Monitoring Flue-Cured Tobacco Leaf Chlorophyll Content under Different Light Qualities by Hyperspectral Reflectance

Rapid assessment of foliar chlorophyll content in tobacco is critical for assessment of growth and precise management to improve quality and yield while minimizing adverse environmental impact. Our objective is to develop a precise agricultural practice predicting tobacco-leaf chlorophyll-a content. Reflectance experiments have been conducted on flue-cured tobacco over 3 consecutive years under different light quality. Leaf hyperspectral reflectance and chlorophyll-a content data have been collected at 15-day intervals from 30 days after transplant until harvesting. We identified the central band that is sensitive to tobacco-leaf chlorophyll-a content and the optimum wavelength combinations for establishing new spectral indices (simple ratio index, RVI; normalized difference vegetation index, NDVI; and simple difference vegetation index, DVI). We then established linear and BackPropagation (BP) neural network models to estimate chlorophyll-a content. The central bands for leaf chlorophyll-a content are concentrated in the visible range (410 680 nm) in combination with the shortwave infrared range (1900 2400 nm). The optimum spectral range for the spectral band combinations RVI, NDVI, and DVI are 440 and 470 nm, 440 and 470 nm, and 440 and 460 nm, respectively. The linear RVI, NDVI, and DVI models, SMLR model and the BP neural network model have respective R values of 0.76, 0.77, 0.69, 0.78 and 0.86, and root mean square error values of 0.63, 1.60, 1.59, 2.04 and 0.05 mg chlorophyll-a/g (fresh weight), respectively. Our results identified chlorophyll-a sensitive spectral regions and new indices facilitate a rapid, non-destructive field estimation of leaf chlorophyll-a content for tobacco.


Introduction
Leaf chlorophyll content is a useful indicator of several aspects of plant health, including nutrient stress, photosynthetic ability, and aging [1] [2]. Leaf chlorophyll is also important for aroma precursors in tobacco. Chlorophyll content is primarily affected by spectral light quality during leaf growth and development [3] [4]. Spectral light quality has been shown to alter chlorophyll in Hordeum [5], pea [3] and Arabidopsis [6].
Detection of chlorophyll in crop leaves in real-time is useful for crop-growth diagnostics and quality monitoring, as well as for the quantitative estimation and simulation of the C and N cycles in agroecosystems [7] [8] [9]. Chlorophyll is easy to measure by laboratory analysis [10] [11] and is frequently measured using a chlorophyll meter, which utilizes as a spectral index and remote sensing techniques, such as the SPAD 502® [Minolta Osaka Co., Ltd., Japan] [12] [13] [14] and remote-sensing techniques. Remote-sensing techniques can be applied to rapidly and accurately estimate crop chlorophyll content in a non-destructive procedure, and therefore it has a strong potential for applications in monitoring growth, yield estimation, and diagnosis of the nutritional status of field crops [15] [16] [17] [18] [19]. To date, few empirical studies evaluated chlorophyll content using remote sensing under different light spectral conditions.
The reflectance signature of crops in the visible light region is primarily affected by pigments. Numerous researchers used hyperspectral data to estimate the pigment content of crops to predict growth, yield, nutritional and health status [20]- [26]. For example, Sachidananda and Deepak [27] proposed a new spectral index model, the normalized difference chlorophyll index (NDCI) to predict chlorophyll-a content. Based on their results, Ju et al. [28] proposed that the red-edge position of leaves could be used to predict chlorophyll content. In addition, numerous researchers used various approaches to investigate hyperspectral inversion in vegetation. Some have used hyperspectral variables and first-order differentials to predict changes in crop chlorophyll density in crops [29] [30] [31]. Others analyzed and compared different spectral variables and radiation transmission models with the objective to improve the accuracy of chlorophyll inversion measurements [32] [33].
Recent studies attempt to develop linear relationships between hyperspectral variables, wavelength regions, and chlorophyll content [34]; however, the absence of a simple linear relationship between hyperspectral variables and chlorophyll content limits the wide use of hyperspectral data for the prediction of crop chlorophyll content. However, error backpropagation (BP) neural networks have powerful nonlinear mapping ability and hence, are promising for application in the prediction of complex, nonlinear relationships involving uncertainty  [41]. However, BP neural network-based methods rarely have been used for the analysis of hyperspectral data as a physiological indicator for crops. Tobacco is an important economic crop in China, which is planted in most of the provinces of China. The quality of tobacco products is affected by differences in spectral light quality in different regions of China. In the present work, we tested tobacco chlorophyll-a content and its corresponding hyperspectral data under different light spectral treatments. Our objectives are to 1) identify the spectral bands influencing chlorophyll-a content in tobacco and to define new spectral indices; 2) quantify the relationship between the content of chlorophyll-a content and a simple ratio vegetation index (RVI), a normalized difference vegetation index (NDVI), and a difference vegetation index (DVI); and 3) evaluate optimal models to monitor tobacco chlorophyll-a content.

