Sei sulla pagina 1di 10

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/259119272

Using the mobile phone as Munsell soil-colour sensor: An experiment under


controlled illumination conditions

Article  in  Computers and Electronics in Agriculture · November 2013


DOI: 10.1016/j.compag.2013.10.002

CITATIONS READS

68 3,044

6 authors, including:

Luis Gómez-Robledo Nuria López-Ruiz


University of Granada University Carlos III de Madrid
37 PUBLICATIONS   282 CITATIONS    25 PUBLICATIONS   478 CITATIONS   

SEE PROFILE SEE PROFILE

Alberto J Palma Luis Fermin Capitan-Vallvey


University of Granada University of Granada
142 PUBLICATIONS   2,129 CITATIONS    343 PUBLICATIONS   5,108 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Smart instruments for biosignal processing based on reconfigurable technologies View project

ECT systems based on reconfigurable technologies View project

All content following this page was uploaded by Manuel Sánchez-Marañón on 09 September 2016.

The user has requested enhancement of the downloaded file.


Computers and Electronics in Agriculture 99 (2013) 200–208

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture


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

Using the mobile phone as Munsell soil-colour sensor: An experiment


under controlled illumination conditions
Luis Gómez-Robledo a,⇑,1, Nuria López-Ruiz b,1, Manuel Melgosa a, Alberto J. Palma b,
Luis Fermín Capitán-Vallvey c, Manuel Sánchez-Marañón d
a
Departamento de Óptica, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain
b
ECsens, Departamento de Electrónica y Tecnología de Computadores, ETSIIT, Universidad de Granada, 18071 Granada, Spain
c
Grupo de Investigación Espectrometría en Fase Sólida, Departamento de Química Analítica, Facultad de Ciencia, Universidad de Granada, 18071 Granada, Spain
d
Departamento de Edafología y Química Agrícola, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain

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

Article history: Soil colour has been determined in most cases by using Munsell soil-colour charts, sometimes with spec-
Received 30 January 2013 trometers, and occasionally with digital cameras. The objective here is to assess whether a mobile phone,
Received in revised form 24 July 2013 which has all the requirements to capture and process digital images, might also be able to provide an
Accepted 2 October 2013
objective evaluation of soil colour under controlled illumination. For this, we have developed an Android
application that takes a picture of a soil sample, allowing the user to select the region of interest and then,
after a RGB image-processing and a polynomial process transform between colour spaces, the Munsell
Keywords:
(HVC) and CIE (XYZ) coordinates appear on the screen of mobile phone. In this way, a commercial HTC
Android
Colorimetry
smartphone estimated the colour of 60 crumbled soil samples between 2.9YR and 2.3Y with a mean error
Digital camera of 3.75 ± 1.81 CIELAB units, taking as a reference the colour measurements performed with a spectrora-
Munsell soil-colour charts diometer. The Munsell hue had the worst estimates (mean error of 2.72 ± 1.61 Munsell units) because of
its geometric mismatch with the RGB colour space and for being defined to illuminant C, different of the
D65 source under which the phone camera took the pictures. Because the measuring errors were lower
than those described in the literature for the visual determination of soil colour, and the application also
worked successfully in a different smartphone than the one used in its development, we think that cur-
rent experimental results encourage the expectations of using smartphones in the field as soil-colour
sensors.
Ó 2013 Elsevier B.V. All rights reserved.

