Use of Fluorescence and Reflectance Spectra for Predicting Okra (Abelmoschus esculentus) Yield and Macronutrient Contents of Leaves

In-field proximal sensing of most major crops nutrients still remains an economical and technical challenge. For this purpose, the use of effective multi-excitation fluorescence and reflectance wavelengths is explored in this work on Okra plant. Visible-near infrared (400 1000 nm) reflectance and multi-fluorescence data were collected at leaf scale in a chemically fertilized field by using an USB spectrometer mounted with an Arduino-based LED driver clip. N, P, K and Ca content of samples leaves were measured using reference methods. Average pods yield and leaves macronutrients content were calibrated using IRIV-PLS regression after spectra pretreatments. Single informative wavelengths bands in reflectance, red and far-red fluorescences were selected for building yield and macronutrient content models. We showed that flowering stage was more suitable for yield prediction. Moderately useful macronutrient models were found in Ca content (RPDval = 1.93, rP = 0.818) and potassium content with RPDval = 1.8, rP = 0.88. P and N yielding prediction performance of RPDval = 1.61 (rP = 0.718 ) and RPDval = 1.46 (rP = 0.56) respectively were less accurate. This study demonstrates potentiality of fluorescence and reflectance spectroscopy for accurate estimation of leaf macronutrient content and crop yield. High selectivity obtained from resulted spectral bands could lead to the development of reliable, rapid and cost-effective devices for nutrient diagnosis. How to cite this paper: Dibi, W.G., Bosson, J., Zobi, I.C., Tié, B.T. and Zoueu, J.T. (2017) Use of Fluorescence and Reflectance Spectra for Predicting Okra (Abelmoschus esculentus) Yield and Macronutrient


Introduction
In modern agriculture, remote sensing offers efficient tool for enhancing crop production while satisfying sustainability requirements by diagnosing plant nutritional status. Indeed, remote sensing is widely used for early information management about crops fertilization needs, yield forecasting, diseases identification and decision support system from plant leaf to plant canopy and farmland to landscape scale. It is therefore quite rightly that methods derived from remote sensing are highly regarded as key components of Fertilizers Best Management Practices (FBMPs) [1]. According to statistical prevision, world population will grow over 9 billion by 2050 [2] [3]. This situation may increase pressure on world agriculture by the use of greater amount of chemical fertilizer. To overcome future challenges in food supply for this expected world population, locks and limits in the applications of field remote sensing technology have to be solved.
The availability of nutrients represents a limiting parameter of great importance in the evaluation of crop yields [1]. Macronutrients are plant nutrients required in important amounts and are constituted by nitrogen (N), potassium (K), calcium (Ca), magnesium (Mg), phosphorous (P), and sulfur (S) [4]. N, P, K, and Mg are used in photosynthesis and respiration, whereas Ca serves to cell division and to cell walls construction [5]. As yet, dynamics of progress on field remote sensing methods currently makes proximal portable or mounted devices directly used in the field to measure N content with acquisition of fluorescence and reflectance data [6] [7] [8]. Nevertheless, in-field determinations of other mineral nutrients are not as effective as in the case of nitrogen. Nutrients such as P, K and S are particularly concerned [3]. About developments in fast spectroscopy for plant mineral analysis, it appears that the use of visible data, combined with the lowest part of near infra-red range had resulted in good calibration models especially in K [9] as well as in Ca and Mg cases [10]. P prediction was poor whereas authors in these studies have employed full spectral range (400 -1000 nm). Using Vis-NIR hyperspectral imaging, Zhang et al. [11] gave acceptable prediction performance of leaf P macronutrient content in region of interest. Since correlations between polyphenol and mineral content in olive leaves were established in Cetinkaya et al. [12] study, we might expect better prediction performance of leaf macronutrient content by using reflectance additionally with fluorescence data.
Okra (Abelmoschus esculentus (L.) Moench. syn. Hibiscus esculentus L.) is a rich source of protein, vitamin, magnesium, potassium, manganese, sodium, calcium, iron, copper and zinc [13] [14] [15]. Apart from its rich nutritional as-  [17]. Although these amounts vary with factors such as: cultivar, plant density, soil type, whether the crop is irrigated or not, the climate and other environmental conditions [18], higher productivity of Okra has been obtained with manure fertilization [19] [20] [21]. In West Africa, the gumbos occupy the second place of vegetable production after tomatoes [22]. Two okra species are cultivated in Ivory Coast: Abelmoschus esculentus and Abelmoschus caillei whose average yield is up to 6 t•ha −1 and potential yields vary between 11 and 13 t•ha −1 [23].
Chemistry or biochemistry data retrieval in spectroscopy measurement is accurately performed in multivariate calibration methods. These procedures are largely known as chemometrics in which efforts are made for developing mathematical and statistical methods to extract relevant, useful and efficient information from raw spectral data. In the optimization of chemometric process, steps such as spectral pretreatment, variable selection and latent factors are highly cases sensitive [24]. Especially in quantitative spectroscopy, single wavelengths selection has been shown to improve precision and accuracy in the calibration process [25]. Moreover, identification of effective wavelengths could offer possi- In the present study, we demonstrate potential of multi-excitation fluorescence combined with active reflectance for sensing yield and leaves macronutrients contents. This main objective is twofold: 1) to identify the best phenologic stage for pods yield prediction; and 2) to develop predictive models of N, P, K, and Ca concentration in okra leaves.

