Download as pdf or txt
Download as pdf or txt
You are on page 1of 35

Original research article

Journal of Near Infrared Spectroscopy


0(0) 1–13
Estimation of critical nitrogen contents in peach ! The Author(s) 2020
Article reuse guidelines:
orchards using visible-near infrared spectral sagepub.com/journals-permissions
DOI: 10.1177/0967033520939319
mixture analysis journals.sagepub.com/home/jns

Mert Dedeoglu

Abstract
The aim of this study was to predict the critical nitrogen (N) content in peach trees using spectrometric measurements. A
nutrient-controlled hydroponics experiment was designed for this purpose. Peach saplings were grown under three N
conditions: deficient, sufficient, and excessive. The reflectance values of a plant leaves were measured using a handheld
field spectroradiometer fitted with a plant probe. The N contents of leaves were determined in the laboratory and Gaussian
mixture discriminant analysis (GMDA) was used to estimate N levels in the leaves from reflectance values. The N levels were
categorized for each of the three different N conditions. The wavelengths at 425 nm, 574 nm, 696 nm, and 700 nm were
found to be diagnostic of the different N levels. The model developed here classified the experimental plants with high
accuracy for NDeficient, 89.28%; NSufficient, 96.30%; and NExcess, 71.42% with 85.71% coefficients. The reliability of the model
was also tested under field conditions using 96 peach trees representing the three different N status. Leaves were analyzed
by reflectance at 425 nm, 574 nm, 696 nm, and 700 nm, which functioned in real N, percentage classes determined based on
the laboratory analyses of the orchard samples, and the data were categorized as NDeficient, NSufficient, and NExcess with a
similarity ratio of 77.78%, 80%, and 67.74%, respectively with the general correct classification rate of 75%. The study
findings showed that the model developed using hyperspectral reflectance data can discriminate different N nutritional
status in plants with an accuracy of 70% and can be applied under field conditions. The results of this research provide a
new perspective for future studies by showing that GMDA with hyperspectral remote sensing may be useful for the
classification of different plant nutrient contents.

Keywords
Hydroponics experiment, mixture discriminant, nitrogen, peach, VNIR spectroscopy
Received 21 February 2020; accepted 10 June 2020

Introductions
The energy that can cause the stretching and bend-
Visible and near infrared (VNIR) spectroscopy is a ing of bonds between the atoms of organic molecules
rapidly developing analytical technique with an appli- within the plants, such as –CH, –OH, –NH, C ¼ O,
cation in agriculture. It has many advantages, such as and –SH, is absorbed in the infrared region.12,13 As a
being rapid and simple to use, and has a low associ- result, significant differences in energy absorbance by
ated cost and good reproducibility and is environ- plants can occur in the visible and near infrared
mentally friendly.1–4 The most common type of region. Using multivariate statistical analysis,14
VNIR spectroscopy in agricultural applications is many researchers have reported studies involving
spectroradiometry, which is used in hyperspectral near infrared spectroscopy and its application to
remote sensing.5,6 Spectroradiometers are used for fruit trees, which show that reflectance characteristics
the quantitative measurement of radiance, irradiance,
are directly related to leaf chlorophyll and mineral
reflectance, and transmission,7 and the key element of
content15–18; therefore, it is possible that nutrient
spectroradiometric methods is to find the unique
reflectance values of objects within the electromagnet-
ic spectrum.8 Different plants have specific reflectance Department of Soil Science and Plant Nutrition, Agricultural Faculty,
values,6,9 due to differences in their chemical compo- Selcuk University, Konya, Turkey
sitions10 and morphological features which contain Corresponding author:
the spatial information about the size and shape of Mert Dedeoglu, Department of Soil Science and Plant Nutrition,
the plants profiles.11 Agricultural Faculty, Selcuk University, Konya, Turkey.
Email: mertdedeoglu@gmail.com
Estimation of Critical Nitrogen Contents in Peach Orchards using VNIR

Spectral Mixture Analysis

Abstract

The aim of this study was to predict the critical nitrogen (N) content in peach trees using the

spectrometric measurements. A nutrient-controlled hydroponics experiment was designed for

this purpose. Peach saplings were grown under three N conditions: deficient, sufficient, and

excessive. The reflectance values of a plant leaves were measured using a handheld field

spectroradiometer fitted with a plant probe. The N contents of leaves were determined in the

laboratory and Gaussian mixture discriminant analysis (GMDA) was used to estimate N levels

in the leaves from reflectance values. The N levels were categorized for each of the three

different N conditions. The wavelengths at 425, 574, 696, and 700 nm were found to be

diagnostic of the different N levels. The model developed here classified the experimental

plants with high accuracy for N Deficient, 89.28%; N Sufficient, 96.30%; and N Excess, 71.42% with

85.71% coefficients. The reliability of the model was also tested under field conditions using

96 peach trees representing the three different N status. Leaves were analyzed by reflectance

at 425 nm, 574 nm, 696 nm, and 700 nm, which functioned in real N, percentage classes

determined based on the laboratory analyses of the orchard samples, and the data were

categorized as N Deficient, N Sufficient, and N Excess with a similarity ratio of 77.78%, 80%, and

67.74%, respectively with the general correct classification rate of 75%. The study findings

showed that the model developed using hyperspectral reflectance data can discriminate

different N nutritional status in plants with an accuracy of ≥70% and can be applied under field

conditions. The results of this research provide a new perspective for future studies by showing

that GMDA with hyperspectral remote sensing may be useful for the classification of different

plant nutrient contents.

1
Keywords: Hydroponics experiment, Mixture discriminant, Nitrogen, Peach, VNIR

spectroscopy

Introductions

Visible and near-infrared (VNIR) spectroscopy is a rapidly developing analytical technique

with an application in agriculture. It has many advantages, such as being rapid and simple to

use, and has a low associated cost and good reproducibility with being environment - friendly.1-
4
The most common type of VNIR spectroscopy in agricultural applications is

spectroradiometry, which is used in hyperspectral remote sensing.5,6 Spectroradiometers are

used for the quantitative measurement of radiance, irradiance, reflectance, and transmission,7

and the key element of spectroradiometric methods is to find the unique reflectance values of

objects within the electromagnetic spectrum.8 Different plants have specific reflectance

values,6,9 due to differences in their chemical compositions10 and morphological features which

contain the spatial information about the size and shape of the plants profiles11.

The energy that can cause the stretching and bending of bonds between the atoms of organic

molecules within the plants, such as –CH, –OH, –NH, C=O, and –SH, is absorbed in the IR

region.12,13 As a result, significant differences in energy absorbance by plants can occur in the

visible and near-infrared region. These differences are identified at each wavelength by

spectroradiometers and are presented as graphical data. Using multivariate statistical analysis,14

many researchers have reported studies involving near-infrared technology and its application

to fruit trees, which show that reflectance characteristics are directly related to leaf chlorophyll

and mineral content;15-18 therefore, it is possible that nutrient deficiencies could also be

determined using spectral methods.19,20 Plant nutrient stress has been studied by researchers for

many years as a major abiotic factor in plant health.4,21,22 Leaf chlorosis is a common symptom

of nutrient deficiency or toxicity that can be observed in the field; however, the detection of

2
chlorosis is subjective and dependent of the experience of the horticulturist. Such

morphological observations lead to fertilization which based on measured data. The

unnecessary use of fertilizers results in soil salinization, water contamination, and substance

accumulation in plants, all of which result in economic losses and environmental pollution.23

Because of these negative consequences that affect living beings either directly or indirectly,

early detection of nutrient stress prior to the occurrence of chlorosis in plants has been an

important goal of hyperspectral sensing studies.18,24,25 These studies have been based on

determining mathematical correlations between the chemical compositions of plants and

spectral reflectance data (spectral signatures) obtained at different wavelengths from

leaves.26,27

The most useful wavelengths used to determine the effects of stress through the spectral

reflectance are the red, red-edge (680–760 nm), and near-infrared (NIR) spectral ranges.28 In

these studies, the most investigated plant nutrient has been nitrogen (N).29-33 N is one of

seventeen essential nutrients required by fruit trees for growth and for the optimal fruit yield

and quality.34 It is also an important element for the production of many compounds, such as

chlorophyll, nucleotides, proteins, alkaloids, enzymes, hormones and vitamins, in plants.35

Early diagnosis of N deficiency in fruit trees can assist in the management of N-fertilization to

improve crop yields. However, the determination of N content is time-consuming and labor-

intensive and requires the use of dangerous chemical reagents (e.g., Kjeldahl method) that may

result in environmental contamination.34 For these reasons, numerous researches have sought

to estimate the N status of plant leaves under field conditions by using ground-based VNIR

spectroscopy.36-38 Ground-based measuring techniques include reflectance measurements using

handheld spectroradiometers or digital cameras can perform hyperspectral detection by

examining green plant leaves from a certain height (canopy) or by direct contact with the plant

tissue.24 It is known that there is a high positive correlation between the reflectance values