1. Introduction the closest match between the soil sample and one of the standard
chips contained in the Munsell soil-colour charts, i.e. artificially
Specifying colour by the Munsell system is usual for artists, coloured papers mounted on constant hue charts, showing value
designers, scientists, engineers, and government regulators (ASTM, (lightness) and chroma (colour intensity) variations in the vertical
2008). For example, in the natural sciences, it has been used to and horizontal directions, respectively. Thus, the Munsell designa-
identify and record the colours of specimens such as human skin, tion of that chip (hue H, value V, and chroma C) is assigned to the
flowers, foliage, minerals, and soils. Specifically in soil science, soil sample under study. Several accuracy problems, however, have
Munsell colour has far-reaching implications for the examination, been reported previously in relation to identifying the colour of soil
description, and classification of soils (Soil Survey Staff, 1999; IUSS specimens using Munsell charts (Sánchez-Marañón et al., 1995,
Working Group WRB, 2006), as well as in studying soil genesis and 2005, 2011), all of them related to the three main factors affecting
evaluation, being a valuable indicator of soil structure and compo- the psychophysical character of colour: illumination conditions,
nents (Sánchez-Marañón et al., 2004). Although the specialized sample characteristics, and the observer’s sensitivities (Berns,
work in soil colour has been based on instrumental measurements 2000).
(Torrent and Barrón, 1993; Viscarra Rossel and Webster, 2011), the Currently, the increasing demand of soil data for applications
prevalent practice in soil science is visual determination. This seeks such as precision agriculture and dynamic models for monitoring
environmental change has spurred the development of proximal
soil sensors (Viscarra Rossel et al., 2011). The rationale is to collect
⇑ Corresponding author. Address: Departamento de Óptica, Universidad de
larger amounts of data using simpler, cheaper, faster, and less labo-
Granada, Cuesta del Hospicio s/n, 18071 Granada, Spain. Tel.: +34 958241903.
rious techniques than conventional soil analyses, even at the cost
E-mail address: luisgrobledo@ugr.es (L. Gómez-Robledo).
1
Both authors contributed equally to this work.
of less accuracy. The point is that many more measurements will

0168-1699/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.compag.2013.10.002
L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208 201

