Computers and Electronics in Agriculture 116 (2015) 109–117

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

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

Hyperspectral vegetation indices for predicting onion (Allium cepa L.)yield spatial variability

http://dx.doi.org/10.1016/j.compag.2015.06.0140168-1699/� 2015 Elsevier B.V. All rights reserved.

⇑ Corresponding author. Tel.: +39 0874404709; fax: +39 0874404855.E-mail address: stefanomarino@unimol.it (S. Marino).

S. Marino ⇑, A. AlvinoDepartment of Agricultural, Environmental and Food Sciences (DAEFS), University of Molise, Via De Sanctis, I-86100 Campobasso, Italy

a r t i c l e i n f o

Article history:Received 5 December 2014Received in revised form 5 June 2015Accepted 17 June 2015

Keywords:Onion yieldVegetation indicesSpatial variabilityAgronomic traits

a b s t r a c t

An agronomic research was conducted to evaluate the spatial variability of an onion crop, with the aim totest Vegetation indices (VIs) as a tool to detect different yield areas.

Eleven VIs were derived from geo-referred hyperspectral readings taken at bulbification stage. Eight VIsshowed significant regressions with yield, and grouped in four clusters according to statistical analysis(H = high; Ms and Mi as medium superior and inferior; L = low). Maps were elaborated with ordinaryKriging. At a visual assessment, many VIs appeared similar to yield map. The surface analysis revealedthat all VIs accurately detected an L area (top of maps) characterized by heavy soil constrains, and theH area on the left side of the map (button and upper part). The best estimation of the total field yieldwas obtained by the so-called Soil-line vegetation indices and in particular by TSAVI. This study rein-forces the possibility of assessing onion yield by spectroradiometric measurements at field scale.

� 2015 Elsevier B.V. All rights reserved.

1. Introduction

Horticulture crops play a significant role in improving the pro-ductivity of land, generating employment, enhancing exports,improving economic conditions of the farmers and entrepreneursand providing food and nutritional security to the people (Ushaand Singh, 2013). Onions (Allium cepa L.) are one of the world’s old-est cultivated vegetable, widely cultivated all over the world, withparticular diffusion in the Asian continent and in Europe. Onion ismainly used as a flavouring to enhance the taste and savour ofnumerous dishes and in many countries it is also used as a fresh,cooked, and dehydrated vegetable (Kumar et al., 2007a).

Growth and development of onion relies on interactions amonggenotype, agronomic practices and pedo-climatic conditions.Agronomic practices can have a significant influence on onion bio-mass, yield, yield components (weight, number and bulb diameter)and quality (Martìn de Santa Olalla et al., 2004; Kumar et al.,2007b). Achieving maximum crop yield at minimum cost with alower consumption of resources is one of the goals of agriculturalproduction and environmental protection. Early detection of agro-nomic constrains may give a significant impulse to augment thequality and the quantity of crop yield (Marino et al., 2015). A fur-ther positive impulse to optimized crop management (Qarallahet al., 2008) comes from the Variable Rate Applications of inputs.

Site-specific crop management offers the potential to improvecrop efficiency by tailoring inputs to address relevantwithin-field variability. Successful precision crop managementstrategies can lead to improved crop yield, increased profitability,and decrease of associated adverse environmental and healthimpacts (Mulla, 2013).

Remote and proximal sensing techniques can provide an instan-taneous, non-destructive, and quantitative information about theagricultural crop conditions and crop spatial variability during cropcycle (Marino et al., 2014a). From hyperspectral data have beendeveloped numerous spectral vegetation indices (VIs) (Bassoet al., 2011), which are more sensitive than individual bands ofcrop spectra to monitor crop status (Qi et al., 1994).

Nowadays VIs are used to monitor plant conditions, to estimateplant nutrient status, to detect abiotic and biotic stresses, to assesplant growth rate; to predict biomass and yield (e.g. Li et al., 2014;Peñuelas et al., 1993; Marino et al., 2014c).

The best management of certain areas of the field can enhancethe average value of onion yield. The identification during the cul-tural cycle of the area with the lowest bulbs production couldallow proper field management resulting in increased productionor, if the area had some irresolvable problem, in reducing input(fertilizers, water, etc.).

Several studies hypothesized the use of VIs to predict yield onMaize, Corn, Wheat, Barley and tomato (e.g. Marino and Alvino,2014). There are no evidence in the literature on onion, exceptfor some papers that have studied the relationship between VIs

110 S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117

and Leaf area indices (Gupta et al., 2000; Bosch Serra and Casanova,2000). Furthermore, as reported by Gausman and Allen (1973),spectroradiometric characteristics of row crop and in particularof onion crop are very special, in comparison with other 30 plantspecies, onion showed the highest leaf thicknesses, the lowestmean values of infinite reflectance and the lowest scattering coef-ficient, which was not correlated anyway to leaf thickness.

The experiment was carried out by using hyperspectral radio-metric measurements, with the following objectives: (i) to assessthe yield spatial variability of an onion crop (ii) to test the abilityof Vegetation Indices, at bulbification stage, to predict the yieldlevel of different crop areas.