3
obtained from spectrometry techniques and the N content.6,21,39 It has been reported that the

amount of N applied to soil was estimated with high accuracy (PLS, r = 0.945) in orange leaves

(Citrus sinensis L.).40 Similarly, wavelengths of 540–560 nm (visible green) and 990–1010 nm

(near-infrared) were found to be an indicator for active iron levels in apple, cherry, and peach

trees with an accuracy of 76.70%, 75.28%, and 78.69%, respectively.4 In addition, studies

demonstrated that the VNIR both alone or in combination with visible spectroscopy has

capability and accuracy as a predictive tool for the analysis of citrus products. Research were

examples for the analysis of acidity, soluble solids, and firmness in mandarins;41 for the

measurement of soluble solids content in oranges,42 and citrus fruits before harvest;43 and for

the classification and analysis of citrus oils.44 Additionally, the results of another study showed

that ground-based hyperspectral techniques can also produce reliable results in predicting the

N status of the plants;27,45,46 however, it has been suggested that more studies using ground-

based hyperspectral measurement techniques in different plots and study designs are required

in order to develop product-specific models for the determination of nutrient

concentrations.17,45,47,48

Therefore, it was aimed that a model was developed for estimating critical N content in peach

trees using hyperspectral reflectance measurements of the N content in integration with

Gaussian mixture discriminant analysis (GMDA). N was chosen as a nutrient because it is the

most requirement ingredient in fertilizers and plays a critic role in the optimal development of

peaches (Prunus persica L. Batsch). Peach was selected as the test plant because it is a globally

popular fruit and has high nutritional value.49,50 It was also a product that has considerable

economic value in our region.

4
Materials and Methods

The study has five stages which 1) growing the peach saplings, 2) hyperspectral measurements,

3) leaf sampling and laboratory analysis, 4) data analysis and modeling, 5) testing the reliability

of the method.

Study area

The study was based on the comparison the results of laboratory analyses and reflectance

measurements on peach leaves grown using hydroponics experiment also and cultivated in

Egirdir in the Isparta region of Turkey. Nitrogen deficiency is a common and significant

nutritional problem reported in numerous fruit orchards in this region.51

This area has Mediterranean climate characteristics with an annual rainfall of 524 mm and the

average temperature of 12.5 °C. Based on multi-year averages, the highest temperatures occur

in July, with an average of 23.8 °C. December is the wettest month (79.6 mm), whereas August

is the driest (6.5 mm).

Growing conditions and experimental design

One age seedling was planted into perlite, and nutrient solutions were poured into an automated

computer-controlled hydroponics system. The solutions were periodically analyzed and a

targeted that nutrient concentration was 6.3 pH and electrical conductivity of 1.2–1.8 dS/m.

More than 50% of the calculated as standard nutrient content was delivered to the hydroponic

system to achieve EC and pH targets.52

Table 1. Content of standard hydroponic culture solution (1000 liters)

Chemical fertilizers Gram


Mono Potassium Phosphate; KH2PO4 263
Potassium Nitrate; KNO3 583
Calcium Nitrate; Ca(NO3)2 1003
Magnesium Sulphate; MgSO4 513
Iron Chelate EDTA; C10H12FeN2O8 79
Manganese Sulphate; MnSO4 6.10
Boric Acid; H3BO3 1.70
Copper Sulphate; CuSO4 0.39
Ammonium Molybdate; (NH4)6Mo7O24 0.37
Zinc Sulphate; ZnSO4 0.44

5
The plants were divided into three groups based on the applied N content, which were based

on the work of Jones et al.53 and classified as low, optimal and high for one-year-old peach

samplings (Table 2). The experiment was conducted three times, with 21 plants in each

replicate. Samples were taken four times, once each in the months of May, June, July, and

August.

Table 2. The classes of nitrogen nutrient content of peach leaves.53

Element Deficient Sufficient Excess


N (%) <2.99 3.00-3.50 >3.50

In orchards, samples were collected both from healthy trees and from those exhibiting

symptoms of nutrient deficiency. The young offshoots leaves were used as samples. Samples

were collected from 96 different trees in eight separate orchards, and this was repeated three

times over the month of June. In total, 252 leaf samples obtained from the experiment were

used to make a model. Also, 288 leaves samples obtained from orchards were used for

accuracy assessment. Upon the calculation of replication averages, 84 and 96 chemical analysis

results and spectral reflectance values were used for the experiment field and the orchards,

respectively.

Spectral measurements, leaf sampling, and laboratory analysis

A Plant Probe accessory, combined with the Leaf Clip assembly (Analytical Spectral Devices

Inc., Boulder, CO, USA), coupled in the high-resolution spectroradiometer ASD FieldSpec

Handheld was used to measure the spectral reflectance of peach leaves. The Leaf Clip

assembly is designed for use with the Plant Probe accessory, which has an integrated halogen

bulb that emits radiation over the 400 to 1000 nm spectral range. Its design minimizes

measurement errors associated with the stray light. Leaf Clip includes a gripping system for

holding the target sample in place without inflicting damage to the leaf. It also has a two-sided

6
rotating head (2 cm in diameter) that allows both bidirectional reflectance and transmittance

measurements.54

The process of methodology was presented in Figure 1 as a diagram for generating the model

and N% classification based on spectral measurement by using plant probe which has an active

light source with the accessories of leaf clip holding the lamina.

Fig 1. Schematic chart of the methodology.

Reflectance values were obtained from each sample using 3 separate measurements. Spectral

measurement and leaf samples were collected between 12:00 and 14:30 hours local time in

every sampling period. Spectral reflectance was collected from the ends of leaves by placing

the laminas between the leaf veins under the active light source of plant probe. Special attention

was paid that the leaves in the orchard samples had not biological or mechanical damage. In

7
addition, care was taken to ensure that no contaminants such as dust or pesticide residues were

present on the leaves. Barium sulfate-doped (BaSO4) spectralon was used as a white reference

and was measured at the beginning of each survey and every 10–15 minutes thereafter.17

ViewSpec Pro software (Analytical Spectral Devices, Inc., Boulder CO, USA) was used to

process the reflectance values. The nitrogen status of the leaves was determined chemically

using the Kjeldahl method in the laboratory.55 Nutrient contents of the plants were shown in

Table 3. The nitrogen contents of samples obtained from the experiment were ranged between

1.77% and 4.27%. The samples were found to distribute on all N classes, including low,

optimal, and high according to the classification of Jones et al.53

Table 3. Descriptive statistical information N contents of the leaf samples

Statistics Experiment Orchard


N,% Deficient Sufficient Excess Deficient Sufficient Excess
Number of samples 28 27 29 45 20 31
Minimum 1.77 3.01 3.51 1.91 3.00 3.51
Maximum 2.97 3.49 4.27 2.98 3.50 4.84
Mean 2.53 3.30 3.95 2.64 3.26 3.95
SE Mean 0.067 0.029 0.029 0.038 0.035 0.066
StDev 0.357 0.151 0.157 0.257 0.157 0.369
Variance 0.123 0.028 0.025 0.066 0.025 0.136
CoefVar 14.15 4.58 3.99 9.73 4.81 9.36
Kurtosis -0.93 -0.51 1.15 0.29 -1.30 0.38
Skewness -0.64 -0.71 0.16 -0.90 -0.03 0.90
Def.(Exp.-Orc.) Exc.(Exp.-Orc.) Suff.(Exp.-Orc.)
Two-Sample T-Test p 0.163 0.460 1.000

In order to determine whether the model developed using the training data could be used on

the test data, an independent two-sample t-test was applied among the N contents determined

from the experiment field and orchard samples.56 p values for NDeficient–NSufficient–NExcess groups

were found to be 0.163 (p > 0.05), 0.460 (p > 0.05), and 1.00 (p > 0.05), respectively. Thus, it

was determined that the model produced from the experiment field samples could be used on

the orchard samples. The histogram graphs obtained as a result of the laboratory analysis of the

experiment field and orchard samples using the descriptive statistical methods of Minitab 1657

were presented in Figures 2. Distributions of N levels indicating that it did not adhere to a strict

8
normal distribution; thus, the model fulfilled the conditions required for the implementation of

a mixture distribution model.58

Histogram (with Normal Curve) of N from Experiment Histogram (with Normal Curve) of N from Orchards
16
18
Mean 3.267 Mean 3.194
14 StDev 0.6365 16 StDev 0.6379
N 84 N 96
12 14

10 12
Frequency

Frequency
10
8
8
6
6
4
4
2 2
0 0
2.0 2.5 3.0 3.5 4.0 4.5 2.0 2.4 2.8 3.2 3.6 4.0 4.4 4.8
N, % N, %

Fig 2. Distribution of N levels from experiment and orchards samples

The laboratory analyses for the orchard samples revealed an N distribution between 1.91 and

4.84% as expected, which encompassed the three different classes of N found in trees in the

region. Similarly, histogram graphs of the orchard samples showed these data did not conform

to a strict normal distribution. This showed that the sampling size and the test data set

represented a population containing all the N classes in this study. In order to achieve the

intended objectives of the study with the 96 samples obtained from orchards, the N contents of

three different classes with random distribution were obtained.