counteract their lower accuracy. Since soil colour is related to soil and (iv) to assess the accuracy achieved by the mobile phone with
components, and, by extension, properties or conditions depending respect to that of commercial spectrometers.
on them (Soil Survey Staff, 1993), proximal soil-colour sensing may
provide an integral way of comparing soils, including evolution,
2. Materials and methods
degradation, pedoclimate, and fertility analyses. Digital cameras
have already been proposed as soil-colour sensors (Aydemir
2.1. Developing custom image processing for a mobile phone
et al., 2004; Viscarra Rossel et al., 2008; O’Donnell et al., 2011)
due to the possibility of gaining reliable colour information from
Although there is a wide range of operating systems for devel-
RGB digital images. Currently, the only colour measure by pixels
oping applications in mobile phones such as Android, Symbian,
is based on RGB signals, and several authors (Meyer et al., 2004;
BlackBerry or Windows Mobile, the former (Android 2.2, Google,
Viscarra Rossel et al., 2006; León et al., 2006; Rodriguez-Pulido
EEUU) was selected here for several reasons. Firstly, Android is well
et al., 2012) have reported computational solutions that allow dig-
established in most communication devices, including mobile
ital images to be transformed into standard colour spaces from
phones, netbooks, tablets, and even in electrical appliances such
each pixel of the digital RGB image.
as microwaves and washing machines (Puder and Antebi, 2013).
The requirements for soil-colour measurement are available or
Secondly, Android uses an open code and a free license, allowing
can be implemented in the current typical mobile phone. This in-
a wide community of developers to expand and improve its func-
cludes a high-resolution digital camera, with a low-power high-
tionality. Thirdly, Java is the language used by Android, enabling
performance processor at running frequencies of up to 1 GHz,
the use of libraries and other applications previously developed
sophisticated operating systems offering multi-tasking, Java sup-
for this language.
port, and options for installing and running externally developed
Our software was designed to read colour information from dig-
applications. Initially, mobile phones with built-in cameras were
ital images in JPEG format captured with the built-in camera of a
used as imaging devices to collect and transmit digital data to an
smartphone and to carry out processes of colorimetric analysis.
off-site laboratory or external computer, which processed the
The smartphone HTC Desire HD (HTC Corporation, Taiwan),
information and returned the analysis results to the phone. Thus,
4.8400  2.6800  0.4600 in size and 164 g in weight, has the necessary
mobile phones have been used, for example, for biological and
computing power to run the program, also including a 4.300 touch-
forensic applications (Cadle et al., 2010), telemedicine (Martinez
sensitive screen with a resolution of 800  480 pixels and an inte-
et al., 2008), and for bad-smell monitoring of the living environ-
grated camera up to 8 megapixels with CMOS autofocus sensor. For
ment (Nakamoto et al., 2009). However, only recently, has a mobile
the device setting, we fixed the ISO parameter (sensitivity of the
phone been used for on-site data processing to determine specific
image sensor to light) to 100, according to the daylight illumina-
analyte concentrations from single-use chemical reactive mem-
tion conditions (6500 K), the white balance in the daylight option
branes considering hue changes from blue to magenta (García
for high luminance (1580 lx), and the flash in the off position.
et al., 2011).
The smartphone saved the picture in JPEG format with a size of
The potential of the mobile phone suggests that it may answer
1952  3264 pixels, a weight of around 1 MB, a resolution of 72
the increasing demand of objective soil-colour data. As this elec-
dots per inch, and using 8-bits per channel (24 bits in total for
tronic device becomes available for everyone, in contrast to the
the red, green, and blue channels).
more complex and costly colorimeters and spectrometers, its use
Once our application is installed in the mobile phone the user
as a proximal sensor in the field becomes more feasible. However,
interface on the screen may serve to show all the functionalities
never before has a mobile phone been colorimetrically tested to
(Fig. 1). At program startup, a main menu shows four options
discriminate among the colour gamut encompassed by the Munsell
(Fig. 1a). To normalize any colour to a white, in accordance with
soil charts. Thus, against a dichotomous choice of colours (García
colorimetry (CIE, 2004), a calibration option was included to deter-
et al., 2011), the mobile should now distinguish between different
mine the RGB values from the photo of a reference white (Konica
reddish, brownish, and yellowish hues, from dark to light and of
Minolta PTFE, Fig. 1b and c). A square frame over the picture could
variable intensity. On the other hand, the few works that have
be moved, expanded, or contracted, in order to select the region of
measured soil colour with digital cameras (e.g. Viscarra Rossel
interest to process. The colour information of each pixel inside the
et al., 2008) needed external software for calculations, when ad
cropped area was directly read by the phone, our program giving
hoc solutions performing the immediate computational conversion
the statistical mode for R, G, and B. For an analysis of a heteroge-
on the same platform that produced the RGB images would be
neous colour image, the mode has proved more suitable than the
desirable. Moreover, as outlined in the literatures (Hong et al.,
mean value (García et al., 2011), removing the undesirable effect
2001; Westland and Ripamonti, 2004), the RGB colour space is de-
of noisy pixels. The same RGB information could be extracted from
vice-dependent, and we cannot be sure of the effectiveness of stan-
each picture of colour samples (Fig. 1d and e), which, after being
dard transformation equations (Wyszecki and Stiles, 2000) to
normalized with the reference-white coordinates (Fig. 1c), can be
establish colorimetric coordinates from the RGB information. Final-
shown and saved in the phone (Fig. 1f). This normalization was
ly, we also suspect the possible influence of natural daylight condi-
made for each pixel following Eq. (1), where n (8 in our case) rep-
tions (Sánchez-Marañón et al., 2011), which deserves a separate
resented the number of bits per pixel for each colour channel
study. Accordingly, before going to the field, laboratory experience
(Salmeron et al., 2012).
is necessary.
In the present work, our overall goal was to investigate in the RGBnormalized ¼ 2n RGBacquired =RGBwhite ð1Þ
laboratory the potential of a mobile phone (smartphone) to capture
soil-colour images and process them with a colorimetric rationale,
returning the Munsell notations corresponding to the digital RGB 2.2. Transformation equations from RGB to XYZ and HVC
captured images. To do so, our specific objectives were: (i) to de-
velop a custom image-processing application for mobile phone; As shown in Fig. 1f, we wanted the mobile-phone application to
(ii) to build a model and estimate its parameters for establishing provide not only RGBnormalized (henceforth RGB) values, but also the
Munsell notations from RGB measurements; (iii) to implement colour designation in the CIE (XYZ) and Munsell (HVC) systems.
the image processing and the conversion model RGB ? Munsell, While the Munsell parameters are widely used in soil science,
making them work together as software in the mobile phone; the tristimulus values XYZ are the basis of instrumental colorime-
202 L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208