2. Materials and methods

2.1. Experimental set-up

The study was carried out in 2010 in Central Italy (latitude N41�52.855, longitude E 14�55.728, 30 m a.s.l.) on a plain dedicatedto horticultural crops. The onion seedling was the Red Mech (ISISementi, Italy), a long-day hybrid, suitable for spring sowing, usu-ally harvested in late July and August. The bulbs of good size, char-acterized by high homogeneity, have high-top form, full at thecollar; the outer skins are thick and show an intense and brilliantpurplish red color. Red Mech can be stored all winter long and isparticularly suitable for mechanical harvesting.

The hybrid was sown on January 25th, at a density of75 seeds m�2 on a 50 m � 400 m field (2 hectares); harvest startedfrom the first week of August and lasted one week.

Fertilization was performed using fertilizers and doses com-monly used for the cultivation of onion in central Italy and weedswere controlled with specific chemical herbicides.

Crop was irrigated by sprinkler with pipes 40 mm in diameterand the spray heads 50–22 PSI (NAAN, Israel). Daily maximumand minimum temperatures and rainfall were recorded througha standard agro-meteorological station (Skye instruments Ltd.,Llandrindod Wells, UK), placed beside the experimental field(Table 1). The daily average temperature during the study period(January–August) ranged from 6.5 �C to 25.1 �C, with the minimumtemperatures recorded in March (�4.7 �C) and the maximum tem-perature recorded in July (37.4 �C). The total rainfall was about370 mm, with a uniform distribution. The reference evapotranspi-ration (ET0) of the crop cycle was about 856 mm. As a rough esti-mation, rainfall was about the 40% of ET0; the total amount ofwater applied was 3500 m3 ha�1.

2.2. Measurements

The soil texture was recorded by means of 32 geo-refereed sam-ples. The soil was classified as clay and clay loam, with 37% of sand(S.D. ±6.69) 20% of silt (S.D. ±3.6) and 42% of clay (S.D. ±5.0), with

Table 1Ten-day averages rainfall, minimum (T min) maximum (T max) and mean temper-atures (T mean) and evapotranspiration (ETo) during crop seasons.

Rainfall T min T max T mean ET0

January 40.6 �2 17.5 6.5 27February 66.4 �2.8 22.3 8.2 41March 43.2 �4.7 26 9.7 68April 49.0 0.9 25.8 12.7 123May 45.4 6.8 29.9 17.6 153June 63.0 8.9 36.7 21.3 177July 37.6 13.7 37.4 25.1 212Augusta 1.80 16.8 30 23.6 54.6

a August: first decade.

medium organic matter (1.83%), medium content in total N(1.16 g kg�1), moderately alkaline (pH 7.61). The soil of the upperpart of the field (starting from 360 m to 400 m) was stony,although the texture analysis did not show significant differences.

At harvest time (50 ± 60% of the leaves were dead), the diame-ter, weight and the number of bulbs were determined at 64geo-referenced points. Fresh weight of below and above groundplants was obtained from sampling plots of one square meter.Dry mass accumulation was obtained leaving plants in aforced-draft oven at 75 �C to constant weight (Marino et al., 2013).

2.3. Crop spectroradiometric measurements

The radiance over canopy was measured with an ASD FieldSpecHand-Held Pro spectroradiometer (Analytical Spectral Device,Boudler, CO, USA), and converted to spectral reflectance by divid-ing the radiance reflected by the target by that reflected by a stan-dard white reference plate (spectralon). This instrument covers theportion of the spectrum between 350 and 1100 nm, with a spectralsampling distance of <1.5 nm. Spectral reflectance was measuredat nadir in cloud-free days. The height of the optics was, on average80 cm above the canopy. In order to minimize the effects of thesun’s position, reflectance measurements were taken in about2 h, near solar noon (between 11:00 and 13:00 solar time).

A total of 64 georeferred spectral measurements were taken at140 days after transplanting. Hyperspectral data were smoothedwith the Savitzky–Golay filter (Savitzky and Golay, 1964), toimprove the quality of canopy spectra. The Savitzky–Golay filteris based on least squares polynomial fitting across a moving win-dow within the data. After filtering, the following VIs were calcu-lated: NDVI, GNDVI, SAVI, TSAVI, MSAVI, OSAVI, WDVI, PVI, WI,TCARI, SAVI2 from canopy spectra according to the formulasshown in Table 2. For a detailed discussion of these indices, thereader is referred to the references cited in this table.

2.4. Statistical analysis

The 64 yield data were analyzed by a clustering method, usingHierarchical clustering Ward’s minimum variance approach.Cluster analysis is a collection of statistical methods that identifiesgroups of samples that behave similarly or show similar character-istics. Cluster analysis classifies a set of observations into two ormore mutually exclusive unknown groups based on combinationsof interval variables. The purpose of the cluster analysis is to dis-cover a system that can classify observations into groups, in whichthe group members have properties in common. Agglomerativehierarchical cluster methods produce a hierarchy of clusters fromsmall clusters of very similar items to large clusters that includeitems that are more dissimilar. Hierarchical methods usually pro-duce a graphical output known as a dendrogram, or tree thatshows this hierarchical clustering structure (Ward, 1963).Agglomerative clustering begins by finding the most similar twogroups, based on the distance matrix, and subsequently mergingthem into a single group. This procedure is repeated, step by step,until all the samples have been added to a single large cluster. Thefinal partition is identified by a distance criterion (Fernández andGómez, 2008). Starting from the bottom part of the dendrogram,the researcher decides to stop the agglomeration process whensuccessive clusters are too far apart to be merged.