Data analysis

The N contents of the experiment samples were categorized with the corresponding reflectance

measurements into three different N classes. Thus, in order to develop a model, a training data

set was created from 84 observations (n), consisting of 601 reflectance variables (p) between

the wavelengths of 400 and 1000 nm for each observation. However, the sample size being

lower than the number of variables (n < p) is an important factor that decreases the power of a

model for multivariate analyses that do not fulfill the normal distribution requirement.58 In fact,

9
multivariate modeling studies are used for making inferences to provide the maximum benefit

with the minimum number of variables.59 Therefore, feature selection was performed with the

absolute correlation to determine which of the 601 reflectance values in the training data set

were most appropriate for N classification, and the most effective wavelengths were

determined. In this way, the n > p requirement was fulfilled by dimension reduction. The scaled

hyperspectral reflectance values corresponding to three different N classes were analyzed by

GMDA based on a multivariate normal distribution in Matlab 201860 statistical software by

obje = gmdistribution.fit(X,k) code.

GMDA is a statistical method based on the representation of data as a weighted composition

of one or more Gaussian distributions.61,62 Suppose we have 𝐾 classes in our population and

the number of subclasses in each class is 𝐺𝑘 , 𝑘 = 1, … , 𝐾. Then probability density function in

the mixture discriminant model based on multivariate Gaussian mixture model has been

formulated with the following equation:

𝑓(𝐱, 𝛙) = ∑ 𝜋𝑘 𝑓𝑘 (𝐱, 𝛉𝑘 )
𝑘=1

where 𝐱 is 1 × 𝑝 dimensional observation vector and 𝜋𝑘 is prior probability of 𝑘th class (𝜋𝑘 ∈
(0,1) and ∑𝐾
𝑘=1 𝜋𝑘 = 1). 𝛙 = (𝜋1 , … , 𝜋𝐾 , 𝛉1 , … , 𝛉𝐾 ) is the vector containing all unknown

parameters of the mixture discriminant model based on multivariate Gaussian mixture model.
𝑓𝑘 (𝐱, 𝛉𝑘 ) is probability density function of multivariate Gaussian mixture model for 𝑘th class
and 𝑓𝑘 (𝐱, 𝛉𝑘 ) is defined and presented below.

𝐺𝑘
1 1 −1 𝑇
𝑓𝑘 (𝐱, 𝛉𝑘 ) = ∑ 𝜋𝑘𝑔 1/2
𝑒𝑥𝑝 {− (𝐱 − 𝝁𝑘𝑔 )𝚺𝑘𝑔 (𝐱 − 𝝁𝑘𝑔 ) }
(2𝜋)𝑝/2 |𝚺𝑘𝑔 | 2
𝑔=1

where 𝛉𝑘 = (𝜋𝑘1 , … , 𝜋𝑘𝐺𝑘 , 𝛍𝑘1 , … , 𝝁𝑘𝐺𝑘 , 𝚺𝑘1 , … , 𝚺𝑘𝐺𝑘 ) is the vector containing unknown
parameters of multivariate Gaussian mixture model for 𝑘th class. Where 𝜋𝑘𝑔 is mixture rate of

10
𝑘 𝐺
𝑔th subclasses in the 𝑘th class ( 𝜋𝑘𝑔 ∈ (0,1) and ∑𝑔=1 𝜋𝑘𝑔 = 1). In the 𝑘th class, 𝝁𝑘𝑔 and 𝚺𝑘𝑔
are denoted mixture rate, mean vector, and covariance matrix of 𝑔th subclass, respectively.
The prior probability 𝜋𝑘 can be estimated from the training data or other sources.62 The
estimation of the prior probability 𝜋𝑘 from the training data is defined by
𝑛𝑘
𝜋̂𝑘 =
𝑛
where 𝑛𝑘 is the number observation for the 𝑘th class of the training data. The maximum
likelihood estimations of parameters πkg , μkg and Σkg used in 𝑙𝑜𝑔𝐿(𝛉𝑘 ) calculations for 𝑘th
class can be estimated from training data using the expectation-maximization (EM)
algorithm.63,64 The EM algorithm maximizes the conditional log-likelihood, logL(θk ), for the
𝑘th class of the training data as
𝐺𝑘 𝑛𝑘
𝑙𝑜𝑔𝐿(𝛉𝑘 ) = ∑ ∑ 𝑃𝑘 (𝐺 = 𝑔|X 𝑘 = 𝐱𝑘𝑖 ){𝑙𝑜𝑔𝜋𝑘𝑔 + 𝑙𝑜𝑔𝑓𝑘𝑔 (𝐱𝑘𝑖 , 𝛉𝑘𝑔 )}
𝑔=1 𝑖=1

where 𝐱𝑘𝑖 is 1 × 𝑝 dimensional observation vector of 𝑖th in the 𝑘th class of the training data.
An observation is classified into 𝑘th class, which has the highest posterior probability, based
on Bayes rule, denoted by
𝜋𝑘 𝑓𝑘 (𝐱, 𝛉𝑘 )
𝑃(𝐾 = 𝑘|X = 𝐱) = .
𝑓(𝐱, 𝛙)

Testing the reliability of method

The reliability refers to the study is correct classification rate which can vary from 30% to 99%

according to the purpose of the disciplines such as medical, marketing, sociology, spectroscopy

etc., because they were used different sample size and properties in statically analysis.65,66,67,68

Therefore, it is not possible to say anything definite about the confidence interval of the correct

classification rate. However, some research has stated that the correct classification rate of 85%

or more were “reliable” or “high accuracy” in models used on spectral data from obtained plant

tissue.69,70,71,72,73,74. Thus, a conclusion was made from past research and reliability intervals

were divided into three ranges according to the correct classification rates between > 85% as

high, 85-70% as satisfactory and < 70% as needs to be improved.

11
Result and Discussion

As a result of the measurements and sampling procedures performed in the experiment field,

the N contents of the leaf samples were homogeneously distributed into three different N

classes. Thus, the models were produced represent all the nitrogen thresholds of the peach trees.

In fact, it is advised that varying doses of the related nutrients should be formed in plants for

the estimation of plant nutrient status with spectral reflectance.4 In this study, the distribution

of N contents obtained from plants in randomly-selected orchards was similar to experiment

samples. The similarity has increased the confidence in the statistical interpretation of the

study. The characteristic curves of different N classes were drawn for hyperspectral reflectance

values obtained from the experiment field (Figure 3).

Fig 3. Average reflectance curves representing different N classes in experiment samples

Examination of the mean spectral curves of leaf samples belonging to three different N classes

revealed that there was a difference in the reflectance values between 425 and 750 nm between

the classes. Considering their known characteristics, the wavelengths with the highest

correlations between N and reflectance values were found to be 425, 572, 696, and 700 nm,

12
which exhibited correlations of 0.50, 0.40, 0.43, and 0.42, respectively (Figure 4). Thus, the

reflectance values of these four wavelengths were used for modeling.

Fig 4. Correlation diagram between N contents and reflection values

This event showed that the decrease in reflectance values was resulted from increasing of N

levels, chlorophyllase enzyme activity and hence chlorophyll content.75 In addition, there is a

direct correlation between the amount of chlorophyll a and b and carotenoid pigment with

maximum absorption at 425, 455, and 485 nm.76 Studies have shown that there is a very high

increase in reflectance during N starvation, which can reach 90% depending on the degree of

deficiency.21,77-79

The reflectance curves of the orchard peach samples categorized based on the three critical N

classes exhibited similar spectral characteristics to the experiment samples. (Figure 5).

13
Fig 5. Average reflectance curves representing different N classes in orchards samples

The spectral curves show the correlation between reflectance and N content and the

wavelengths between 400 and 750 nm were similar to those in previous studies.80 It has been

reported that variations in the plant spectral reflectance between approximately 425 and 750

nm can be considered as acceptable indicators of plant nutrient stress.81 It has also been reported

that the wavelength ranges of 520–560, 650–690 nm (visible region), and 710–730 nm (red-

edge region) are useful for estimating different N levels from leaf spectra. The variation in the

reflectance ratios of wavelengths of 420–750 nm has been directly associated with

photosynthetic activity, and 670 nm is the absorption band of chlorophylls a and b,20,29,82,83

whereas 425–485 nm is the reflectance zone of carotenoid pigment.76 In this study, it was

observed that the variations in the carotenoid pigment reflectance zone (approximately 425

nm), green wavelength peak (approximately 574 nm), red-edge minimum chlorophyll

absorption reflectance point (approximately 696 nm) and near-infrared region ( approximately

700 nm) caused differences in spectral reflectance, in particular, energy absorptions due to the

N–chlorophyll/carotenoids correlation.