Experimental Design and Treatments
We measured the response of tobacco grown under different light spectral con- The light quality (spectral composition) of different light treatments was measured by ASD equipped with a whole-light cosine receiver. The cosine receiver was aligned level with the sky in open-field conditions prior to each test to American Journal of Plant Sciences convert the spectral measurements. Each light treatment was measured nine times at different positions and heights to generate the average light quality (spectral composition) ( Table 1).

Light-Quality Measurements
The light quality (spectral composition) of different light treatments was measured by ASD equipped with a whole-light cosine receiver. The cosine receiver was aligned level with the sky in open-field conditions prior to each test to convert the spectral measurements. Each light treatment was measured nine times at different positions and heights to generate the average light quality (spectral composition) ( Table 1).

Plant Collection and Measurement
After the field measurements of leaf spectral reflectance, the test plants were transferred to the laboratory for measurement to measure chlorophyll-a content. Samples (~0.2 g) were collected using a 4 mm diameter leaf tissue punch, which was uimmersed into 95% ethanol solution, and left be extracted for chlorophyll-a for 24 h in the dark. After the dark treatment, the leaves had a white-green color. Leaf pigment density was measured using a Jasco 560-V colorimetric spectrophotometer (Jasco, Tokyo, Japan). The concentration of extracted chlorophyll-a is calculated from absorbance values at 665, 649, and 470 nm.

Index Definition
Two-band indices were evaluated in three ways: 1) a simple ratio vegetation index (RVI); 2) a normalized difference vegetation index (NDVI); 3) a simple difference vegetation index (DVI).

Statistical Analyses
Multiple comparisons of chlorophyll-a content between the different light treatments were conducted using SPSS software version 16 data has been validated using the data from the 2016 harvest to identify the differences between the models and to check the generic capacities of the prediction models. Error backpropagation (BP) neural network analysis is one of the most widely used artificial neural network analyses because of its ability to solve nonlinear problems. This multilayer feedforward network, reverses the output error into the input layer with the need to disclose the input samples, while the correcting for the weight of each layer of neurons continuously to make the export result constantly close to the target export.

Effect of Different Light Qualities on Chlorophyll-a Content of Tobacco Leaves
The chlorophyll-a content in the five spectrally filtered light treatments shows a changing trend over time, increasing from 30 to 90 days after transplantation and then decreasing from days 90 to 135 ( Figure 1). Chlorophyll-a content in CK declines from day 60 after transplantation. Differences in chlorophyll-a content between the six treatments followed the order R > B > Y > G > W > CK in all stages. The difference is more expressed from day 90 to 135 after transplanting. There is no difference at day 60 after transplanting. These results illustrate that the filtered delays the degradation time of chlorophyll-a and increases its content compared to the control. In relation to the composition of the irradiance spectrum, the decrease in ultraviolet, blue-violet light, and the increase in near-infrared light compared with CK is favorable for chlorophyll-a synthesis, delaying the degradation of chlorophyll-a and the maturation of tobacco.