2.4.1. Assessing normalityShapiro–Wilk method, Lilliefors method, Anderson–Darling

method and D’Agostino et al. method were used to check thenormality assumption as described by Marino et al. (2014a). Therejection of one or more of this test is a symptom ofnon-normality distribution.

Tabl

e2

Refle

ctan

cein

dex,

form

ulat

ion

and

refe

renc

esof

vege

tati

onin

dice

sob

tain

edfr

omsp

ectr

alre

flect

ance

mea

sure

men

ts.

Veg

etat

ion

indi

ces

Form

ula

tion

Ref

eren

ces

Tran

sfor

med

soil

-adj

ust

edve

geta

tion

inde

x(T

SAV

I)(1

.234

4⁄(

R80

0�

1.23

44⁄R

670�

0.01

83))

/(1.

2344⁄R

800

+R

670�

1.23

44⁄0

.018

3)B

aret

etal

.(19

89)

Soil

-adj

ust

edve

geta

tion

inde

x(S

AV

I)(R

800�

R67

0)/(

R80

0+

R67

0+

0.5)

)⁄(

1+

0.5)

Hu

ete

(198

8)N

orm

aliz

eddi

ffer

ence

vege

tati

onin

dex

(ND

VI)

(R80

0�

R67

0)/(

R80

0+

R67

0)R

ouse

etal

.(19

74)

Perp

endi

cula

rve

geta

tion

inde

x(P

VI)

1/R

AD

Q(1

.234

4^2

+1⁄(

R80

0�

1.23

44⁄R

670�

0.01

83))

Ric

hard

son

and

Wie

gan

(197

7)O

ptim

ized

soil

-adj

ust

edve

geta

tion

inde

x(O

SAV

I)(1

+0.

16)⁄(

R80

0�

R67

0)/(

R80

0+

R67

0+

0.16

)R

onde

aux

etal

.(19

96)

Wei

ghte

ddi

ffer

ence

vege

tati

onin

dex

(WD

VI)

R80

0�

(1.2

344⁄R

670)

Cle

vers

,198

9M

odifi

edso

il-a

dju

sted

vege

tati

onin

dex

(MSA

VI)

(R80

0�

R67

0)/(

R80

0+

R67

0+

(1+

(1–1

.234

^2⁄R

800�

(1.2

344⁄R

670)⁄(

R80

0�

R67

0)/(

R80

0+

R67

0)))

)Q

iet

al.(

1994

)G

reen

nor

mal

ized

diff

eren

ceve

geta

tion

inde

x(G

ND

VI)

(R80

0�

R55

0)/(

R80

0+

R67

0)G

itel

son

etal

.(19

96)

Seco

nd

soil

-adj

ust

edve

geta

tion

inde

x(S

AV

I2)

R80

0/(R

670

+0.

0183

/1.2

344)

Maj

oret

al.(

1990

)Tr

ansf

orm

edch

loro

phyl

lab

sorp

tion

inre

flec

tan

cein

dex

(TC

AR

I)3⁄(

(R70

0�

R67

0)�

0.2⁄(

R70

0�

R55

0)⁄(

R70

0/R

670)

)K

imet

al.(

1994

)W

ater

inde

x(W

I)R

900/

R97

0Pe

ñu

elas

etal

.(19

93,1

997)

S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117 111

2.4.2. Non-parametric ANOVA by Kruskal–WallisThe Kruskal–Wallis test (Kruskal and Wallis, 1952) is most com-

monly used when there is one nominal variable and one measure-ment variable, and the measurement variable does not meet thenormality assumption. It can be interpreted as the non-parametricalternative to classical one-way Analysis of Variance test. Aone-way ANOVA may yield inaccurate estimates of the P-valuewhen the data are very far from normal distribution. The Kruskal–Wallis test does not make assumptions about normality. Like mostnon-parametric tests, it is performed on ranked data, so the mea-surement observations are converted to their ranks in the overalldata set. The KW test verifies whether three or more independentgroups have same distribution. All statistical procedures were com-puted using the statistical packages IBM SPSS 21 (IBM Inc. USA) andOriginPRO 8 (Origin Lab Corporation, Northampton, MA 01060,USA).

2.4.3. Regression analysisRegression analysis, coefficients of determination, significance

levels and RMSE were computed on the 64 geo-referred yield dataand VIs ones, using the statistical package Origin PRO 8 (Origin LabCorporation, Northampton MA 01060 USA).

2.4.4. Geostatistical analysisOrdinary kriging (OK) is a commonly used linear method of

spatial prediction to provide estimates of variables at unvisitedsites. The procedure uses information from neighbouring pointsto predict at a target point [1]; weights are assigned to these pointsbased on their distance from the target. The equation for ordinarypunctual kriging is:

Z�OKðx0Þ ¼Xn

i¼1

WiZðxiÞ ð1Þ

where Z*OK (x0) is the OK estimate at an unsampled location (x0), n is

the number of samples in the search neighbourhood, wi are theweights assigned to the ith observation z(xi).