14
These results show that different N applications result in statistically significant differences in

reflectance curves; therefore, peach trees can be used for statistical analysis of the detection of

critical N classes.

For the analysis, the wavelength reflectance values at 425, 574, 696 and 700 nm obtained from

the experiment plants given three different doses were selected as training data for each class.

Then, In the study, parameter estimates of mixed discriminant model were obtained based on

training data. Suppose we have 3 classes such as deficient, sufficient and excess in our

population and the number of subclasses in each class is 3. Thus, parameter vector 𝛙

containing all parameter of the mixture discriminant model based on multivariate Gaussian

mixture model is shown Figure 6.

Fig 6. The all parameters of the mixture discriminant model based on multivariate Gaussian
mixture model

The mixture discriminant analysis model based on the multivariate normal distributions used
in the study is as follows.

3
3 1 1 𝑇
−1
𝑓(x, ψ) = ∑ 𝜋𝑘 ∑ 𝜋𝑘𝑔 1/2
𝑒𝑥𝑝 {− (x − 𝜇𝑘𝑔 )Σ𝑘𝑔 (x − 𝜇𝑘𝑔 ) }
𝑘=1 (2𝜋)𝑝/2 |Σ𝑘𝑔 | 2
𝑔=1

According to the number of observations in the N% classes, the preliminary probabilities were
28 27 29
determined as 84 , 84 and 84
, respectively.

15
The GMDA model obtained in the study to show the parameter estimates for the mixture of
multivariate normal distribution in each N% classes for θ̂𝑘 was as follows.

28 27 29
̂) =
𝑓(x, ψ 𝑓1 (x, θ̂1 ) + 𝑓2 (x, θ̂2 ) + 𝑓3 (x, θ̂3 )
84 84 84

The hyperspectral reflectance at wavelengths of 425, 574, 696 and 700 nm in the all classes of

nitrogen content from training dataset were presented Table 4.

Table 4. Dataset used to obtain estimation parameters (θ̂𝑘 ) for the N classes

NDeficient NSufficient NExcess


𝑋1,425 𝑋1,576 𝑋1,696 𝑋1,700 𝑋2,425 𝑋2,576 𝑋2,696 𝑋2,700 𝑋3,425 𝑋3,576 𝑋3,696 𝑋3,700
0.042 0.167 0.176 0.244 0.045 0.112 0.116 0.159 0.042 0.151 0.157 0.227
0.041 0.192 0.187 0.267 0.046 0.140 0.148 0.214 0.039 0.124 0.109 0.155
0.042 0.174 0.176 0.248 0.043 0.197 0.205 0.286 0.051 0.142 0.125 0.172
0.035 0.146 0.145 0.213 0.045 0.142 0.147 0.211 0.050 0.099 0.099 0.134
0.037 0.219 0.211 0.301 0.039 0.107 0.095 0.131 0.044 0.115 0.114 0.163
0.039 0.226 0.214 0.308 0.042 0.139 0.115 0.160 0.051 0.121 0.118 0.164
0.039 0.098 0.101 0.140 0.046 0.130 0.134 0.191 0.048 0.144 0.127 0.180
0.039 0.137 0.136 0.199 0.042 0.139 0.117 0.162 0.042 0.116 0.125 0.178
0.044 0.150 0.128 0.180 0.042 0.105 0.103 0.146 0.047 0.112 0.095 0.128
0.040 0.141 0.133 0.199 0.046 0.110 0.111 0.157 0.040 0.133 0.127 0.187
0.036 0.158 0.133 0.191 0.052 0.167 0.188 0.268 0.045 0.137 0.134 0.192
0.040 0.099 0.111 0.154 0.043 0.112 0.120 0.166 0.043 0.107 0.111 0.156
0.037 0.095 0.103 0.143 0.042 0.096 0.098 0.138 0.045 0.133 0.117 0.160
0.040 0.107 0.112 0.154 0.042 0.139 0.119 0.162 0.052 0.137 0.115 0.156
0.045 0.101 0.106 0.145 0.040 0.112 0.116 0.167 0.041 0.121 0.121 0.176
0.045 0.103 0.107 0.144 0.043 0.109 0.108 0.154 0.048 0.121 0.102 0.139
0.040 0.179 0.195 0.276 0.044 0.128 0.113 0.156 0.049 0.137 0.126 0.182
0.048 0.167 0.189 0.264 0.041 0.101 0.099 0.142 0.041 0.114 0.110 0.157
0.043 0.23 0.202 0.282 0.051 0.129 0.108 0.146 0.049 0.133 0.128 0.183
0.036 0.095 0.097 0.135 0.036 0.134 0.133 0.197 0.041 0.108 0.110 0.158
0.046 0.168 0.191 0.268 0.044 0.141 0.126 0.175 0.045 0.113 0.112 0.160
0.045 0.192 0.215 0.291 0.049 0.123 0.126 0.180 0.046 0.146 0.147 0.212
0.049 0.245 0.29 0.394 0.042 0.114 0.104 0.148 0.045 0.140 0.140 0.201
0.038 0.095 0.104 0.143 0.035 0.110 0.127 0.181 0.049 0.127 0.112 0.152
0.044 0.133 0.117 0.162 0.047 0.131 0.132 0.192 0.044 0.103 0.102 0.146
0.046 0.138 0.138 0.201 0.047 0.108 0.110 0.155 0.050 0.152 0.143 0.208
0.046 0.163 0.180 0.251 0.040 0.102 0.100 0.142 0.046 0.145 0.136 0.194
0.045 0.160 0.168 0.24 0.044 0.117 0.115 0.168
0.045 0.108 0.110 0.154

16
̂𝑘 ) of the Gaussian mixture model for the
The maximum likelihood estimations parameters ( 𝛉

all classes of nitrogen content were given Table 5-6-7.

̂1 )
Table 5. Estimations parameters for the NDeficient class ( 𝛉

𝜋̂11 0.357126
𝛍
̂11 0.040799 0.107599 0.108599 0.149999
1.06E-05 3.38E-05 1.70E-05 2.18E-05
3.38E-05 0.000315 0.000139 0.000206
̂11
𝚺
1.70E-05 0.000139 7.18E-05 0.000105
2.18E-05 0.000206 0.000105 0.000156
𝜋̂12 0.355008
𝛍
̂12 0.044806 0.170198 0.184296 0.258418
8.05E-06 4.69E-05 7.93E-05 9.90E-05
4.69E-05 0.000859 0.001238 0.001545
̂12
𝚺
7.93E-05 0.001238 0.001828 0.002278
9.90E-05 0.001545 0.002278 0.002847
𝜋̂13 0.287866
𝛍
̂13 0.038913 0.187371 0.179538 0.256611
6.69E-06 4.27E-05 4.09E-05 5.11E-05
4.27E-05 0.00108 0.000922 0.001244
̂13
𝚺
4.09E-05 0.000922 0.000939 0.001279
5.11E-05 0.001244 0.001279 0.001753

̂2 )
Table 6. Estimations parameters for the NSufficent class ( 𝛉

𝜋̂21 0.276326
𝛍
̂21 0.045226 0.142927 0.147052 0.209294
1.32E-05 3.68E-05 6.67E-05 0.000102
3.68E-05 0.000739 0.000926 0.00129
̂21
𝚺
6.67E-05 0.000926 0.001204 0.001692
0.000102 0.00129 0.001692 0.002388
𝜋̂22 0.290881
𝛍
̂22 0.04504 0.110304 0.108028 0.152263
1.31E-05 3.30E-05 2.07E-05 2.44E-05
3.30E-05 0.000104 5.80E-05 6.91E-05
̂22
𝚺
2.07E-05 5.80E-05 6.65E-05 9.80E-05
2.44E-05 6.91E-05 9.80E-05 0.000151
𝜋̂23 0.432793
𝛍
̂23 0.04132 0.123603 0.11745 0.165048
8.80E-06 6.44E-06 -1.14E-05 -2.57E-05
̂23
𝚺 6.44E-06 0.000199 5.64E-05 6.49E-05
-1.14E-05 5.64E-05 7.76E-05 0.000117

17
-2.57E-05 6.49E-05 0.000117 0.000194

̂3 )
Table 7. Estimations parameters for the NExcess class ( 𝛉

𝜋̂31 0.413787
𝛍
̂31 0.042417 0.114917 0.113833 0.163167
3.74E-06 -9.6E-06 -4.3E-06 -7.8E-06
-9.6E-06 6.31E-05 3.97E-05 6.94E-05
̂31
𝚺
-4.3E-06 3.97E-05 4.75E-05 7.7E-05
-7.8E-06 6.94E-05 7.7E-05 0.000129
𝜋̂32 0.310345
𝛍
̂32 0.046667 0.142778 0.137556 0.197667
5.78E-06 -3.5E-06 -1.5E-05 -2.2E-05
-3.5E-06 3.77E-05 4.42E-05 6.78E-05
̂32
𝚺
-1.5E-05 4.42E-05 9.49E-05 0.000145
-2.2E-05 6.78E-05 0.000145 0.000222
𝜋̂33 0.275868
𝛍
̂33 0.049125 0.124 0.110375 0.150625
4.86E-06 4.38E-06 7.45E-06 1.1E-05
4.38E-06 0.000171 0.000107 0.000153
̂33
𝚺
7.45E-06 0.000107 9.7E-05 0.000142
1.1E-05 0.000153 0.000142 0.00021

The posterior probabilities were obtained according to Bayes rule for classification using

parameter estimates by the following equations.

