View metadata, citation and similar papers at core.ac.

uk brought to you by CORE

provided by Biblioteca Digital da Produção Intelectual da Universidade de São Paulo (BDPI/USP)

Universidade de São Paulo

Biblioteca Digital da Produção Intelectual - BDPI

Departamento de Física e Ciências Materiais - IFSC/FCM Artigos e Materiais de Revistas Científicas - IFSC/FCM

2014-06

Use of artificial vision techniques for diagnostic

of nitrogen nutritional status in maize plants

Computers and Electronics in Agriculture, Amsterdam : Elsevier BV, v. 104, p. 63-70, June 2014
http://www.producao.usp.br/handle/BDPI/50229

Downloaded from: Biblioteca Digital da Produção Intelectual - BDPI, Universidade de São Paulo
 Computers and Electronics in Agriculture 104 (2014) 63–70

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Use of artificial vision techniques for diagnostic of nitrogen nutritional

status in maize plants
L.M. Romualdo a,⇑, P.H.C. Luz a, F.F.S. Devechio a, M.A. Marin a, A.M.G. Zúñiga b, O.M. Bruno b, V.R. Herling a
a
Department of Animal Science, College of Animal Science and Food Engineering, University of São Paulo (FZEA/USP), Duque de Caxias Norte Street, 225, Post Box 23,
13630-000 Pirassununga, SP, Brazil
b
Department of Physics, Institute of Physics of São Carlos, University of São Paulo (IFSC/USP), Trabalhador são-carlense Avenue, 400, Poste Box 369, 13560-970 São Carlos, SP, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: The identification of the nutritional status of maize by foliar chemical analysis requires sampling of leaves
Received 7 June 2013 when the plant is in an advanced stage of development, hindering corrective action in ongoing cultiva-
Received in revised form 19 March 2014 tion, if deficiency detection of a specific nutrient occurs. An artificial vision system (AVS) is a set of meth-
Accepted 23 March 2014
ods used for analysis and interpretation of images. Therefore, an AVS is being developed to identify
nutrient deficiencies at different stages of plant development, especially in the early stages of growth,
which may contribute to early diagnosis and correction in the same cycle of growth. The objective was
Keywords:
to evaluate methods of digital image processing to develop the AVS to diagnose induced nitrogen defi-
Zea mays
Nutritional diagnosis
ciency in maize leaves. The experiment was done in greenhouse and the treatments were N doses
Macronutrient (0.0; 3.0; 6.0 e 15.0 mMol L1) combined with three growing stages (V4, V7 and R1). The images of maize
Computational vision leaves were digitized in 1200 dpi. After scanning, leaves were chemically analyzed for N content and was
determined the dry mass of plants. The studied methods in AVS were: Volumetric Fractal Dimension
(VFD), Gabor Wavelet (GW) and VFD with canonical analysis. The omission and reduction of nitrogen
in maize plants resulted in typical symptoms of N deficiency. The AVS was able to identify levels of nitro-
gen deficiency in the early stages of development of corn, with global percentage of right of 82.5% at V4
and 87.5% at V7. The GW technique with color images resulted in the better method for features
extraction.
Ó 2014 Elsevier B.V. All rights reserved.

1. Introduction The Nitrogen (N) deficiency symptoms begin as leaf chlorosis,

from tip to base as an inverted ‘‘V’’ pattern; dryness from the tip
The improvement of methods for identification of the nutri- of the old leaves towards the central nervure; necrosis followed
tional status of plants, combined with the need for improvements by leaf tearing, and thinning of stalks (Taiz and Zeiger, 2006;
in the efficient use of nutrients present in soil or applied through Marschner, 2012). Nitrogen is a constituent of all amino acids,
fertilization to obtaining high yields by crops, have led to search amides, proteins, nucleic acids, nucleotides, polyamides and cyto-
for new technologies, of the viewpoints agronomic, economic, chromes, and is part of chlorophyll molecule (Malavolta, 2006).
environmental or operational. Mineral nutrients have essential role Lack of N delay cell division at the growing gems, decreasing leaf
in plant metabolism in such a way that in cases where one of the area and plant size, with prejudice of grain yield (Coelho, 2007).
nutrients is not in adequate amount the whole metabolism is prej- The evaluation of nutritional state of the plants is usually done
udiced. Such disturbs usually can be identified by symptoms in through chemical analysis or visual evaluation. Leaf chemical anal-
plant development, such as chlorosis and necrosis, reduced growth ysis imply sampling at certain phenological stage, which in prac-
and similar anomalies (Malavolta, 2006). tice does not allow to take remediation actions for the crop being
cultivated. Visual diagnose is subjected to errors in interpretation
because visual symptoms can occur simultaneously and also be
⇑ Corresponding author. Address: Sorocaba Street, 985, 14075-130 Ribeirão Preto, confounded with pest attack or disease. Therefore, visual diagnose
SP, Brazil. Tel.: +55 19 35654267. depends strongly on the operator’s experience. The use of biomet-
E-mail addresses: [email protected] (L.M. Romualdo), phcerluz ric measurements of leaves, related to the shape, area, texture and
@usp.br (P.H.C. Luz), [email protected] (F.F.S. Devechio), mario.marin@ ribs, among others, provide quantitative, objective and accurate
timacagro.com.br (M.A. Marin), [email protected] (A.M.G. Zúñiga), bruno@ information that can be employed in the study of plant nutrition.
ifsc.usp.br (O.M. Bruno), [email protected] (V.R. Herling).