Weights are assigned to each sample such that the estimationor kriging variance, E = [{Z⁄(x0) � Z(x0)}2], is minimized and theestimates are unbiased (Webster and Oliver 2007). The weightsdepend on the relative positions of the samples in the neighbour-ing both to one another and to the target point, and on the vari-ogram. The latter describes the spatial correlation and covariancestructure between data points for each variable [2]. The variogramcan be computed by Matheron (1965) usual method of moments asfollows:

y ¼XmðhÞ

i¼1

zðxi þ hÞ � zðxiÞ½ �2 ð2Þ

where y is the semivariance between two observation points, z(xi)and z(xi + h), separated by a distance h, and m(h) is the number ofpairs at h.

The best variogram model for each parameter was selectedbased on cross validation. Cross validation was performed usingmean error (ME), root mean square error (RMSE) and standardizedmean square error (SMSE) (Delhomme, 1978; Merino et al., 2001)(Table 3). The semivariogram was computed using GS+ version 8,while kriging was done using surfer version 9 (Golden software,Golden, Colorado, USA) according to Selvaraja et al. (2012), withx; y representing the UTM coordinates (expressed in meters), andz the parameter values.

Table 3Semivariogram parameters (model, nugget effect, sill, range) and cross validation (mean square ME, root mean square error RMSE and standardized mean squared error) used toonion yield, and vegetation indexes kriging map.

Model Nugget Sill Range ME Residual

RMSE SMSE

Yield Spherical 0.343 0.782 61.1 0.640872 0.800545 1.001889TSAVI Spherical 0.0032 0.0164 59.0 0.0146 0.12067 1.000008SAVI Spherical 0.0042 0.0073 61.2 0.0073 0.08544 1.000082NDVI Spherical 0.0028 0.0111 54.0 0.0111 0.10539 1.000062PVI Spherical 0.0001 0.0002 52.6 0.0002 0.01413 1.000110OSAVI Spherical 0.0040 0.0084 51.1 0.0094 0.09709 1.000002WDVI Spherical 0.0027 0.0036 59.9 0.0052 0.07232 1.195542MSAVI Spherical 0.0009 0.0016 62.6 0.0011 0.03260 1.000102GDVI Spherical 0.0016 0.0045 59.0 0.0052 0.07232 1.000002

Table 4Cluster centroids: mean (StdDev) of 64 georeferenced yield data, split into fourdifferent clusters: high (H), medium superior (Ms), medium inferior (Mi) and low (L)according to hierarchical clustering analysis with Euclidean distance and Ward’s linkaggregation.

H Ms Mi L

Yield (kg h�1) 8.29 (0.53) 7.20 (0.21) 6.57 (0.21) 5.36 (0.74)

Fig. 1. Spatial distribution of onion yield (kg m�2), divided into high (H), mediumsuperior (Ms), medium inferior (Mi) and low (L) zones, as modeled by ordinarykriging.

112 S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117

3. Results and discussion

The field average value of onion yield crop was 6.57 kg m�2, withthe lowest data of 3.00 kg m�2and the highest yield 9.40 kg m�2,these values indicate the presence of a significant variability inthe field. Yield data were split in cluster groupings to better under-stand the crop field variability. The cluster selection procedureidentified four clusters, labelled as (H) High-value cluster; (Ms)Medium-superior-value cluster; (Mi) Medium-inferior-value clus-ter and (L) Low-value cluster. The yield onion mean related to Hcluster was 8.29 kg m�2, 13% higher than Ms cluster (7.20 kg m�2),20% higher than Mi cluster (6.57 kg m�2) and 35% higher than Lcluster (5.36 kg m�2) (Table 4).

According to cluster selection analysis, georeferenced yield datawere spatialized with geostatistical analysis (Ordinary Kriging) tobest investigate the yield spatial variability of the onion field crop.Fig. 1 shows yield onion map, three H zones were detected on theleft side of the field, the first one starting from 10 to 40 m (on thelong side), the second one from 110 to 120 m and the third areastarting from 270 to 340 m, spreading up to 20 m into the field.The total surface of H area was calculated to be 1300 m2 (6% oftotal surfaces). Two Ms area surrounding the H area, with asurfaces of about 4200 m2 (21% of total surfaces) were detected.The widest surface has been defined as the area Mi, with about11,700 m2, which included the right part of the field from 0 to150 m, and most of the field starting from 150 m to 360 m. Theincidence of the Mi zone was the highest, and equal to the 59%of the total area of the crop field. The less productive areas (14%of total surfaces) have been identified in the lower right part of

Fig. 2. Spectral reflectance at vis–NIR wavelength of onion crop at high (H),medium superior (Ms), medium inferior (Mi) and low (L), onion yield level.

Fig. 3. The onion bulbs yield plotted versus TSAVI, SAVI, NDVI, PVI, OSAVI, WDVI, MSAVI and GNDVI.

S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117 113

114 S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117

the field (two areas) and in the upper part starting from 360 m,with a total area of about 2700 m2.