28 ̂
𝑃(𝐾 = 1|X = 𝐱) = 84 𝑓1 (𝐱, 𝛉1 )
28 ̂ 27 ̂ 29 ̂
84 𝑓1 (𝐱, 𝛉1 ) + 84 𝑓2 (𝐱, 𝛉2 ) + 84 𝑓3 (𝐱, 𝛉3 )

27 ̂
𝑃(𝐾 = 2|X = 𝐱) = 84 𝑓2 (𝐱, 𝛉2 )
28 ̂ 27 ̂ 29 ̂
84 𝑓1 (𝐱, 𝛉1 ) + 84 𝑓2 (𝐱, 𝛉2 ) + 84 𝑓3 (𝐱, 𝛉3 )

29 ̂3 )
𝑓2 (𝐱, 𝛉
𝑃(𝐾 = 3|X = 𝐱) = 84
28 ̂1 ) + 27 𝑓2 (𝐱, 𝛉
̂2 ) + 29 𝑓3 (𝐱, 𝛉
̂3 )
𝑓1 (𝐱, 𝛉
84 84 84

18
The class assignment matrix from multivariate probability function for experiment samples

was given Table 8.

Table 8. Confusion matrix for training data

Actual classes
Deficient Sufficient Excess
Predict Deficient 27 4 0
classes Sufficient 0 21 3
Excess 1 2 26
Total 28 27 29

It was found that the feature vector X (reflectance rates separated as train data for different N

levels) was assigned to three groups, developed as train data of all reflectances obtained from

84 observations between 400 and 1000 nm in the experiment plants, as NDeficient, NSufficient, and

NExcess with an accuracy of 96.43%, 77.78%, and 89.66%, respectively, and the general correct

classification rate was found “high reliable” according to range of reliability which were

mentioned in the section of testing the reliability of method with a value of 88%.

The developed models predicted the N classes of the experiment plants with high accuracy.

After there the models were tested under field conditions. At this stage, the reflectance values

were considered as a whole without being categorized to N classes. In this way, it was

determined with which the accuracy of the reflectance values (deficient, 45; sufficient, 20; and

excess, 31) of 96 different plants between 400 and 1000 nm were categorized into three

different N classes. For this purpose, the reflectance on 425, 574, 696 and 700 nm for the actual

and prediction classes of test data according to posterior probability values were given Table

9.

Table 9. Actual and prediction Nitrogen classes of the test data in peach leaves based on the
parameter estimation by mix of multivariate normal distributions
Actual Prediction
𝑿𝟒𝟐𝟓 𝑿𝟓𝟕𝟔 𝑿𝟔𝟗𝟔 𝑿𝟕𝟎𝟎 𝑷(𝑲 = 𝟏|𝐗 = 𝐱) 𝑷(𝑲 = 𝟐|𝐗 = 𝐱) 𝑷(𝑲 = 𝟑|𝐗 = 𝐱)
Classes Classes
0.042 0.167 0.176 0.244 1 0.982 0.018 0.000 1
0.041 0.192 0.187 0.267 1 1.000 0.000 0.000 1

19
0.042 0.174 0.176 0.248 1 0.529 0.471 0.000 1
0.035 0.146 0.145 0.213 1 0.976 0.024 0.000 1
0.037 0.219 0.211 0.301 1 1.000 0.000 0.000 1
0.039 0.226 0.214 0.308 1 1.000 0.000 0.000 1
0.039 0.098 0.101 0.140 1 0.998 0.001 0.000 1
0.039 0.137 0.136 0.199 1 0.448 0.084 0.468 3*
0.044 0.150 0.128 0.180 1 0.883 0.117 0.000 1
0.040 0.141 0.133 0.199 1 0.915 0.000 0.085 1
0.036 0.158 0.133 0.191 1 0.844 0.156 0.000 1
0.040 0.099 0.111 0.154 1 0.864 0.135 0.001 1
0.037 0.095 0.103 0.143 1 1.000 0.000 0.000 1
0.040 0.107 0.112 0.154 1 0.948 0.049 0.004 1
0.045 0.101 0.106 0.145 1 0.580 0.211 0.210 1
0.045 0.103 0.107 0.144 1 0.830 0.168 0.002 1
0.040 0.179 0.195 0.276 1 1.000 0.000 0.000 1
0.048 0.167 0.189 0.264 1 1.000 0.000 0.000 1
0.043 0.230 0.202 0.282 1 1.000 0.000 0.000 1
0.036 0.095 0.097 0.135 1 1.000 0.000 0.000 1
0.046 0.168 0.191 0.268 1 1.000 0.000 0.000 1
0.045 0.192 0.215 0.291 1 1.000 0.000 0.000 1
0.049 0.245 0.290 0.394 1 1.000 0.000 0.000 1
0.038 0.095 0.104 0.143 1 1.000 0.000 0.000 1
0.044 0.133 0.117 0.162 1 0.618 0.373 0.009 1
0.046 0.138 0.138 0.201 1 0.438 0.241 0.322 1
0.046 0.163 0.180 0.251 1 0.999 0.001 0.000 1
0.045 0.160 0.168 0.240 1 0.578 0.422 0.000 1
0.049 0.123 0.131 0.185 1 0.043 0.951 0.006 2*
0.042 0.127 0.105 0.146 1 0.020 0.980 0.000 2
0.042 0.100 0.103 0.140 1 0.974 0.016 0.010 1
0.050 0.149 0.126 0.173 1 0.051 0.000 0.949 3*
0.041 0.146 0.145 0.215 1 0.999 0.000 0.001 1
0.044 0.144 0.121 0.170 1 0.635 0.365 0.000 1
0.047 0.229 0.236 0.328 1 1.000 0.000 0.000 1
0.056 0.253 0.261 0.348 1 1.000 0.000 0.000 1
0.044 0.180 0.188 0.264 1 0.304 0.696 0.000 2*
0.045 0.113 0.115 0.155 1 0.155 0.845 0.000 2*
0.049 0.139 0.114 0.161 1 0.000 0.002 0.998 3*
0.056 0.114 0.117 0.155 1 1.000 0.000 0.000 1
0.060 0.092 0.099 0.134 1 0.000 0.000 1.000 3*
0.048 0.146 0.122 0.171 1 0.417 0.028 0.555 3*
0.033 0.162 0.157 0.236 1 1.000 0.000 0.000 1
0.043 0.143 0.118 0.167 1 0.101 0.899 0.000 2*
0.042 0.142 0.144 0.209 1 0.995 0.002 0.003 1
0.042 0.139 0.115 0.160 2 0.082 0.918 0.000 2