http://dx.doi.org/10.1016/j.compag.2014.03.009
0168-1699/Ó 2014 Elsevier B.V. All rights reserved.
64 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70

In this context, computational models of artificial vision and math- (FZEA/USP) at the Pirassununga-SP campus. The crop tested was
ematical models can contribute to a more detailed analysis of foliar maize (Zea mays L.), hybrid DKB 390Ò, grown in a hydroponic sys-
structures through benchmark techniques and information extrac- tem in nutrient solution.
tion (Backes and Bruno, 2008). Experimental design was fully random, in a 4  3 factorial with
An artificial vision system or computational vision may be four replications. The factors established for the study consisted of
defined as a set of methods and techniques through computer sys- four levels of N concentration in nutrient solution: T1 = complete
tems, which are able to interpret an image automatically or semi- omission (0 mMol L1), T2 = 20% of full level (3.0 mMol L1),
automatically (Punam and Udupa, 2001). And are constituted in T3 = 40% of full level (6.0 mMol L1) and T4 = full level
the steps: acquisition, image segmentation, feature extraction (15 mMol L1), combined in three time sampling of collecting
and classification/identification (Bruno, 2000; Gonzales and leaves (phenological stage), when plants were with four fully
Woods, 1993). In literature there are several methods of taxono- expanded leaves (V4), seven leaves (V7) and silking (R1), totaling
mies for image segmentation. However, the most traditional is 48 experimental units (4 level of N  3 sampling stage  4 replica-
adopted by Gonzales and Woods (1993), which defines three cate- tions). The concentration of elements in nutritional solutions was
gories of segmentation: thresholding, based segmentation on determined after a previous trial having the Hoagland and Arnon
edges and based on regions. Other methods are based on colors (1950) solution as reference.
(Lim and Lee, 1990; Gonzalez et al., 1990; Moreira, 1999), the ones Seeding was done in plastic trays filled with washed sand until
using neural networks and fuzzy logic (Bezdek et al., 1999) and two weeks after emergence. After that, two plants were transferred
those based on genetic algorithms (Ankenbrandt et al., 1990). to 3.5 L pots (experimental units) and supported by a piece of foam
Chromatic analysis is considered the information concerning the and maintained there until stage V7, when they were transferred
color. Among the methods used in this category of segmentation, to 10 L pots until stage R1. After transplanting, plants were kept
can highlight the analysis of chromatic histograms (Cheng et al., in a 50% diluted nutrient solution for them to adapt, during 5 days.
2001), chromatic moments (Stricker and Orengo, 1995) and set Nutrient solutions were replaced 15 days after transplanting, and
of colors on mappings (Cheng et al., 2001). The segmentation pro- henceforward weekly. When needed, pot solution was leveled with
cess selects regions for images analysis. The feature extraction step deionized water and pH adjusted between 5.0 and 6.0 with HCl 1 N
consists of methods to assess the regions of the images for their and NaOH 1 N. Each pot had air bobbling during 10 s at each 30 s
characterization. This step, the texture analysis considers the dis- interval.
tribution and organization of pixels in a certain image region. In In each collecting times established, 16 pots (experimental
this context, methods that use measures like fractal dimension units) were sampled (sample destructive). Sampled material was
and lacunarity can measure the complexity of images, and gener- split in (a) above ground portion (shoot, Sh); (b) root (Rh); (c) index
ate digital signatures to characterize the image (Backes and leaf (IL) of the growing stage (V4 = leaf 4; V7 = leaf 7, and
Bruno, 2008). R1 = opposite leaf below the first ear) and (d) old leaf (OL). At all
The application and use of image analysis is utilized in agricul- sampling times (V4, V7 and R1), the OL collected in the corn plant
ture for some time now, mostly in precision agriculture. Yang et al. corresponded to older leaf that was not senescent.
(2000) applied artificial neural networks (ANN), trained with the The shoots, index leaves and old leaves after collection of
back-propagation algorithm, to the development of a model capa- images, were rinsed in deionized water. The roots were washed
ble of distinguishing young maize plants of weeds plants. Burks in running water and then were washed in the following order:
et al. (2005) used the same algorithm to study the recognition of water with a neutral detergent, deionized water, hydrochloric acid
weeds plants. Baesso et al. (2007) assessed the use of spectral indi- (HCl) diluted in deionized water and finally deionized water. Pos-
ces, taken from digital images, to discriminate different doses of teriorly, the plant parts were dried in an oven with forced air circu-
nitrogen in common bean. Sena Júnior et al. (2008) identified lation at a temperature of 65 °C, for approximately 72 h, to
through image analysis nutritional stages of wheat plants. Silva determine the dry mass and then grind with 2 mm screen and
et al. (2014) evaluated different methods for feature extraction in saved in plastic bags for further analyses of N determination
images of maize leaves in the V4 stage, grown in greenhouse under according to methodology described in Bataglia et al. (1983).
nutritional deficiency induced of magnesium. With the obtained
results, it was found global percentage of right 76% with reliable 2.2. Statistical analysis for dry mass and N concentration
Kappa index. Silva Júnior et al. (2012) determined the percentage
of vegetation cover of weeds plants in the crop beans, under the Statistical analysis model below was accomplished for the dry
no-tillage and conventional, using digital image processing and mass yield and accumulation of nitrogen in shoot and root of maize
geostatistics. Abrahão et al. (2013) conducted a study of classifiers plants, and N concentration in index leaves. Results were statisti-
based on different combinations of bands and spectral indices of cally processed using variance analysis. According to Steel et al.
original images to discriminate foliar nitrogen and chlorophyll, also (1997) the statistical model used was as follows
in the crop beans.
Yijk ¼ m þ Ei þ Nj þ ENij þ eijk ð1Þ
Being so, the use of image processing in agriculture can be an aux-
iliary important tool in the soil and plant management. The present In the model, Yijk is the observed value at parcel that received ij
research was focused in the development of an AVS for N diagnosis in treatment on repetition k; m is the average; Ei is the effect of the
maize plants, for further validation in field. The objective of this development stage of maize; Nj is the effect of nitrogen levels at
study is to present an evaluation of methods of image processing parcel; eijk is the effect of uncontrolled factors at parcel that
used to identify and evaluate induced N deficiency in maize. received ij on repetition k. When the F test was significant
(P 6 0.05) only for Nj, has been done only one polynomial regres-
sion analysis for all stages. When the F test was significant
2. Material and methods (P 6 0.05) for ENij, in the other words, when there was interaction
between the levels of N applied and the developmental stages,
2.1. Experiment in greenhouse the splitting aimed to study the levels within of the Ei for the N con-
tent in the index leaves. And for the accumulated N in different
The experiment was done in a greenhouse of the Animal Science plant parts were studied developmental stages within each level
Department of the College of Animal Science and Engineering Food of N (Nj) in the nutrient solution.
 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70 65