Spectroradiometric measurements were taken at bulbificationstage to identify ex-post the best Vegetation Indices significantlyrelated to yield for evaluating the potential of VIs to detect fieldvariability in onion yield before harvest. Readings were taken atbulbification stage according to the findings of Martìn de SantaOlalla et al. (2004), who noticed very little variations of someindices (e.g. NDVI) when measured at the bulbification stages.Martìn de Santa Olalla et al. (2004) has found that, the little varia-tion could be due to the reason that leaves are nearly vertical andthus offer less interaction cross-section and the plant-soil exposureproportion in the top-down radiometric measurement wouldchange very little. In fact, Gupta et al. (2000) states that the verticalarchitecture of canopy is responsible for the slower rate ofdecrease for Vegetation indices (NDVI in particular). After the bulb-ification stage, the moisture content of the leaves decreases andconsequently its stem begins to shrink. Gupta et al. (2000) foundthat the chlorophyll absorption in the red band comes down(increasing the reflected signal in the red band), and reflected sig-nal in the near-infrared band decreases, as well. After bulbificationstage, the leaves begin to fall, resulting in fall of indices value; withthe result that the significance of correlation among LAI, yield andVIs decreases (Gupta et al., 2000; Marino et al., 2013). Furthermore,the use of remote sensing on onion crop growth at the early stageshas not yet been successful because, as reported by Bosch Serraand Casanova (2000), onion crop grows in rows and its biomassat the early stages is very small.

Starting from the spectroradiometric measurements, the firststep was to detect differences in the spectra vegetation curves,the mean of all the georeferenced spectra of four clusters (H, Ms,Mi and L) are presented in Fig. 2. The curves showed differencesamong clusters: up to 700 lm a lower reflectance for the H sam-ples was recorded compared to Ms, Mi and L respectively. From700 lm and up to 1075 lm the situation was reversed, with thehigher reflectance samples recorded by H followed by Ms, Miand L samples. The difference among reflectance curves is gov-erned by internal leaf structure, leaf surface properties and bythe concentration and distribution of bio-chemical component(Peñuelas et al., 1997). Starting from geo-referred reflectance mea-surements, eleven VIs were calculated and correlated to the equiv-alent geo-refereed yield samples. Out of eleven VIs, eight indices(Fig. 3) showed significant regression curves (R2 0.61–0.67). Thebest relation was found by TSAVI (R2 0.67), followed by SAVI, PVIand NDVI indices (R2 = 0.66), followed by WDVI (R2 = 0.64) and atlast OSAVI, GNDVI (R2 = 0.62) and MSAVI (R2 0.61). In our case,the linear model was the best at describing the relationship

Table 5Mean of 64 georeferenced data, yield and VIs data, split into four different zones: high(H), medium superior (Ms), medium inferior (Mi) and low (L) processed by Kruskal–Wallis one-way analysis of variance (* Significant at the 0.05 probability level;** Significant at the 0.01 probability level; n.s. – not significant).

H Ms Mi L Kruskal–Wallis KW (corr.ties)

Yield 8.29 7.20 6.57 5.36 –** –**

TSAVI 0.75 0.71 0.66 0.45 –** –**

SAVI 0.54 0.51 0.44 0.30 –** –**

NDVI 0.77 0.74 0.69 0.52 –** –**

PVI 0.74 0.74 0.76 0.78 –** –**

OSAVI 0.65 0.62 0.56 0.40 –** –**

WDVI 0.32 0.30 0.24 0.15 –** –**

MSAVI 0.16 0.15 0.12 0.08 –** –**

GNDVI 0.67 0.64 0.62 0.51 –** –**

SAVI 2 5.80 5.70 5.10 4.80 n.s. n.s.TCARI 0.27 0.27 0.26 0.23 n.s. n.s.WI 1.15 1.13 1.11 1.10 n.s. n.s.

between indices and onion yield, according to several studies thathave established that the linear relationship provides better corre-lation among VIs and crops growth and yield (Còrcoles et al., 2013;Gupta et al., 2000).

The different clusters selection zones computed by Kruskal–Wallis one-way analysis-of-variance were also reported in Fig. 3to analyze and evaluate the distribution of clusters in the regres-sions. It is important to emphasize and reiterate that the clusterswere created on the yield basis and statistical tests were used toverify (a posteriori) the dependence conditioning of the differentvariables, compared to the four clusters. The clustering did not pro-vide ready explanations for the yield variability but indicated thepresence of yield variability, and can help to determine theyield-limiting factors on the field and appropriate managementzones (Vrindts et al., 2005). It is also important to remark thatthe average cluster of eight VIs were significantly different fromeach other (Table 5), with besides TSAVI, OSAVI, NDVI and SAVIindices with a greater range of the average minimum and maxi-mum reflectance values.

The Water Index, SAVI2 and TCARI do not have a significant cor-relation with plant yield, probably because the reflectance differ-ences among minimum and maximum value in all samplingdates are too small. In literature, different studies confirmed thepossibility to have no correlation due to very close values to eachother (Römer et al., 2012).