20
0.046 0.130 0.134 0.191 2 0.037 0.933 0.031 2
0.042 0.139 0.117 0.162 2 0.037 0.962 0.001 2
0.042 0.105 0.103 0.146 2 0.000 0.934 0.066 2
0.046 0.110 0.111 0.157 2 0.000 0.970 0.030 2
0.052 0.167 0.188 0.268 2 0.008 0.992 0.000 2
0.043 0.112 0.120 0.166 2 0.212 0.760 0.028 2
0.042 0.096 0.098 0.138 2 0.000 0.994 0.006 2
0.042 0.139 0.119 0.162 2 0.000 0.995 0.005 2
0.040 0.112 0.116 0.167 2 0.004 0.596 0.399 2
0.043 0.109 0.108 0.154 2 0.000 0.345 0.655 3*
0.044 0.128 0.113 0.156 2 0.478 0.496 0.025 2
0.041 0.101 0.099 0.142 2 0.004 0.834 0.162 2
0.051 0.129 0.108 0.146 2 0.000 0.529 0.471 2
0.036 0.134 0.133 0.197 2 0.653 0.329 0.019 1*
0.044 0.141 0.126 0.175 2 0.656 0.338 0.007 1*
0.049 0.123 0.126 0.180 2 0.000 1.000 0.000 2
0.042 0.114 0.104 0.148 2 0.000 0.262 0.738 3*
0.035 0.110 0.127 0.181 2 0.002 0.998 0.000 2
0.047 0.131 0.132 0.192 2 0.153 0.751 0.095 2
0.051 0.411 0.434 0.547 3 1.000 0.000 0.000 1*
0.045 0.163 0.173 0.244 3 0.624 0.376 0.000 1*
0.042 0.109 0.108 0.155 3 0.000 0.083 0.917 3
0.037 0.086 0.094 0.134 3 0.476 0.524 0.000 2*
0.044 0.112 0.144 0.200 3 0.763 0.237 0.000 1*
0.041 0.119 0.113 0.165 3 0.020 0.046 0.935 3
0.041 0.110 0.106 0.154 3 0.006 0.309 0.684 3
0.040 0.121 0.125 0.186 3 0.028 0.172 0.801 3
0.047 0.343 0.366 0.478 3 1.000 0.000 0.000 1*
0.037 0.127 0.125 0.187 3 0.287 0.691 0.022 2*
0.038 0.139 0.138 0.205 3 0.479 0.034 0.486 3
0.041 0.088 0.099 0.142 3 0.001 0.999 0.000 2*
0.050 0.355 0.393 0.506 3 1.000 0.000 0.000 1*
0.035 0.111 0.116 0.173 3 0.294 0.706 0.000 2*
0.045 0.133 0.117 0.160 3 0.077 0.600 0.323 2*
0.052 0.137 0.115 0.156 3 0.000 0.000 1.000 3
0.041 0.121 0.121 0.176 3 0.003 0.023 0.973 3
0.048 0.121 0.102 0.139 3 0.000 0.047 0.953 3
0.049 0.137 0.126 0.182 3 0.000 0.000 1.000 3
0.041 0.114 0.110 0.157 3 0.000 0.129 0.870 3
0.049 0.133 0.128 0.183 3 0.000 0.000 1.000 3
0.041 0.108 0.110 0.158 3 0.001 0.352 0.647 3
0.045 0.113 0.112 0.160 3 0.000 0.180 0.820 3
0.046 0.146 0.147 0.212 3 0.013 0.016 0.971 3
0.045 0.140 0.140 0.201 3 0.014 0.046 0.940 3

21
0.049 0.127 0.112 0.152 3 0.006 0.005 0.989 3
0.044 0.103 0.102 0.146 3 0.000 0.158 0.842 3
0.050 0.152 0.143 0.208 3 0.000 0.000 1.000 3
0.046 0.145 0.136 0.194 3 0.000 0.000 1.000 3
0.044 0.117 0.115 0.168 3 0.029 0.128 0.843 3
0.045 0.108 0.110 0.154 3 0.000 0.079 0.921 3
* Incorrectly classified values

Reflectance of the orchard samples were assigned to the N% classes with the highest posterior

probability and was shown in Table 10.

Table 10. Confusion matrix according to mixture discriminant model for test data

Actual class
Deficient Sufficient Excess
Deficient 35 1 2
Predict
class

Sufficient 6 16 8
Excess 4 3 21
Total 45 20 31

The reflectance values of the orchard samples were used as data set for the three different

models. These data were assigned to the three classes, NDeficient, NSufficient and NExcess, with an

accuracy of 77.78%, 80.00% and 67.74%, respectively and the general correct classification

rate was found “satisfactory “ according to range of reliability with a value of 75%. The fact

that the test data (96 plants) randomly selected for all classes and functioned in three different

models were assigned to their classes with satisfying accuracy was significant in terms of the

study objectives. Studies have reported that accuracy coefficient values of ≥0.70 are considered

to be "robust" in the estimation of the N contents of plants using spectral reflectance values and

that they are considered to be “reliable” for the created models.45,84 Based on the results

presented here, the success of the GMDA in studies on the classification of spectral data,85,86

could also be achieved by ground-based hyperspectral measurements. Distributions consisting

of two or more components are called mixture distributions, and mixture distribution models

provide a mathematical approach to creating statistical models for measurements collected of

different aspects of natural events that are fully random in many fields.87 Thus, they are

22
considered to be useful for the analysis of spectral data, since each spectrum can be separated

from a specific spectral data.86,88,89 However, there is a lack of literature on the use of mixture

discriminant models in the categorization and estimation of nutrient levels in plants using

spectral techniques. In fact, stepwise multiple linear regression90, partial least squares

regression,77 and multivariate linear regression,14 models have been especially commonly used

in many studies conducted in this field. It has been suggested that estimates by the models

produced as a result of multiple regression analyses give results with high accuracy coefficients

(r2) in the plots and plant species (in themselves), whereas their predictive power decreases

significantly with the same plant species in different.17,20 The usage of mixture discriminant

model was ensured an assignment of samples in the N classes. Whereas in many studies,

models were set up for a determination of N content. Similarly, it has been suggested that

differential N application could be performed based on leaf spectra and discriminate analysis

(which has a coefficient of determination of 0.94–0.97) under field condition.81 This study was

showed that mixture discriminant models developed using hyperspectral reflectance data can

determine the critical class ranges of N nutrients in plants with high accuracy and be applied

under field conditions. For the farmers, the rapid prediction of the nitrogen classes in the field

conditions can be considered as valuable information as much as the results of laboratory

analysis.

Conclusion

In this study, a non-destructive, low cost, rapid, eco-friendly and practical model was

developed and tested under the field conditions for the determination of N classes of peach

trees. Hyperspectral reflectance data collected using a handheld spectroradiometer and

modeled using GMDA, were considerable accurate for the determination and monitoring of N

classes of peach trees in agricultural lands during the vegetation period. Many studies have

23
been conducted using spectral techniques in different plant species for determining N content

and similarly, the reflectance variations between the 400-750 nm wavelengths have been

previously associated with N content. In the present study, plants were assigned to one of three

critical N classes rather than directly estimating the N content in peach leaves. This assignment

was succeeded with high accuracy using a wider distribution model, created for experiment

samples, and used in different locations. In this way, a different approach has been presented

for the prediction of plant nutrient status using hyperspectral data. In addition, it was found that

the GMDA was a successful and powerful mathematical approach for modeling used. Thus,

the results of this study provide a different perspective for future studies. It can be suggested

that GMDA integrated with hyperspectral remote sensing is useful in the classification of

different nutrient contents. However, it was seen that some values especially NExcess class made

wrong assignments with error rates ranging from 20% to 35% in classification by the model.

This situation may have occurred due to (1) the difference between the number of observations

used in the training and the test groups, and it could be eliminated by increasing the number of

observations, (2) some critical values, which are the same or very close to the limit values of

the class ranges such as 3.40 to 3.60, (3) although the experimental samples were grown under

controlled conditions, some factors which are a necessity of working with living tissue and

which we cannot control such as instantaneous changes of temperature and radiation intensity,

and (4) orchard samples were grown in their natural environment and under traditional fertilizer

- irrigation applications, in which the error rate was expected to increase. However, the findings

showed that reasonable and remarkable rates were obtained ranging from 67.74% to 80.00%

in N classes assignment of randomly selected variables. In addition to these results, further

investigations on mixture discriminant analyses in controlled studies (e.g., nutrient

interactions) using different management systems will increase their reliability.

24
Acknowledgement

I thank Dr. Murat ERİŞOĞLU for statistical comments and advices on the manuscript.

References

1. Jacobsen S, Søndergaard I, Møller B, et al. A chemometric evaluation of the underlying

physical and chemical patterns that support near infrared spectroscopy of barley seeds as a tool

for explorative classification of endosperm genes and gene combinations. J. Cereal Sci 2005;

42: 281-299.

2. He Y, Li X and Deng X. Discrimination of varieties of tea using near infrared spectroscopy

by principal component analysis and BP model. J. Food Eng. 2007; 79: 1238-1242.

3. Wu D, Feng L, He Y, et al. Variety identification of Chinese cabbage seeds using visible

and near-infrared spectroscopy. T ASABE 2008; 51: 2193-2199.

4. Basayigit L, Dedeoglu M and Akgül, H. The prediction of iron contents in orchards using

VNIR spectroscopy. TURK J AGRIC FOR 2015; 39: 123-134.

5. Wang A, Xie L. Technology using near infrared spectroscopic and multivariate analysis to

determine the soluble solids content of citrus fruit. J. Food Eng. 2014; 143: 17-24.

6. Wójtowicz M, Wójtowicz A and Piekarczyk J. Application of remote sensing methods in

agriculture. Commun. Biometry Crop. Sci. 2016; 11: 31-50.