Plant Material, Environmental Conditions and Soil Sampling
The seeds of gumbo from the variety GB1230 provided by the Centre National de Recherche Agronomique (CNRA) were used in this study. boratory for 15 days and clods were crushed by hand and gravels removed. Subsequently, the soil was slightly crushed and passed through a 2 mm calibrated sieve. Analyses were performed on soil fractions smaller than 2 mm. pH was measured in a water-soil (ratio 2/5) solution with a pH-meter. The methods of Walkley and Black [27], Kjeldahl [28] and Thomas [29] were respectively used for the determination of organic carbon, total nitrogen and organic ammonium.
Cation exchange capacity (CEC) and exchangeable cations (Ca 2+ , Mg 2+ , Na + , K + ) were extracted by the ammonium acetate method [29]. Total phosphorus is extracted by perchloric acid while assimilable phosphorus was estimated with Olsen method modified Dabin. Chemical properties of samples analysed indicate a very low fertile soil with fairly rich organic matter according to the critical classes of chemical elements [30]. The mean soil pH was 6.

Field Experimental Design and Fertilizers Management
The  Table 1.
Fertilizers were prepared using urea (45% N), potassium chloride (60% K 2 O) and tricalcic phosphate (27% P 2 O 5 ). They were applied once and locally for each plant of a plot at twenty days after sowing corresponding to 2 -3 leaves stage.
Each fertilizer proportion was calculated on the basis of the optimal requirement per plant knowing the elementary surface of application according the formula [31]: With x: fertilizer proportion per hectare; S: elementary surface (m 2 ); C: content of fertilizing unit (%).

Spectra Acquisition
The experimental setup was detailed in a previous study [32] and was adaptation from Brydegaard et al. [33] one. It used an USB4000 spectrometer Ocean Optics, a laptop and an Arduino microcontroller. The Led clip is composed of three Light-emitting-diodes (Leds) exciting the fluorescence at 380 nm (UV), 520 nm (green) and 630 nm (red) while the emitted fluorescence light was detected in the red and far-red spectral regions using a long-pass filter at 650 nm. A white LED emitting (400 -700 nm) was utilized for reflectance measurement. A schematic drawing of the experimental setup used is given in Figure 1.