The significant regression between the vegetation indices andyield gave no information about the spatial yield variability andon the ability of the VIs to identify areas with onion yield loss,for this reason, a VIs maps to analyze the spatial distribution ofthe best indices were elaborated. Fig. 4 shows the eight mapsderived from the indices significantly correlated with yield. By avisual assessment all VIs showed only two H zones on the left sideof the field, characterized by different surfaces and different accu-racy in the identification of the yield H area.

All maps also identified with different accuracy, the areasurrounding the H area (Ms); instead, all indices overestimatethe production of an area located on the left side of the field andat the central part of the field. This was probability due to amomentary state of best plant health (e.g. effect of fertilizations)compared to the surrounding area, that has not been reflected inthe Onion yield at harvest. Except for the GNDVI, the other indiceshave clearly identified the Mi area, and as we expected, in somecases VIs have overestimated onion yield and in other cases haveunderestimated it. The L area placed in the bottom right part ofthe field has not been identified by VIs, except for GNDVI, whilethe L area placed in the upper part of the field has been well iden-tified by all the indices. Soil analysis on the L area at the end of thefield showed soil constrains due to the high presence of skeleton,the soil was alkaline (pH = 8 , 15), with high values of CaCO3

(21%), that explains the differences in yield respect to other fieldareas and emphasizes the need to differently manage this area.Calcareous soils, as reported by Alam et al. (2010) negatively affect,in onion, the macronutrients uptake (Zn, B, Mo, Mn, Cu, Cl) and, asa consequence onion growth and yield.

As found by Maguire (2000), precise information on the chargesin soil type could be used for example to modify the seed rate ofonion seeds to improve the size uniformity of the onion crop,and increasing marketable yield.

The other two L areas on the bottom part to the right of the fieldhave not been identified by the indices, except for the NDVI whichunderestimates the areas, and that GNDVI with an overestimationof the areas. TSAVI, SAVI, OSAVI, MSAVI and NDVI seemed able toreturn the best response, partly confirmed by surface computing.TSAVI is the most closely index to the yield map, with an underes-timation of H, Ms and L zones (�29%, �7% and �23% respectively)and an overestimation of Mi zones (+11%) (Fig. 5). The Transformed

Fig. 4. Spatial distribution of TSAVI, SAVI, NDVI, PVI, OSAVI, WDVI, MSAVI and GNDVI indices divided into high (H), medium superior (Ms), medium inferior (Mi) and low (L)zones, as modeled by ordinary kriging.

S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117 115

Fig. 5. Surface incidence (%) of onion yield and vegetation indices (TSAVI, SAVI,NDVI, PVI, OSAVI, WDVI, MSAVI and GNDVI) maps divided into high (H), mediumsuperior (Ms), medium inferior (Mi) and low (L) zones.

116 S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117

Soil Vegetation Index is widely used for the ability to minimize thecontribution of soil reflectance to the canopy reflectance and istherefore used in different crops to predict crop growth and yieldvariables. In agreement with what found by Qi et al. (1994),Rondeaux et al. (1996), Huete (1988) on other crops, TSAVI per-formed better than other indices to detect crop variability. SAVI,OSAVI and NDVI were the indices that best identified, behind theTSAVI, the spatial variability of the field. In detail, SAVI have under-estimated the H, Ms and L area (incidence of surfaces about �25%of difference from yield surface) and have overestimated the Miarea (about 25% of difference from yield surface). OSAVI andNDVI have underestimate H, Mi and L areas (difference from yieldsurface: about �75% for H area, �2% for Mi area and �30% for Larea) and have overestimate the Ms area (difference from yieldsurface: about +50%).

All other indices have less identified the surfaces of differentareas. However important differences were found in the identifica-tion of several areas, despite this, it would seem that the use ofsome VIs could be useful to identify and to mark zones with similarcharacteristics, and a valid support to improve field monitoringand field management.

Among VIs, the best performance was obtained by the so-calledSoil-line vegetation indices, which use the information of soil linein NIR-Red reflectance to reduce the effect of the soil on canopyreflectance [Soil Adjusted Vegetative Index (SAVI); Optimized SoilAdjusted Vegetative Index (OSAVI); Transformed Soil AdjustedVegetative Index (TSAVI)]. This appears to be a normal result, giventhat as reported by Tei et al. (1996), ground cover is a useful indi-cation of light interception for prostrate leaves, but is less usefulfor crops with erect leaves like onion because only a small fractionof incident radiation is vertical. Onion showed a lower early rela-tive growth rate than other crops; this was due partly to the lowlight interception ability of the crop canopy and partly to the lowinitial radiation use efficiency compared with that of the othertwo crops. On the other hand, thanks to a uniform distribution ofthe radiation inside the canopy, to the cessation of leaf develop-ment after the start of bulbing, and to the lower costs of storageand maintenance in the later phase of growth, onion showed highradiation use efficiency and was able to produce a large amount of

dry matter. Its growth limits seem to be the low light interceptiondue to the leaf posture and to the relatively short duration of a highground cover compared with the length of bulbing process(Brewster, 1990).