2.3. Artificial visual system cultivated in each experimental unit, 8 IL and 8 OL were obtained
for each treatment.
In Fig. 1 presents a diagram of the proposed system, with steps In each leaf sampling event, three areas were considered to
steps of images analysis according Gonzales and Woods (1993). image crop in each leaf: bottom, medium third and tip, both for
The proposed system consists in four main parts: early processing, IL and OL (Fig. 2). The leave image is divided into three equal parts
texture sample generator, texture feature extraction and pattern and the 5% of the boundary region is not considered to avoid over-
recognition. Several leaves are scanned by the same time using a lap. At the end, 24 classes were created and 50 images were
scanner, providing a digital image. In the first step of the system, obtained for each of these classes. Therefore, each class had images
the early processing part consists in algorithms to remove the from one area of the leaf for each level o N supply of IL and OL. In
background and after that, the foreground leaves have to be split this way, there were 40 images for training and 10 for test for each
and oriented horizontally. For each leave, a texture window is class, and in total, 960 images for training and 240 for test. From
taken. In this task, a large number of small windows are extracted each of those scanned area, windows with 80  80 pixels were ori-
from the leave image. For each leave, a texture window is taken. In ented to the horizontal position (Fig. 3). From each image, win-
this task, a large number of small windows are extracted from the dows with 80  80 pixels from specific spots where the
leave image. The samples pass by statistical analysis and if they are deficiency symptoms are clearer were extracted. Such image win-
not representative taking into account the whole leave, it is disre- dows were saved without file compression.
garded. To check the samples are representative, it is compared as According to Zúñiga (2012), color and texture are the main
a mean of all windows obtained from the leave. These tasks (win- visual characteristics associated to maize nutrition. The following
dows extractions and selection) are performed automatically by methods proposed by Zúñiga (2012) for AVS were tested in the
the ‘‘Texture sample generator module’’. On the ‘‘texture feature present study for the texture analysis: Volumetric Fractal Dimen-
extraction’’ module, texture analysis methods are applied into each sion (VFD), Gabor Wavelet (GW) and VFD with canonical analysis
window and the feature vector is created by each one. On the last (VDFCA).
part of the system, the feature vectors can be used for training the The VDF works with binary images but was adapted here to
model or to yield a report, checking in which class it belongs. It is work with the images captured by the scan, as proposed by
done using a supervised classification model based on the naive Backes et al. (2009), in which the image signature is calculated to
Bayes method. work with all reE values, where E is the set of Euclidean distances
Leaves were scanned in a regular desktop scanner (HP Scanjet for a maximum radius rmax. In such method the radius ranged from
3800) at 1200 dpi resolution and saved in tiff format. In each sam- 1 to 20.
pling event (according to plant stage), IL and OL of 16 pots were pffiffiffi pffiffiffi
collected (4 treatments and 4 replications). Since two plants were E ¼ 1; 2; 3; . . . ; r max ð2Þ

Fig. 1. Image analysis procedure used in this study.