Spectra Preprocessing
Before starting chemometrics analyzes, three preprocessing methods namely gaussian filtering, standard normal variate transformation (SNV) and second derivation were applied to each spectra type (fluorescence with red, green, UV Led and reflectance with white Led). SNV standardized each spectrum by its own average and standard deviation for quantity correction [34]. It changes relation between peaks and enhances noisy signals. If prior to SNV one peak varies, after SNV all peaks vary. Derivatives are a form of high-pass filter and frequency-dependent scaling and are often used when lower-frequency (i.e., smooth and broad) features such as baselines are interferences and higher-frequency (i.e., sharp and narrow) features contain the signal of interest. Hruschka [35] showed that the derivative transformation methods could enhance the signal-to-noise ratio and resolve overlapping absorption features. Additionally, the derivative hyperspectra of crops could also eliminate or minimize the effects of leaf water content, noise, and atmosphere on N estimations of crops. Rieppo et al. [36] noted that second derivative spectroscopy allows more specific identification of small and nearby lying absorption peaks which are not resolved in the original spectrum, thereby offering means to increase the specificity of absorption peaks for certain molecules of the tissue.

Content Measurement of N, P, K and Ca
After the spectral measurement of leaves from each experimental plot in the field, those leaves were collected and destructively sampled to determine their macronutrient content. Chemical analysis of plants tissue were conducted in the Laboratoire de Pédologie and Laboratoire des Procédés Industriels, de Synthèses, de l'Environnement et des Énergies nouvelles (LAPISEN) at National Polytechnic Institute Houphouet Boigny (INPHB). The reference method for nitrogen content was the Kjeldahl [28]. Dried and crushed leaves samples of 0.4 g are mineralized and put in solution with nitric and chlorhydric acid [37]. Phosphorus content was determined by UV/Vis Spectrometer Jasco V-530 Inc. Calcium and potassium content was determined by flame atomic absorption spectrometry air-acetylene using spectrophotometer VARIAN type AA20, Australia [38].

Chemometric Analysis
A general workflow procedure has guided the development of macronutrient models. In this context, Figure 2 gives a substantial overview. Four main methods were employed. In chemometric process, initial samples are firstly screened with Monte Carlo method for detecting potential outliers. After eliminating these dubious data, the remaining samples are splitted in training and test data according to Kennard-Stone data partition method. These two subsets are used as input parameters in IRIV algorithm for selecting optimal reflectance and fluorescence wavelengths. Finally, PLSR using normal samples was conducted with optimal spectral data obtained from IRIV process. Then, calibration models and their validation are generated.

Monte Carlo Method of Outliers Detection
Applied at first on QSAR/QSPR data set, the MC method is proved to be an approach for diagnosing dubious samples contained in raw dataset [39]. This method was used to increase model performance prediction and consisted in stud- In Equation (4) At each iteration, the algorithm selects the point furthest from the points already retained. This process is executed again until selecting the N desired points. The aim of KS algorithm is to find a representative and uniform subset selection of data.

IRIV Selection Method
IRIV stands for Iteratively Retaining Informative Variable. This method was developed by considering the data matrix x which contains N samples in rows and p variables in columns, and y, of size 1 N × , denoting the measured property of interest [26]. Binary matrix sampling consists of generating a binary matrix M of dimension K P × which is assigned randomly with the number "1" or "0" to each column. Each row of the binary matrix corresponds to an individual sampling, where "1" denotes the samples selected for modeling and "0" for unselected samples. By the process of inclusion and exclusion of the binary values "1" or "0", the matrix 1 M is obtained by changing for the first column all "0" to "1" and vice versa. Then a subset (variable ith column) is evaluated by PLS and a statistical distribution of the RMSECV at inclusion and exclusion is obtained. These statistical distributions of RMSECV for a variable are tested with the assumption that the distribution of the data is the same for two groups in order to identify the importance of this variable. Mann-Whitney's non-parametric U-test is used for this purpose. According to the p-value obtained, the variables are classified into 4 categories: highly informative, weakly informative, interfering and non-informative. After removing the last two types of variables, a new subset is generated and the same calculation cycle is repeated. At the end of the iteration, when no non-informative or interfering variable exists, a background elimination procedure is carried out to output the optimal variables selected.