4. Conclusion

Onion hybrid Red Mech showed good average bulb yield(6.57 kg m�2), characterized by a high yield variability, rangingfrom 3 kg m�2 to 9.40 kg m�2. Cluster analysis on georeferencedyield data identified four clusters with increasing productivity(H = high; Ms and Mi as medium superior and inferior; L = low).The yield map was elaborated according to Ordinary Kriging andit has clearly identified four productive zones characterized fromthe following mean values, 8.29 kg m�2 for H yield area, 7.2 kg m�2

for Ms area, 6.57 kg m�2 for Mi and 5.36 kg m�2 for L area.Spectroradiometric georeferenced measurements were

recorded at onion bulbification with the aim to identifyVegetation Indices correlated to onion yield. Eight of eleven VIsshowed a significant linear regression with yield (R2 from 0.61 to0.67).

However, only four indices (TSAVI, SAVI, OSAVI and e NDVI)were able to detect yield spatial variability according to four clus-ters. The above-mentioned indices well detected the surfaces withlowest yield (L area) in the upper part of the field. The H yield sur-faces area (6% of total surface) were very similar to those of TSAVIand SAVI (5% of total surface). Ms and Mi yield area were well iden-tified only by TSAVI index that showed the best performanceamong VIs. Given the onion plant characteristics, also linked toan incomplete soil cover, TSAVI seemed to be more effective thanothers indices to identify onion yield spatial variability; it wasunfortunately unable to identify several yield spot areas (e.g. lowerthan 150 m2).

References

Alam, M.N., Abedin, M.J., Azad, M.A.K., 2010. Effect of micronutrients on growth andyield of onion under calcareous soil environment. Int. Res. J. Plant Sci. 1 (3), 56–61.

Baret, F., Guyot, G., Major, D., 1989. TSAVI: a vegetation index which minimizes soilbrightness effects on LAI and APAR estimation. In: Proceedings of 12th CanadianSymposium on Remote Sensing and IGARSS’89, Canada, 10–14 July 1989, pp.1355–1358.

Basso, B., Cammarano, D., Cafiero, G., Marino, S., Alvino, A., 2011. Cultivardiscrimination at different site elevations with remotely sensed vegetationindices. Ital. J. Agron. 2011 (6), e1.

Bosch Serra, A.D., Casanova, D., 2000. Estimation of onion (Allium Cepa, L.) biomassand light interception from reflectance measurements at field level. In:Proceedings of XXV IHC – Part 9, pp. 53–69 (Acta Horti. 519).

Brewster, J.L., 1990. Physiology of crop growth and bulbing. In: Rabinovitch, H.D.,Brewster, J.L. (Eds.), Onions and Allied Crops. Vol. I. Botany, Physiology andGenetics. CRC Press, Boca Raton, Florida, USA, pp. 53–88.

Clevers, J.G.P.W., 1989. The application of a weighted infrared-red vegetation indexfor estimating leaf area index by correcting soil moisture. Remote Sens. Environ.29, 25–37.

Còrcoles, J.I., Ortega, J.F., Hernàndez, D., Moreno, M.A., 2013. Estimation of leaf areaindex in onion (Allium cepa L.) using an unmanned aerial vehicle. Biosyst. Eng.115, 31–42.

Delhomme, J.P., 1978. Kriging in the hydrosciences. Adv. Water Resour. 1, 251–266.Fernández, A., Gómez, S., 2008. Solving non-uniqueness in agglomerative

hierarchical clustering using multidendrograms. J. Classif. 25 (1), 43–65.Gausman, H.W., Allen, W.A., 1973. Optical parameters of leaves of 30 plant species.

Plant Physiol. 52 (1), 57–62.Gitelson, A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in the

remote sensing of global vegetation from EOS–MODIS. Remote Sens. Environ. 58(1996), 289–298.

Gupta, R.K., Prasad, T.S., Vliayan, D., 2000. Relationship between LAI and NDVI forIRS LISS and LANDSAT TM bands, 2000. Adv. Space Res. 26 (7), 1047–1050.

Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25,295–309.

Kim, M.S., Daughtry, C.S.T., Chappelle, E.W., McMurtrey III, J.E., Walthall, C.L., 1994.The use of high spectral resolution bands for estimating absorbedphotosynthetically active radiation (Apar). In: Proceedings of the 6thSymposium on Physical Measurements and Signatures in Remote Sensing,January 17–21, 1994, France, pp. 299–306.

S. Marino, A. Alvino / Computers and Electronics in Agriculture 116 (2015) 109–117 117

Kruskal, W.H., Wallis, W.A., 1952. Use of ranks in one-criterion variance analysis. J.Am. Stat. Assoc. 47 (260), 583–621.

Kumar, S., Imtiyaz, M., Kumar, A., 2007a. Effect of differential soil moisture andnutrient regimes on postharvest attributes of onion (Allium cepa L.). Sci. Hortic.112, 121–129.

Kumar, S., Imtiyaz, M., Kumar, A., Singh, R., 2007b. Response of onion (Allium cepa L.)to different levels of irrigation water. Agric. Water Manag. 89, 161–166.

Li, F., Miao, Y., Feng, G., Yuan, F., Yue, S., Gao, X., Liu, L., Liu, B., Ustin, S.L., Chen, X.,2014. Improving estimation of summer maize nitrogen status with red edge-based spectral vegetation indices. Field Crop. Res. 157, 11–123.