66 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70

Fig. 2. Separation of leaves for leaves for image analysis considering three parts: tip, middle and bottom.

sion matrices were generated to assess the amount of right classi-

fications made by AVS, and also to know classes that were difficult
to classify. In addition, were assessed the percentage of images cor-
rectly classified or Global Percentage of Right (GPR) and Kappa
index (K). According to Everitt and Dunn (2001), the Kappa index
indicates the correlation between GPR and truth. This index is used
in this study to measure the confidence of classification. The Kappa
index is evaluated as follows: 0–0.2: not trust; 0:21–0:41: low;
0.41–0.60: moderate; 0.61–0.8: trust; 0.81–1.0: worthy trust.

3. Results and discussion

3.1. Visual symptoms

Plants subjected to lack of N showed chlorosis in early leaves

which evolved to drying and necrosis. These leaves did not elon-
gate and plants did not evolved (Fig. 4a). Similar symptoms were
observed by Ferreira (2012) in the hybrid maize BRS 1010 culti-
vated in a sand plus vermiculite substrate at 1:1 ratio, in the com-
plete lack of N.
Strong chlorosis in OL, from tip to base and in inverted ‘‘V’’
Fig. 3. Segmentation process. shape was observed in plants cultivated with 3.0 mMol L-1 of N
(Fig. 4b). Such scenario evolved to tip drying advancing through
pffiffiffi pffiffiffi the central nervure and finally necrosis and tearing, as well as thin
fðr max Þ ¼ ½log Vð1Þ; log Vð 2Þ; log Vð 3Þ; . . . ; Vðr max Þ; ð3Þ stalks. Such typical symptoms were also observed in plants grown
in 40% of N, after the V7 stage.
The Gabor bi-dimensional transform is a bi-dimensional gaussian
function modulated into a senoidal oriented with a certain fre-
quency W and in certain direction h. Its bi-dimensional form in 3.2. Nitrogen in the plant and dry mass yields
the space domain g(x, y) and frequency G(u, v) is given by the follow-
ing equations: From the results obtained, it was found that there were signif-
! " !# icant effects (P < 0.01) of the levels, times and interaction on the
1 1 x2 y2 nitrogen concentration in index leaves, N accumulated and dry
gðx; yÞ ¼ exp  þ þ 2pjWx ð4Þ matter yields in shoot and root, with the increment of nitrogen
2prx ry 2 rx2 r2y
in nutrient solution (Figs. 5 and 6).
(" # The N concentration in IL at the stages V4, V7 and R1 as a func-
1 ðu  WÞ2 v 2 tion of doses is showed in Fig. 5. It can be observed that the great-
Gðu; v Þ ¼ exp  þ 2 ð5Þ
2 r2u rv est N concentration was observed during the stage V4 with the N
dose of 9.0 mMol L1. For V7 and R1, the maximum concentration
The Gabor transformed can be adapted as a wavelet. In such a case of N was around 12.0 mMol L1. According to Malavolta (2006),
these equations are used as a mother wallet. In the next step, a filter adequate levels of N in IL at R1 for maize should be between 28
dictionary can be obtained by dilation and rotation and gz(x, y) and 35 g kg1. The N values at V4 (63 g kg1) should be related
through the function generated as proposed by Manjunath and to the fact that leaves are small at this stage, but such value does
Ma (1996): not relate to N accumulation in terms of dry mass, as can be seem
g mn ðx; yÞ ¼ am gðx0 ; y0 Þ ð6Þ in Fig. 4a which shows that at 24 days (V4) the smallest N accumu-
lations were observed for all doses. According to Büll (1993) the
where a > 1 and m, n refers to the scale and orientation, with percent content of N in young plant tissues of maize is greater than
m = 0,1, ... , M – 1en = 0, 1, ... , N – 1 M is the total number of scales in the other plant growing stages, although the plant requirement
and N is the total number of orientations. for such nutrient is small due to the short size of the plants.
For all methods the Naive Bayes classification and the cross val- The V4 stage of development was the time that had lower dry
idation learning method were used. For the evaluation, the samples mass yields, with maximum production of shoot and root of
were separated randomly into n groups of roughly equal size and 2.84 e 0.96 g planta1 respectively, obtained with the level of N
was made ‘‘to let an outside group’’ the cross-validation which in the nutrient solution of about 10.5 mMol L1. In V7 was
can also be called a ‘‘n-fold cross-validation’’ test scheme. Samples 28.4 g planta1 (shoot) and 12.32 g planta1 (root), obtained with
were independent for each class, and these samples did not appear the level of N in the nutrient solution of 13.4 e 10.2 mMol L1,
in the same training and testing. In each processing, 80% (960 respectively. The tasseling (R1 development stage) was time that
images, 40 images per class) of the images were used for training showed the highest biomass yield in the plants, with a maximum
and 20% (240 images, 10 per class) for testing ‘‘blind’’. The confu- of 75.4 g planta1 in shoot and 26.4 g in root, obtained with the
 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70 67