PLSR Analysis
Partial Least-Squares Regression (PLSR) has become a standard tool in chemometry and is widely applied in various fields such as chemical [44], economical [45], biological [46] [47] [48] and social sciences [49]. In this study, PLSR analysis was performed to establish regression models for the determination of the target chemical concentrations measured in fresh gumbo leaves (variable matrix ) based on the optimal variable spectra (variable matrix X) obtained from the IRIV process. Autoscale technique was the main preprocessing method employed for X-Block and Y-Block. Cross-validation method was based on venetian blinds using 5-folds split for K model, 6-folds for N and Ca models, then 8-fold for P one. Models for macronutrients content developed were built with a number of samples in calibration set comprising 67% to 75% of whole data.

Model Evaluation
To check whether the model satisfy to standard qualities in multivariate regression, we referred to some common chemometrics terms: Bias, Ratio Performance Deviation (RPD), correlation coefficient in calibration (R C ) and prediction (R P ), root mean square error (RMSE) of calibration (RMSEC) and prediction (RMSEP).
According to Williams [50], correlation coefficient, SEP, bias and RPD are the most useful statistics for evaluating calibration performance. Moreover, using Malley et al. [51] criteria based on correlation coefficient and RPD on validation data, we are able to classify the success of predictions. Indeed these criteria propose a guideline scale for describing the performance of calibrations for environmental samples as follows: Excellent, R P > 0.95, RPD > 4; Successful, R P = 0.9 -0.95, RPD 3 -4; Moderately Successful, R P = 0.8 -0.9, RPD 2.25 -3; and Moderately Useful, R P = 0.7 -0.8, RPD 1.75 -2.25.

Software
All  On the level of nutritional information, both calibration and validation set covered large value ranges. This is due to fertilization design, leaf position and period of time run out during the first and the second acquisition. Moustakas et al. [53] showed that in Okra (Abelmoschus esculentus (L.) Moench.) major nutrient uptake was rapidly increasing between 30 and 60 days after transplanting. This ensures higher concentration value, mainly observed in the second acquisition samples set. Table 2 summarizes yield and macronutrient contents used in this study.

Spectral Features of Preprocessed Okra Fresh Leaves
After SNV and 2 nd derivative, the preprocessed spectra are successively combined to form a single spectrum. Autoscale of those combinations of multi fluorescence and reflectance spectra is shown in Figure 3.
On Figure 3 examination, we can notice enhanced peaks in number and intensity more in fluorescence spectra than in reflectance one. These peaks may have been caused by different contents of nutritional parameters as derivative spectroscopy has been proved to be an efficient method which enhances resolution in spectroscopic data [54].

Outliers Detection Using MC Method
Analysis of Figure 4 shows that there are different outliers in number and type by chemical content. Some similar outliers are found in calcium as well as in potassium and phosphorus data.
Nitrogen samples are very low because chemical analysis at first data acquisition performed on P, K and Ca has consumed almost samples. Besides some of

Okra Best Phenologic Stage for Fruits Yield Prediction
After partitionning yield and spectra data with Kennard-Stone algorithm, we applied IRIV algorithm to find out the best phenologic stage for pods yield prediction. YieldAcq1 is related to first spectra acquisition at vegetative stage and   [57] one. Their study on grain maize prediction by using leaves reflectance have indicated that coefficients of determination (r 2 ) between predicted and actual grain yield were highest for measurements conducted at anthesis and milk-grain stage. Thus, we conclude that fluorescence and reflectance measured at flowering stage are suitable for gumbo fruits yield prediction.   Ca and K models have better calibration success according to RPD value calculated on validation data. These models are classify as moderately useful according to Malley et al. [51] criteria although the r P coefficients above 0.8 were high enough to be considered moderately successful (Table 4) [58], Ward et al. [59] and Aldana et al. [60] results ones in which 400 -2500 nm spectral ranges were used to predict dry ground leaves samples. Highest RPD val value of 6.1 for potassium has been found in Menesatti et al. [10] work using VIS-NIR in 400 -1000 nmon fresh orange leaves but Ca RPD val = 1.5 was low. In Figure 5 Klančnik et al. [61] has shown that reflectance spectra were significantly af- and anthocyanin content. Increase of anthocyanin content in petals due to potassium uptake has been reported [62]. An increase in chlorophyll, and carotenoid contents was also found as a result of potassium treatments excepting for higher concentration [63]. Calcium and potassium models developed denote strong relevance of the selected wavelengths in fluorescence and reflectance mode that could be used for accuracy measurements of leaf calcium and potassium content.