7. Blackburn GA, Ferwerda JG. Retrieval of chlorophyll concentration from leaf reflectance

spectra using wavelet analysis. Remote Sens. Environ. 2008; 112: 1614-1632.

8. Chen J, Li F, Wang R, et al. Estimation of nitrogen and carbon content from soybean leaf

reflectance spectra using wavelet analysis under shade stress. Comput. Electron. Agric. 2019;

156: 482-489.

25
9. Rossel RAV, Walvoort DJJ, McBratney AB. et al. Visible, near infrared, mid infrared or

combined diffuse reflectance spectroscopy for simultaneous assessment of various soil

properties. Geoderma 2006; 131: 59-75.

10. Blackburn GA. Hyperspectral remote sensing of plant pigments. J. Exp. Bot. 2007; 58: 855-

867.

11. Liao W, Ochoa D, Gao L, et al. Morphological Analysis for banana disease detection in

close range hyperspectral remote sensing images. In: IGARSS, Cape Town, South Africa, July

12-17, 2009, Proceedings, ISBN 978-1-4244-3395-7.

12. Chang CW, Laird DA, Mausbach MJ et al. Near-infrared reflectance spectroscopy–

principal components regression analyses of soil properties. Soil Sci. Soc. Am. J. 2001; 65: 480-

490.

13. Pasquini C. Near infrared spectroscopy: fundamentals, practical aspects and analytical

applications. J. Braz. Chem. Soc. 2003; 14: 198-219.

14. Krishna G, Sahoo RN, Singh P, et al. Comparison of various modelling approaches for

water deficit stress monitoring in rice crop through hyperspectral remote sensing. Agric. Water

Manage. 2019; 213, 231-244.

15. Ferrari M and Quaresima V. A brief review on the history of human functional near-infrared

spectroscopy (fNIRS) development and fields of application. Neuroimage 2012; 63: 921-935.

16. Nicolai, BM, Beullens K, Bobelyn E, et al. Nondestructive measurement of fruit and

vegetable quality by means of NIR spectroscopy: A review. Postharvest Biol. Technol. 2007;

46: 99-118.

17. Huang S, Miao Y, Yuan F, et al. Potential of RapidEye and WorldView-2 satellite data for

improving rice nitrogen status monitoring at different growth stages. Remote Sens. 2017; 9:

227.

26
18. Maimaitiyiming M, Ghulam A, Bozzolo A, et al. Early detection of plant physiological

responses to different levels of water stress using reflectance spectroscopy. Remote Sens. 2017;

9: 745.

19. Agelet LE, and Hurburgh CR. A Tutorial on near infrared spectroscopy and its calibration.

Crit. Rev. Anal. Chem. 2010; 40: 246-260.

20. Li X, Liu X, Liu M, et al. A hyperspectral index sensitive to subtle changes in the canopy

chlorophyll content under arsenic stress. Int. J. Appl. Earth Obs. Geoinf. 2015: 36: 41-53.

21. Ayala-Silva, T, and Beyl CA. Changes in spectral reflectance of wheat leaves in response

to specific macronutrient deficiency. Adv. Space Res. 2005; 35: 305-317.

22. Khaled AY, Abd Aziz S, Bejo SK, et al. Spectral features selection and classification of oil

palm leaves infected by Basal stem rot (BSR) disease using dielectric spectroscopy. Comput.

Electron. Agric. 2018; 144: 297-309,

23. Karaçal İ and Tüfenkçi Ş. New approaches on plant nutrition and fertilizer-environment

relationship. In: Agricultural Engineering VII. Technical Congress, TMMOB, Chamber

Agricultural Engineering, Ankara, Turkey, 11 Jan-15 Jan 2010 (pp.11-15)

24. Jackson RD. Remote sensing of biotic and abiotic plant stress. Annu. Rev. Phytopathol.

1986; 24: 265-287.

25. Bagheri N, Ahmadi H, Alavipanah S, et al. Soil-line vegetation indices for corn nitrogen

content prediction. Int. Agrophys. 2012; 26: 103-108.

26. Li F, Miao Y, Hennig SD, et al. Evaluating hyperspectral vegetation indices for estimating

nitrogen concentration of winter wheat at different growth stages. Precis. Agric. 2010; 11: 335-

357

27. Başayiğit L, Dedeoğlu M, Akgül H, et al. Investigation of N deficiency in cherry trees using

visible and near-infrared spectra part of the spectrum in field condition. SPECTROSC SPECT

ANAL 2017; 37: 293-298.

27
28. Min M and Lee W. Determination of significant wavelengths and prediction of nitrogen

content for citrus. Trans ASAE 2005; 48: 455-461.

29. Kokaly RF and Clark RN. Spectroscopic Determination of Leaf Biochemistry Using Band-

Depth Analysis of Absorption Features and Stepwise Multiple Linear Regression. Remote

Sens. Environ. 1999; 67: 267-287.

30. Zhai Y, Cui L, Zhou X, et al. Estimation of nitrogen, phosphorus, and potassium contents

in the leaves of different plants using laboratory-based visible and near-infrared reflectance

spectroscopy: comparison of partial least-square regression and support vector machine

regression methods. Int. J. Remote Sens. 2013; 34: 2502-2518.

31. Sanches IDA, Souza FCR and Kokaly RF. Spectroscopic remote sensing of plant stress at

leaf and canopy levels using the chlorophyll 680 nm absorption feature with continuum

removal. ISPRS J. Photogramm. Remote Sens. 2014; 97: 111-122.

32. Gredilla A, de Vallejuelo SFO, Elejoste N, et al. Non-destructive Spectroscopy combined

with chemometrics as a tool for Green Chemical Analysis of environmental samples: A review.

TrAC, Trends Anal. Chem. 2016; 76: 30-39.

33. Yuan M, Couture JJ, Townsend PA, et al. Spectroscopic determination of leaf nitrogen

concentration and mass per area in sweet corn and snap bean. Agron. J. 2016; 108: 2519-2526.

34. Wang J, Shen C, Liu N, et al. Non-destructive evaluation of the leaf nitrogen concentration

by in-field visible/near-infrared spectroscopy in pear orchards. Sensors 2017; 17: 538.

35. Marschner H, Kirkby E, and Cakmak I. Effect of mineral nutritional status on shoot—root

partitioning of photoassimilates and cycling of mineral nutrients. J. Exp. Bot. 1996; 47: 1255-

1263.

36. Bausch W and Khosla R. QuickBird satellite versus ground-based multi-spectral data for

estimating nitrogen status of irrigated maize. Precis. Agric. 2010; 11: 274-290.

28
37. Heege H, Reusch S and Thiessen E. Prospects and results for optical systems for site-

specific on-the-go control of nitrogen-top-dressing in Germany. Precis. Agric. 2008; 9: 115-

131.

38. Portz G, Molin J and Jasper J. Active crop sensor to detect variability of nitrogen supply

and biomass on sugarcane fields. Precis. Agric. 2012; 13: 33-44.

39. Watt MS, Pearse GD, Dash JP, et al. Application of remote sensing technologies to identify

impacts of nutritional deficiencies on forests. ISPRS J. Photogramm. Remote Sens. 2019; 149:

226-241.

40. Menesatti P, Antonucci F, Pallottino F, et al. Estimation of plant nutritional status by Vis–

NIR spectrophotometric analysis on orange leaves [Citrus sinensis (L) Osbeck cv Tarocco].

Biosyst. Eng. 2010; 105: 448-454.

41. Gomez AH, He Y, and Pereira AG. Non-destructive measurement of acidity, soluble solids

and firmness of Satsuma mandarin using Vis/NIR-spectroscopy techniques. J. Food Eng. 2006;

77: 313-319.

42. Cayuela JA. Vis/NIR soluble solids prediction in intact oranges (Citrus sinensis L.) cv.

Valencia Late by reflectance. Postharvest Biol. Technol. 2008; 47: 75-80.

43. Zude M, Herold B, Roger JM, et al. Non-destructive tests on the prediction of apple fruit

flesh firmness and soluble solids content on tree and in shelf life. J. Food Eng. 2006; 77: 254-

260.

44. Steuer B, Schulz H and Lager E. Classification and analysis of citrus oils by NIR

spectroscopy. Food Chem. 2001; 72: 113-117.

45. Fitzgerald G, Rodriguez D and O’Leary G. Measuring and predicting canopy nitrogen

nutrition in wheat using a spectral index—The canopy chlorophyll content index (CCCI). Field

Crops Res. 2010; 116: 318-324.

29
46. Eitel JUH, Vierling LA, Litvak ME, et al. Broadband, red-edge information from satellites

improves early stress detection in a New Mexico conifer woodland. Remote Sens. Environ.

2011; 115: 3640-3646.