a b

Fig. 4. Nitrogen deficiency symptoms in maize plants subjected to complete lack of N at 24 days after showing (a), and plants grown with 20% N at 35 days after emergency
(b).

3.3. Artificial visual system (AVS)

Tables 1 and 2 are shown the global percentage of right (GPR)

and confidence in the classification (Kappa index) for the tech-
niques studied (VFD, GW and VDFCA) for feature extraction in
leaves of maize in different stages of crop development (V4, V7
and R1).
Based on the results was observed that the GW technique
applied to color images was the one with the best performance,
with 82.5% rights for V4 at the bottom of index leaves (Table 1),
87.5% rights for V7 using the image windows at the bottom of
old leaves (Table 2) and 98% of rights for R1 using the third middle
of index leaves (Table 1). The confidence in the classification
(Kappa index) with the use of this technique had an index above
0.91, therefore classified as worthy trust (Tables 1 and 2). Accord-
ing to Table 1 and 2, the GW technique using gray scale had GPR of
94% and Kappa index of 0.99 (worthy trust), but with the use of
Fig. 5. Nitrogen concentration in index leaves of maize plants submitted to
color images for this same technique, for features extraction, found
nitrogen levels at different stages of development.
results even better (GPR 98%). These results confirm the impor-
tance of color information in this study in order to characterize
the nutritional symptoms on leaves of corn. Similar results were
N level in the nutrient solution of respectively 10.4 e observed by Silva et al. (2014) using the same methods of this
11.15 mMol L1 (Fig. 6a, e and b). study to patterns recognize of nutrient deficiencies for magnesium.
According to the results, it is observed that the omission and The confusion matrices for GW in each growing stage of maize are
reduction of N compromised the dry mass of maize plants during shown in Table 3.
the studied period. The nitrogen content of the plant influences The confusion matrices show in their diagonal the amount of
the distribution pattern of carbohydrates within the plant, which images that were correctly classified in each set of 50 images. In
affects growth and crop yield, and therefore there is a strong corre- the lines, it can be seen the amount of images that the AVS
lation between N content and biomass production (Lawlor, 2002). assigned to another class (wrong of interpretation).
Accumulation of N can be observed in the results showed in The confusion matrices of the best results obtained with the
Fig. 6 in the shoot (c) and root (d). The greatest amount of N was bottom of old leaves (Table 3), indicate that the complete omission
observed at day 66 (R1 stage), in plants cultivated in the full dose (zero level) was the simplest class rating, with rights of 96% and
of N (Fig. 6c, e and d). According to Coelho (2007), N requirement 98% for V4 and V7 times respectively, and the class corresponding
by plants varies enormously during the growing stages of a plant to the level of 6.0 mMol L1 of N (T3) had the highest wrong of clas-
but reaches a maximum during the beginning of grain formation. sification in V4 stage. The full level and 3.0 mMol L1 of N (T1 e T2)
Duarte et al. (2003) and Pinho et al. (2009) also reported small N showed good distinction between them and between other classes.
accumulation of N during the initial stages of growth, and a strong The results also indicate that it is possible to obtain a good separa-
increase at 42 days and from 42 days forward a constant and linear tion of classes considering the zero and 3.0 mMol L1 levels, which
increase until the final stages of the development of maize. How- are the levels of severe and moderate deficiency in the plant.
ever, as noted in the dry matter yield (Fig. 6a, e and b), at 24 Almost all classes had the greatest number of errors in the T3 class.
DAE (V4), the inadequate supply of nitrogen affected the develop- That can be related to the fact that the leaves in this treatment had
ment of corn and consequently the biomass production. According images very similar to the full level of N and to the T2
to Coelho (2007), nitrogen deficiency, when the plant has a height (3.0 mMol L1) during the V4 stage (Table 3).
of about 20 cm, it will cause reduction in the number of grains, It is worth to highlight that the levels 3.0 and 6.0 mMol L1 are
compromising the final production. the most difficult to be distinguished by the human eye, because
68 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70

Fig. 6. Dry mass yield in shoot (a) and root (b); nitrogen accumulation in shoot (c) and root (d) of maize plants submitted to nitrogen levels at different stages of development.