Maguire, S., 2000. Variable rate precision drilling. MSc by Research Thesis, CranfieldUniversity at Silsoe, Silsoe, Bedford, UK.

Major, D.J., Baret, F., Guyot, G., 1990. A ratio vegetation index adjusted for soilbrightness. Int. J. Remote Sens. 11, 727–740.

Marino, S., Alvino, A., Cocozza, C., Tognetti, R., 2015. Use of proximal sensing andvegetation indexes to detect the inefficient spatial allocation of drip irrigation ina spot area of tomato field crop. Prec. Agric. http://dx.doi.org/10.1007/s11119-015-9396-7.

Marino, S., Alvino, A., 2014. Proximal sensing and vegetation indices for site-specificevaluation on an irrigated crop tomato. Eur. J. Remote Sens. 47, 271–283.

Marino, S., Alvino, A., Cocozza, C., Tognetti, R., 2014c. Effects of inefficient spatialallocation of irrigation water on fruit yield, leaf physiology and spectralreflectance in a tomato crop. Acta Hortic. 1038, 185–192.

Marino, S., Aria, M., Basso, B., Leone, A.P., Alvino, A., 2014a. Use of soil andvegetation spectroradiometry to investigate crop water use efficiency of a dripirrigated tomato. Eur. J. Agron. 59, 67–77.

Marino, S., Basso, B., Leone, A.P., Alvino, A., 2013. Agronomic traits and vegetationindices of two onion hybrids. Sci. Hortic. 155, 56–64.

Martìn de Santa Olalla, F., Domìnguez-Padilla, A., López, R., 2004. Production andquality of the onion crop (Allium cepa L.) cultivated under controlled deficitirrigation conditions in a semi-arid climate. Agric. Water Manag. 68,77–89.

Matheron, G., 1965. Les variables regionalisees et leur estimation: une applicationde la theorie de fonctions aleatoires aux sciences de la nature. Masson et Cie,Paris.

Merino, G.G., Jones, D., Stooksbury, D.E., Hubbard, K.G., 2001. Determination ofsemivariogram models to krige hourly and daily solar irradiance in WesternNebraska. J. Appl. Meteorol. 40, 1085–1094.

Mulla, D.J., 2013. Twenty five years of remote sensing in precision agriculture: keyadvances and remaining knowledge gaps. Rev. Biosyst. Eng. 114.

Peñuelas, J., Filella, I., Briel, C., Serrano, L., Savé, R., 1993. The reflectance at the 950–970 nm region as an indicator of plant water status. Int. J. Remote Sens. 14,1887–1905.

Peñuelas, J., Piñol, J., Ogaya, R., Filella, I., 1997. Estimation of plant waterconcentration by the reflectance water index WI (R900/R970). Int. J. RemoteSens. 18, 2869–2875.

Qarallah, B., Shoji, K., Kawamura, T., 2008. Development of a yield sensor formeasuring individual weights of onion bulbs. Biosyst. Eng. 100, 511–515.

Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A modified soiladjusted vegetation index. Remote Sens. Environ. 48, 119–126.

Richardson, A.J., Wiegand, C.L., 1977. Distinguishing vegetation from soilbackground information. Photogramm. Eng. Remote Sens. 43, 1541–1552.

Römer, C., Wahabzada, M., Ballvora, A., Pinto, F., Rossini, M., Panigada, C., Behmann,J., Léon, J., Thurau, C., Bauckhage, C., Kersting, K., Rascher, U, Plümer, L., 2012.Early drought stress detection in cereals: simplex volume maximisation forhyperspectral image analysis. Funct. Plant Biol. 39, 878–890.

Rondeaux, G., Steven, M., Baret, F., 1996. Optimization of soil-adjusted vegetationindices. Remote Sens. Environ. 55, 95–107.

Rouse, J.W., Haas Jr., R.H., Schell, J.A., Deering, D.W., 1974. Monitoring vegetationsystems in the great plains with ERTS. In: Proceedings of 3rd ERTS-1Symposium, Greenbelt, MD, vol. 1, pp. 9�317

Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplifiedleast squares procedures. Anal. Chem. 36 (8), 1627–1639.

Selvaraja, S., Balasundram, S.K., Vadamalai, G., Husni, M.H.A., 2012. Spatialvariability of orange spotting disease in oil palm. J. Biol. Sci. 12, 232–238.

Tei, F., Aikman, D.P., Scaife, A., 1996. Growth of lettuce, onion and red beet. 2.Growth Model. Ann. Bot. 78, 645–652.

Usha, K., Singh, B., 2013. Potential applications of remote sensing in horticulture – Areview. Sci. Hortic. 153, 71–83.

Vrindts, E., Mouazen, A.M., Reyniers, M., Maertens, K., Maleki, M.R., Ramon, H., DeBaerdemaeker, J., 2005. Management zones based on correlation between soilcompaction, yield and crop data. Biosyst. Eng. 92 (4), 419–428.

Ward Jr., J.H., 1963. Hierarchical grouping to optimize an objective function. J. Am.Stat. Assoc. 58, 236–244.

Webster, R., Oliver, M.A., 2007. Geostatistics for Environmental Scientists, seconded. John Wiley Sons Ltd., England, 315p.