Fig. 1. Screen of the HTC smartphone running the Android app for measuring soil colour: (a) main menu; (b) calibration with a reference white; (c) RGB coordinates of the
reference white; (d) selection of a region of interest for one chip of the Munsel soil colour charts using a square frame; (e) a picture with a soil sample; and (f) colour
coordinates of a soil sample.

try. The two possible solutions to find XYZ and HVC from the origi- pany, 2000) which were completely new. The measurements of
nal RGB were either standard transformation equations as those each chip were undertaken with the HTC smartphone running
collected by Viscarra Rossel et al. (2006) or empirical equations our image-processing application for determining its RGB vari-
established from our colour measurements. ables, as well as with a spectroradiometer Konica Minolta
The standard equations were processed by using the ColoSol CS2000 (Tokyo, Japan) for registering the reflected spectral-power
software (Viscarra Rossel, 2006) while, for developing empirical distribution between 380 and 780 nm at 2 nm steps. For calculat-
equations, we used the colour measurements of 238 chips from a ing the XYZ tristimulus values, we assumed the CIE 1964 Standard
recent edition of Munsell soil-colour charts (Munsell Color Com- Observer (CIE, 2004). In a dark room, the chips of each Munsell

Fig. 2. Measuring process: (a) geometry, and (b) a photograph under the experimental conditions with a Munsell chart and the reference white.
L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208 203