Table 1
Global Percentage of Rights (GPR) of color and gray scale images using the techniques Volumetric Fractal Dimension (VFD), VFD with canonical analysis (VDFCA) and Gabor
Wavelet (GW), and corresponding Kappa index (K), applied to tip, bottom and middle third (middle) of index leaves of maize plants at stages V4, V7 and R1 grown in different N
doses.

Techniques studied on AVS

VFD VFDCA GW VFD VFDCA GW
Color images Gray scale images
V4 V4
GPR K GPR K GPR K GPR K GPR K GPR K
Tip 70.5 0.87 73.5 0.88 80.0 0.90 64.0 0.85 77.5 0.92 72.0 0.91
Bottom 65.0 0.86 75.0 0.90 82.5 0.95 55.0 0.84 75.0 0.93 71.5 0.91
Middle 65.0 0.87 66.0 0.86 70.0 0.90 52.5 0.79 71.5 0.88 73.5 0.90
V7 V7
Tip 59.0 0.80 79.0 0.93 74.5 0.90 54.0 0.77 70.0 0.87 61.5 0.83
Bottom 50.0 0.76 74.0 0.91 75.0 0.92 53.0 0.77 68.0 0.88 65.0 0.85
Middle 62.5 0.81 70.5 0.89 75.0 0.91 56.0 0.78 70.5 0.88 64.5 0.84
R1 R1
Tip 84.0 0.96 86.5 0.95 85.5 0.98 62.0 0.86 85.5 0.95 86.5 0.97
Bottom 96.5 0.99 92.0 0.97 98.0 0.99 74.5 0.89 91.0 0.98 94.0 0.99
Middle 89.0 0.98 84.0 0.93 91.5 0.98 66.0 0.86 82.0 0.94 85.5 0.95

these levels of N are close to the full dose and therefore, the symp- actions to be taken to rescue the crop from future failure. Although
toms are less perceptible. Even in such cases, the percent rights the GW technique applied to color images of V4 had the best per-
were 76% for the 3.0 mMol L1 and 66% for the 6.0 mMol L1, for formance (bottom of leaves) (Table 1), similar percent of rights was
OL, which can be considered very promising. achieved from OL with the same technique (Table 2). These results
The results in Tables 3 suggest that the detection of N deficiency agree with the literature which describes the symptoms of defi-
in the early stages of maize development is possible, and also that a ciency appearing firstly in old leaves due to N mobility in plants.
good number of rights can be achieved for the levels T1 and T2, The R1 stage showed the Kappa considered ‘‘worthy trust’’ and
which are the severe and moderate levels of N deficiency. The typ- better GPR in all parts of plants. The percent of rights is also very
ical symptoms of N deficiency were obtained with the reduction of close between IL and OL, for the R1 stage (Tables 1 and 2). The
nitrogen in the nutrient solution, which allowed the recognition of AVS results pointing to the IL as having the greatest number of
the nutritional patterns of corn by AVS. rights can be related to the fact that, with the evolution of pheno-
The high number of rights in stage V4 is central to this study, logical stage, there was a better differentiation between classes,
since the early identification of deficiency is essential to allow which can be correlated with the results observed in deficiency
 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70 69

Table 2
Global Percentage of Rights (GPR) of color and gray scale images using the techniques Volumetric Fractal Dimension (VFD), VFD with canonical analysis (VDFCA) and Gabor
Wavelet (GW), and corresponding Kappa index (K), applied to tip, bottom and middle third (middle) of old leaves of maize plants at stages V4, V7 and R1 grown in different N
doses.

Techniques studied on AVS

VFD VFDCA GW VFD VFDCA GW
Color images Gray scale images
V4 V4
GPR K GPR K GPR K GPR K GPR K GPR K
Tip 60.5 0.81 61.5 0.84 74.0 0.94 55.5 0.77 65.5 0.86 61.5 0.83
Bottom 62.5 0.81 68.5 0.87 82.0 0.95 46.5 0.74 73.0 0.90 64.0 0.87
Middle 61.0 0.79 55.5 0.84 80.5 0.94 48.5 0.77 69.5 0.88 66.0 0.86
V7 V7
Tip 64.0 0.87 78.5 0.94 73.0 0.92 55.5 0.75 74.0 0.91 62.0 0.85
Bottom 71.5 0.89 84.5 0.94 87.5 0.96 51.5 0.80 78.0 0.93 76.5 0.92
Middle 74.5 0.91 79.5 0.96 79.5 0.96 65.0 0.85 77.5 0.94 82.0 0.94
R1 R1
Tip 89.0 0.98 81.0 0.94 85.5 0.96 52.0 0.73 82.5 0.95 74.5 0.93
Bottom 77.0 0.93 76.0 0.91 92.0 0.97 49.0 0.72 73.5 0.92 76.0 0.92
Middle 77.5 0.92 75.5 0.90 83.0 0.97 39.0 0.67 72.5 0.90 76.0 0.88

Table 3
Confusion matrix of classification result from bottom of old leaves at stage V4 and V7, and index leaves at R1, of maize grown in different N doses.

