is increasingly driven by use of technology. From high end phenomics platforms to low-cost methods and gadgets that have acceptable level of reliability and reproducibility, phenotyping is increasingly becoming a machine-driven process aided by programming and modelling approaches. In resource constrained breeding programmes, where high end phenomic research is not possible, low-cost gadgets including android/IOS based smartphones with open software support can effectively help in high throughput phenotyping with fair amount of precision. The present approach is quick, cheap and non-destructive and has been validated in rice and beans and can be mainstreamed in phenotyping programmes for screening large diversity panels under various stresses. Despite the fact that the present study was undertaken using a simple approach of multiple regression followed by model fitting, the study could help us develop much more comprehensive methods of handling huge data using more objective modelling analysis to devise low cost, non-destructive methods, the advances in low cost electronic instrumentation can greatly facilitate research efforts especially in small low budget institutions to undertake research efforts into crop response to varied climatic conditions.