Prediction of Macronutrient Content
In contrast, the two last models have low performance prediction owing to their RPD val values (1.6 for P and 1.47 for N). Nevertheless, the correlation coefficient in phosphorus prediction around 0.72 represents a satisfactory result since poor prediction ability has been reported in previous investigations [10] [59] [64].
Our result in validation set is comparatively equivalent to Zhang et al. [11] phosphorus prediction when they applied hyperspectral imaging technique in VIS-NIR region (380 -1030 nm) on oilseed rape leaves region of interest. Photosynthetic process largely relies on P containing compounds [65]. P plays a very important role in the composition of chloroplast and photosynthesis and can be directly involved in the assimilation of photosynthesis and photosynthetic phosphorylation [66]. Thus, photosynthesis is sensitive to low P stress [67]. Variation in P content affects chlorophyll a fluorescence response with red light excitation. Our findings also show that phosphorus level concomitantly induced variation in flavonol and anthocyanin content respectively responsible of fluorescence in far-red band under UV (375 nm) and green (520 nm) light excitation. Involved wavelengths are mentioned in Table 3. This result corroborates Ulrychová et Sosnová [68] study which shows that phosphorus deficiency result in an evident increase of anthocyanin content. But, Stewart et al. [69] work attributed increase of total flavonol in tomato plant leaves mainly to nitrogen deficiency than to low phosphorus content.
In the case of nitrogen prediction, most investigations have established high accuracy measurements by using reflectance spectra [70] [71] [72]. In this work, 7 optimal variables in 2 nd derivative reflectance bands (570 -571; 607 -608 nm) are the most informative regions for nitrogen model development. These regions correspond to negative slope in green-yellow leaf reflectance and are related to LUE (Light Use Efficiency) green spectral region. They represent high selectivity for nitrogen wavelength screening in light of previous studies. Few sampling explains low value of correlation coefficient and RPD on validation. Nevertheless, according to Wang model performance classification [73], RPD and correlation coefficient values respectively in 1.4 -2 and 0.5 -0.75 ranges are acceptable. As shown in Figure 6, the calibration models are representative of macronutrient content variation and present in test set lower bias than calcium and potassium models.
From this study, we can conclude reasonable prediction performance of macronutrient models using combined fluorescence and reflectance informative  wavelengths. Results evidence with representative dataset is given for calcium, potassium and phosphorus Okra leaves content. However models building process has left out many outliers samples. More generalized model should take into account larger data. Since the experiments were conducted under one cultural and ecological environment in the field conditions of three months with variety GB1230 of Okra seeds, methods and monitoring models still need to be verified in other ecological areas.

Conclusion
The results from this study showed that fluorescence and reflectance spectral acquisition at flowering stage are appropriate for okra yield estimation. In addition, optimal wavelengths were selected in red and far-red bands from three ex-Open Journal of Applied Sciences citation fluorescence and specific bands reflectance derivatives spectra especially for yield estimation using spectra acquired at flowering stage as well as for potassium and phosphorus data. Macronutrients content prediction has followed a process trajectory using Monte Carlo outliers detection, Kennard-Stone data partition and IRIV approach combined to PLSR. Moderately useful models are concerned with calcium content (RPD val = 1.93, r p = 0.818) and potassium content (RPD val = 1.8, r p = 0.88). Less accuracy but acceptable with RPD val = 1.6, r p = 0.718 was obtained in phosphorus prediction and RPD val = 1.46, r p = 0.56 in nitrogen prediction. Our work gives a substantial understanding of macronutrient content relationship with some plants biochemicals by linking them to new relevant spectral bands. Thus fluorescence and reflectance spectroscopy could be a promising alternative to acquire a reliable predictive view of the macronutrient concentration of Okra leaves. Furthermore, such method could be optimized and applied on other agricultural cultures in order facilitating the rapid and easy evaluation of crops nutritional status by the development of low-cost portable devices for crops monitoring.