Real classification % Rights

V4 V7 R1
T1 T3 T3 T4 T1 T2 T3 T4 T1 T2 T3 T4
T1 – 0 mMol L1 96.0 2.0 2.0 0.0 98.0 2.0 0.0 0.0 98 0.0 0.0 2.0
T2 – 3 mMol L1 0.0 76.0 24.0 0.0 0.0 68.0 30.0 2.0 0.0 98 2.0 0.0
T3 – 6 mMol L1 8.0 36.0 56.0 0.0 0.0 0.0 90.0 5.0 0.0 0.0 96 4.0
T4 – 15 mMol L1 0.0 0.0 0.0 100 0.0 6.0 0.0 94.0 0.0 0.0 0.0 100

Bold numbers mean percentage of rights within their respective classes corresponding to the N doses.

symptoms in plants, with reflections on the development and archip of FAPESP (State of São Paulo Research Foundation) (Grants
nutrition of corn at this stage. It is noteworthy that for maize, R1 2010/18233-3 and 2009/51835-0).
is the time that best reflects the nutritional status of the crop, so
that the index leaves this stage are drains for grain filling, and is
the standard leaf for analytical determinations for nutritional References
diagnosis.
The results presented here point to the AVS using the GW Abrahão, S.A., Pinto, F.A.C., Queiroz, D.M., Santos, N.T., Carneiros, J.E.S., 2013.
method on color images as being the best for the N status in maize Determination of nitrogen and chlorophyll levels in bean-plant leaves by using
spectral vegetation bands and indices. Rev. Ciênc. Agron. 44 (3), 464–4173.
plants, particularly the IL at V4 stage (82.5%) corresponding to Ankenbrandt, C.A., Buckles, B.P., Petry, F.E., 1990. Scene recognition using genetic
early development stage of corn, but still requiring field trials to algorithms with semantic nets. Pattern Recogn. Lett. 11, 285–293.
validate such result. Backes, A.R., Bruno, O.M., 2008. Fractal and multi-scale fractal dimension analysis: a
comparative study of bouligand-minkowski method. J. Comput. Sci. 7 (2), 74–
83.
4. Conclusion Backes, A.R., Casanova, D., Bruno, O.M., 2009. Plant leaf identification based on
volumetric fractal dimension. Int. J. Pattern Recog. Artific. Intell. 23 (6), 1145–
1160.
Maize plants replied to N levels with respect to symptomatol- Baesso, M.M., Pinto, F.A.C., Queiroz, D.M., Vieira, L.B., Alves, E.A., 2007.
ogy, N concentration in plant tissue and biomass production. Determinação do ‘‘status’’ nutricional de nitrogênio no feijoeiro utilizando
imagens digitais coloridas. Eng. Agríc. 27 (2), 520–528.
The best method for feature extraction in leaves of maize was Bataglia, O.C., Furlani, A.M.C., Teixeira, J.P.F., Furlani, P.R., Gallo, J.R., 1983. Método
Gabor Wavelets in color images. The AVS was able to identify the de análises química de plantas. Boletim Técnico, 78, Instituto Agronômico,
induced levels of nitrogen deficiency in the early stages of develop- Campinas, 48p.
Bezdek, J.C., Keller, J., Krisnapuram, R., Pal, N.K., 1999. Fuzzy Models and Algorithms
ment of corn, so that the bottom of index leaves showed of global for Pattern Recognition and Image Processing. The Handbooks of Fuzzy Sets, 4.
percentage of right (GPR) of 82.5% at V4 stage and 87.5% in the bot- Kluwer Academic Publishers, Boston.
tom of old leaves at V7. However, the middle section of the index Bruno, O.M., 2000. Paralelismo em visão natural e artificial. PhD Thesis. University
of Sao Paulo, Physics Institute of Sao Carlos, Brazil.
leave at the R1 stage was the best GPR (98%) to detect the symp- Büll, L.T., 1993. Nutrição mineral do milho. In: Cultura do milho: fatores que afetam a
toms of nitrogen deficiency by AVS. produtividade. Büll, L.T., Cantarella, H. (Eds.), POTAFOS, Piracicaba, pp. 63–145.
The main idea presented in this work could be adopted for oth- Burks, T.F., Shearer, S.A., Heath, J.R., Donohue, K.D., 2005. Evaluation of neural-
network classifiers for weed species discrimination. Biosyst. Eng. Kidlington 91
ers plants, but it should be pointed out that the performance of the
(3), 293–304.
system is valid for the conditions under which this experiment was Cheng, H.D., Jiang, X.H., Sun, Y., Wang, J., 2001. Color image segmentation: advances
conducted. and prospects. Pattern Recogn. 34, 2259–2281.
Coelho, A.M., 2007. Manejo da adubação nitrogenada na cultura do milho. Circular
Técnica, 96, Embrapa Milho e Sorgo, Sete Lagoas, 11p.
Acknowledgments Duarte, A.P., Kiehl, J.C., Camargo, M.A.F., Reco, P.C., 2003. Acúmulo de matéria seca e
nutrientes em cultivares de milho originárias de clima tropical e introduzidas
de clima temperado. Rev. Bras. de Milho e Sorgo 2 (3), 1–20.
Liliane Maria Romualdo and Pedro Henrique de Cerqueira Luz Everitt, B.S., Dunn, G., 2001. Applied Multivariate Data Analysis, second ed. London
gratefully acknowledges the financial research support and schol- Arnold, London.
70 L.M. Romualdo et al. / Computers and Electronics in Agriculture 104 (2014) 63–70