Stepwise Multiple Linear Regression (SMLR)
Twenty NDVI, 20 RVI and 20 DVI indices with the optimal R2 for the linear regression from the contour maps were taken as the independent variables to establish the SMLR estimation model for chlorophyll-a content. The spectral indices selected by stepwise regression were NDVI (R440, 470), NDVI (R650, 2060) and DVI (430, 460). The R 2 of the stepwise model was 0.78, as shown in Table 2.

The BP Neural Network
The BP neural network is composed of three layers: an input layer, output layer (tobacco leaf chlorophyll-a content), and a hidden layer. The hidden layer has 22 nodes, determined by a "trial and error method". The tansig and purelin func-  The 2017 and 2018 data are used as training samples for this model, and the R 2 of the prediction result is 0.86 (Table 2). In addition, point-to-point mapping results of the BP neural network ( Figure 5(d)) also demonstrates that the BP neural network has a higher accuracy than the linear NDVI, RVI, and DVI models and SMLR model.

Testing of Chlorophyll-a Estimation Models
To determine whether the estimation models could be used under different light conditions, the independent validation dataset from the 2016 experiment was used to test the reliability of the models ( Figure 6). In the 1:1 relationship plot ( Figure 6), the more accurate the predictive equations, the more closely clustered the points are with respect to the theoretical 1:1 correspondence.

Response of Chlorophyll-a Content
Chlorophyll-a is the primary photosynthetic pigment absorbing photosynthetically active radiation (PAR) from sunlight in vegetation canopies [30]. Its concentration strongly influences maximal leaf light-use efficiency (e.g., Waring et al., 1995 [42]). There is little research on changes in chlorophyll-a content in tobacco in relation to changes in the spectral composition of light. In the present study, the content of chlorophyll-a under different light treatments follows the pattern R > B > Y > G > W > CK in all stages. This indicates that changes in the proportions of various spectral regions (e.g., decreased ultraviolet and blue-violet light and increased near-infrared light) enhance chlorophyll-a content and delay its degradation.

Selection of Active Chlorophyll-a Absorption Bands and a New Spectral Index
The development of a new, simple, and reliable spectral index is a challenge and hot research topic in the field of agricultural remote-sensing monitoring. In the  study, we investigated the NDVI, RVI, and DVI indices, as combination of any two bands within the spectral range of 350 -2500 nm, and then construct a new, simple spectral index for chlorophyll-a content derived from the optimal combinations of these spectral ranges. In contrast with other approaches, this method is more appropriate for a given crop and does not require previous knowledge.
Comprehensive and systematic analyses have been performed to determine the active bands of chlorophyll-a content in tobacco during 3 consecutive years under various light emission spectral treatments with two varieties. The spectral bands having a high impact on chlorophyll-a content are 440, 460 and 470 nm. The three wavelength bands are all located in the blue-violet region of the visible spectrum, in which a strong absorption band of chlorophyll-a is located. These findings are consistent with those of a prior report [43]

Accuracy and Generic Potential of the Prediction Models
Various predictive models for monitoring chlorophyll-a content have been established based on new spectral indices, and subsequently validated. All models showed significant levels of accuracy and stability. The ability of the BP neural network to take advantage of point-to-point fitting improved the performance of the BP model significantly. Therefore, the BP neural network can be used for a precise prediction of chlorophyll-a content in flue-cured tobacco under spectrally different light sources.
A 3-year field investigating the impact of light spectral emission on flue-cured tobacco yielded numerous samples with a good representativeness of chlorophyll-a. The BP neural network prediction model generated the most reliable performance of chlorophyll-a, with results validated by a validation dataset.

Conclusion
Abundant data with regard to crop leaf hyperspectra are available. These data can be used to extract information that reflects the physiological status of crops by defined hyperspectral indices. Here, we analyzed the relationship between chlorophyll-a content and NDVI, RVI, and DVI, and identified the spectral bands having the largest impact to monitor chlorophyll-a content in tobacco. We demonstrated that a BP neural network has the highest accuracy and lowest error, and can be used to monitor chlorophyll-a content in flue-cured tobacco. The method described here for establishing new spectral indices for chlorophyll-a content in tobacco can be applied for other biochemical variables and other plant species as well.