47. Haboudane D, Tremblay N, Miller JR, et al. Remote estimation of crop chlorophyll content

using spectral indices derived from hyperspectral data. IEEE Trans. Geosci. Remote Sens.

2008; 46: 423-437.

48. Me C, Balasundram SK and Hanif AHM. Detecting and monitoring plant nutrient stress

using remote sensing approaches: A review. Asian J. Plant Sci 2017; 16: 1-8.

49. Wan S, Zhang B, Zhan J, et al. In silico cloning and sequence analysis of phospholipase

Dα gene from peach fruit. Agric. Sci. China 2009; 8: 1293-1300.

50. Zhang, C, Zhang B, Yu M, et al. Isolation, cloning, and expression of five genes related to

nitrogen metabolism in peach (Prunus persica L. Batsch). J. Hortic. Sci. 2016; 91: 448-455.

51. Köseoğlu A and Açikgöz V. Determination of iron chlorosis with extractable iron analysis

in peach leaves. J. Plant Nutr. 1995; 18: 153-161.

52. Graeff S, Steffens D and Schubert S. Use of reflectance measurements for the early

detection of N, P, Mg, and Fe deficiencies in Zea mays L. J. Plant Nutr. Soil Sci. 2001; 164:

445-450.

53. Jones JrJB, Wolf B and Mills HA. Plant analysis handbook. A practical sampling,

preparation, analysis, and interpretation guide. 1st ed. USA:Micro-Macro Publishing, Inc,

1991. pp.213.

54. Ferreira MP, Grondona AE, Rolim SBA,et al. Analyzing the spectral variability of tropical

tree species using hyperspectral feature selection and leaf optical modeling. JARS 2013; 7:

073502.

55. Kacar B and Inal A. Plant analysis. 1st ed. Turkey: Nobel publication. 2008. pp.892.

30
56. Ritchie GE, Roller RW, Ciurczak EW, et al. Validation of a near-infrared transmission

spectroscopic procedure: Part B: Application to alternate content uniformity and release assay

methods for pharmaceutical solid dosage forms. J. Pharm. Biomed. Anal. 2002; 29: 159-171.

57. Minitab. Statistical software 2010. Minitab Inc. State College, Pennsylvania, USA.

58. Reynolds D. Gaussian mixture models. In: Li SZ and Jain AK (eds) Encyclopedia of

biometrics. 2nd ed. US: Springer, 2015, pp.827-832.

59. Erisoglu U, Erisoglu M and Erol H. Mixture model approach to the analysis of

heterogeneous survival data. Pak. J. Statist 2012; 28: 115-130.

60. Martinez WL and Martinez AR. Computational statistics handbook with MATLAB. 3th ed.

London: Chapman and Hall/CRC, 2015, p.352.

61. Halbe Z and Aladjem M. Regularized mixture discriminant analysis. Pattern Recognit. Lett.

2007; 28: 2104-2115.

62. McLachlan GJ, Lee SX and Rathnayake SI. Finite mixture models. Annu. Rev. Stat. Appl.

2019; 6: 355-378.

63. Strahler AH. The use of prior probabilities in maximum likelihood classification of

remotely sensed data. Rem. Sen. Envi. 1980; 10: 135-163.

64. Dempster AP, Laird NM and Rubin DB. Maximum likelihood from incomplete data via

the EM algorithm. J. of the Royal Stat. Soc. 1977; 39: 1-22.

65. McLachlan GJ and Krishnan T. The EM algorithm and extensions. 2nd edition. New Jersy:

John Wiley & Sons, 2007.

66. Morrison DG. On the interpretation of discriminant analysis. J. of Mark. Res. 1969; 6: 156-

163.

67. Morrison DG. Discriminant analysis and predictive classification. In Jagdish N Sheth (eds)

Multivariate methods for market and survey. Illinois:Marketing Classics Press, 2011, Chapter

5.

31
68. Tinsley HE and Brown SD. Handbook of applied multivariate statistics and mathematical

modeling. London: Academic press, 2000, p.709.

69. Filella I, Serrano L, Serra J, et al. Evaluating wheat nitrogen status with canopy reflectance

indices and discriminant analysis. Crop Science 1995; 35: 1400-1405.

70. Walczak B. Wavelets in chemistry. In BGM Vandeginste and SC Rutan (eds) Data

Handling in science and technology. Amsterdam: Elsevier, 2000, Vol 22.

71. Karimi Y, Prasher S O, McNairn H, et al. Discriminant analysis of hyperspectral data for

assessing water and nitrogen stresses in corn. Transactions of the ASAE 2005, 48: 805-813.

72. Jain N., Ray SS, Singh JP, et al. Use of hyperspectral data to assess the effects of different

nitrogen applications on a potato crop. Precision Agriculture 2007; 8: 225-239.

73. Gómez-Casero M T, López-Granados F, Peña-Barragán JM, et al. Assessing nitrogen and

potassium deficiencies in olive orchards through discriminant analysis of hyperspectral data. J

Am Soci Hort Sci 2007; 132: 611-618.

74. Lehmann JRK, Große-Stoltenberg A, Römer M, et al. Field spectroscopy in the VNIR-

SWIR region to discriminate between Mediterranean native plants and exotic-invasive shrubs

based on leaf tannin content. Remote Sensing 2015; 7: 1225-1241.

75. Zhang HK, Roy DP, Yan L, et al. Characterization of Sentinel-2A and Landsat-8 top of

atmosphere, surface, and nadir BRDF adjusted reflectance and NDVI differences. Remote Sens.

Environ. 2018; 215: 482-494.

76. Merzlyak MN, Solovchenko AE and Gitelson AA. Reflectance spectral features and non-

destructive estimation of chlorophyll, carotenoid and anthocyanin content in apple fruit.

Postharvest Biol. Technol. 2003; 27: 197-211.

77. Basayigit L, Albayrak S and Senol H. Analysis of VNIR reflectance for prediction of macro

and micro nutrient and chlorophyll contents in apple trees (Malus communis). Asian J. Chem.

2009; 21: 1302-1308

32
78. Fernandez-Martínez J, Joffre R, Zacchini M, et al. Near-infrared reflectance spectroscopy

allows rapid and simultaneous evaluation of chloroplast pigments and antioxidants, carbon

isotope discrimination and nitrogen content in Populus spp. leaves. For. Ecol. Manage. 2017;

399: 227-234.

79. Zhou C, Chen S, Zhang Y, et al. Evaluating metal effects on the reflectance spectra of plant

leaves during different seasons in post-mining Areas, China. Remote Sens. 2018; 10: 1211.

80. Foster AJ, Kakani VG, Ge J, et al. Discriminant analysis of nitrogen treatments in

switchgrass and high biomass sorghum using leaf and canopy-scale reflectance spectroscopy.

Int. J. Remote Sens. 2016; 37: 2252-2279.

81. Penuelas J and Filella I. Visible and near-infrared reflectance techniques for diagnosing

plant physiological status. Trends Plant Sci. 1998; 3: 151-156.

82. Kumar L, Schmidt K, Dury S, et al. Imaging spectrometry and vegetation science. In: Van

der Meer FD and de Jong SM(eds) Imaging spectrometry. 1st ed. Dordrecth: Springer, 2002

pp. 111-155

83. Feng W, Guo BB, Wang ZJ, et al. Measuring leaf nitrogen concentration in winter wheat

using double-peak spectral reflection remote sensing data. Field Crops Res. 2014; 159: 43-52.

84. Ju J, Kolaczyk ED and Gopal S. Gaussian mixture discriminant analysis and sub-pixel land

cover characterization in remote sensing. Remote Sens. Environ. 2003; 84: 550-560.

85. Çalış N and Erol H. A new per-field classification method using mixture discriminant

analysis. J. Appl. Stat. 2012; 39: 2129-2140.

86. Manolakis D, Siracusa C and Shaw G. Hyperspectral subpixel target detection using the

linear mixing model. IEEE Trans. Geosci. Remote Sens. 2001; 39: 1392-1409.

87. Gillis D, Bowles J, Ientilucci EJ, et al. A generalized linear mixing model for hyperspectral

imagery. In: SPIE DEFENSE AND SECURITY SYMPOSIUM. Algorithms and Technologies

33
for Multispectral, Hyperspectral, and Ultraspectral Imagery XIV(ed Shen SS and Lewis PE),

Orlando, Florida, USA, 16 March-20 March 2008, 6966, paper no. 69661B

88. Deng S, Xu Y, Li X, et al. An infinite Gaussian mixture model with its application in

hyperspectral unmixing. Expert Syst. Appl. 2015; 42; 1987-1997.

89. Wang Z, Skidmore AK, Wang T, et al. Canopy foliar nitrogen retrieved from airborne

hyperspectral imagery by correcting for canopy structure effects. Int. J. Appl. Earth Obs.

Geoinf. 2017; 54, 84-94.

34

You might also like