Ferreira, M.M.M., 2012. Sintomas de deficiência de macro e micronutrientes de Punam, K.S., Udupa, J.K., 2001. Optimum image thresholding via class
plantas de milho híbrido BRS 1010. Rev. Agro@mbiente 6 (1), 74–83. uncertainty and region homogeneity. IEEE Trans. Patt. Anal. Mach. Intell.
Gonzales, R.C., Woods, R.E., 1993. Digital Image Processing. Addison-Wesley 23 (7), 689–706.
Publishing Company, New York. Sena Júnior, D.G., Pinto, F.A.C., Queiroz, D.M., Santos, N.T., Khroury Júnior, J.K.,
Gonzalez, M., Benitez, D., Suarez, C.P., 1990. Segmentation and recognition in visual 2008. Discriminação entre estágios nutricionais na cultura do trigo com
chromatic spaces. Cybern. Syst. 21, 241–247. técnicas de visão artificial e medidor portátil de clorofila. Eng. Agríc. 28
Hoagland, D.R., Arnon, D.I., 1950. The Water Culture Method for Growing Plants (1), 187–195.
Without Soils. California Agricultural Experimental Station, Berkeley. Silva Júnior, M.C., Pinto, F.A.C., Queiroz, D.M., Gómez, J.G., Navas, L.M.G., 2012.
Lawlor, D.W., 2002. Carbono and niotrogen assimilation in relation to yield: Weed mapping using a machine vision system. Weeds Plants 30 (1), 217–
mechanisms are the key to understanding production systems. J. Exp. Botany. 227.
53, 773–787. Silva, F.F., Luz, P.H.C., Romualdo, L.M., Marin, M.A., Zúñiga, A.G., Herling, V.R., Bruno,
Lim, Y.W., Lee, S.U., 1990. On the color image segmentation algorithm based on the O.M., 2014. A diagnostic tool for magnesium nutrition in maize based on image
thresholding and the fuzzy c-means techniques. Pattern Recogn. 23 (9), 935– analysis of different leaf sections. Crop Sci. 54 (2014), 1–8.
952. Steel, R.G.D., Torrie, J.H., Dickey, D.A., 1997. Principles and procedures of statistics. A
Malavolta, E., 2006. Manual de nutrição de plantas. Agronômica Ceres, São Paulo. Biometrical Approach, third ed. McGraw Hill Co, New York, 666p.
Manjunath, B.S., Ma, W.Y., 1996. Texture features for browsing and retrieval of Stricker, M., Orengo, M., 1995. Similarity of color images. In: Process SPIE Storage
image data. IEEE Trans. Patt. Anal. Mach. Intell. 18 (8), 837–842. and Retrieva, for Image and Video Databases, pp. 381–392.
Marschner, P., 2012. Mineral Nutrition of Higher Plants, third ed. Academic Press, Taiz, L., Zeiger, E., 2006. Plant Physiology. Sinauer Associates, Sunderland.
London. Yang, C.C., Prasher, S.O., Landry, J.A., Ramaswamy, H.S., DiTommaso, A., 2000.
Moreira, J., 1999. Uma proposta de estruturação e integração de processamento de Application of artificial neural networks in image recognition and classification
cores em sistemas artificiais de visão. PhD Thesis. University of Sao Paulo, of crop and weeds. Can. Agr. Eng. 42 (3), 147–152.
Physics Institute of Sao Carlos, Brazil. Zúñiga, A.M.G., 2012. Sistema de visão artificial para identificação do estado
Pinho, R.A.V., Borges, I.D., Pereira, J.L.A.R., Reis, M.C., 2009. Marcha de absorção de nutricional de plantas. MSc. Dissertation. University of Sao Paulo, Mathematics
macronutrientes e acúmulo de matéria seca em milho. Rev. Bras. de Milho e and Computer Science Institute of Sao Carlos, Brazil.
Sorgo 8 (2), 157–173.