chart together with the reference white were placed inside a Gre- by measuring each of the visually selected ones (100 samples) with
tagMacbeth Spectralight III lighting booth (X-Rite, Switzerland) a Konica Minolta 2600d spectrophotometer (Tokyo, Japan) to
equipped with a D65 simulator (Fig. 2), in agreement with the quantify the Munsell colour. This instrument has an illuminating/
standard ASTM (2008) for the Munsell colour determination. We viewing geometry d/8 and two Xenon lamps as the light source,
fixed the position of smartphone and spectroradiometer and making measurements of the light reflected by the specimen sur-
moved the Munsell chart to focus the image on the centre of each face with the specular component excluded between 360 and
chip. Three replicates of each measure were taken. 740 nm at 10 nm intervals. Among the colour indices in the output
Using the measured variables RGB and XYZ, in addition to the of display, this instrument provided the Munsell colour parameters
HVC variables taken from the notation of each chip in the Munsell HVC. The final selection consisted of 40 NCS samples which Mun-
charts, we computed polynomial transformations between colour sell colour parameters ranged between 0.1YR and 5.5Y in hue, be-
spaces (Johnson, 1996) by the pseudo inverse method (Penrose, tween 2.8/ and 8.2/ in value, and between /0.5 and /8 in chroma, all
1955), in order to build statistical prediction models of XYZ and these colours being approximately homogenously distributed over
HVC from RGB. In the matrix relation i = Td, where i (from indepen- the colour space corresponding to the seven Munsell soil-colour
dent device colour) is XYZ or HVC (3  1 matrix), d (from dependent charts.
device colour) is an RGB n  1 matrix, and we calculated the T ma- Subsequently, we tested the application in soil samples. Like the
trix of dimension 3  n (n = number of polynomial coefficients) NCS samples, the colours of our soil samples were within the space
which provided the nearest relationship between vectors i and d. of Munsell soil-colour charts, but now we managed textured and
These transformations were made by using Matlab2009b (Math- heterogeneous colour samples, which implied a more challenging
works, EEUU) and following the steps suggested by Westland and test for the mobile phone. To cover the widest possible soil-colour
Ripamonti (2004). The partition of Munsell hue for the calculations range of our pedogenic environment, we inspected the soil samples
was 10 for 10R, 12.5 for 2.5YR, and so forth up to 22.5 for 2.5Y, and gathered in our laboratories. Finally, 45 samples were chosen, all
25 for 5Y. In addition to statistical coefficients, the accuracy of the from Mediterranean soils with different degrees of development
models was examined by the differences between measured and such as Entisols, Inceptisols, Vertisols, and Alfisols. These samples
predicted values using the CIELAB colour-difference formula DEab had also been used previously in colorimetric studies to evaluate
(CIE, 2004) and the Munsell colour-difference formula DEM (God- the effects of illumination, sample state, and observer on the
love, 1951), depending on whether RGB was transformed to the soil-colour determination by Munsell charts (Sánchez-Marañón
CIE XYZ or Munsell colour systems, respectively. et al., 1995, 2005, 2011). As usual in laboratory colour studies (Tor-
rent and Barrón, 1993), we prepared air-dried samples of fine earth
2.3. Validation of the mobile-phone application (soil particles < 2 mm) for the 45 study cases plus 15 samples taken
from the same group but grinding and homogenizing them in an
To assess the accuracy of mobile-phone application for colour agate mortar until we obtained a powder with particle sizes of less
sensing in different samples from those used to develop it, we em- than 50 lm. After the powder was placed in circular plastic con-
ployed new soil-colour objects and an increasingly stringent vali- tainers (15 mm in diameter and 5 mm thick) with the upper sur-
dation plan. First, we selected colour samples from the Natural face open and levelled, their spectrophotometric colour was
Colour System (NCS, Sweden) Atlas covering the complete colour measured using the Konica Minolta 2600d spectrophotometer,
gamut encompassed in the Munsell soil-colour charts. Although the results ranging from 2.9YR to 2.3Y in hue, 3.8 to 7.0 in value,
the characteristics of these colour samples were similar to those and 1.9 to 5.6 in chroma.
of the chips of Munsell soil-colour charts, i.e. pieces of artificially All colour samples for the validation process were also mea-
coloured plain paper, they were made with different pigments sured with the HTC smartphone and spectroradiometer under
and the resulting colours did not match those of the Munsell chips the same experimental conditions described in Fig. 2. In addition,
but rather matched intermediate colours. The selection process the colour of soil samples was also tested with another mobile
among available 1950 NCS colours was initially visual, followed phone, a Samsung Galaxy S2 (Samsung, South Korea) smartphone.

Table 1
Polynomial models for transforming RGB to XYZ and HVC. The r, DEab , and DEM are, respectively, Pearson’s correlation coefficient, mean CIELAB and mean Godlove colour
difference between measured and predicted values in the 238 chips of Munsell soil-colour charts.

Terms of the polynomial r for XYZ models r for HVC models DEab DEM
[1,R,G,B] 0.9811 0.9090 9.65 2.03
[1,R,G,B,RGB] 0.9943 0.9101 6.03 1.53
[1,R,G,B,RG,RB,GB] 0.9965 0.9246 3.44 1.12
[1,R,G,B,RG,RB,GB,RGB] 0.9965 0.9259 3.39 1.12
[1,R,G,B,RG,RB,GB,R2,G2,B2] 0.9973 0.9547 2.14 1.03
[1,R,G,B,RG,RB,GB,R2,G2,B2,RGB] 0.9973 0.9562 2.07 1.02
[1,R,G,B,RG,RB,GB,R2,G2,B2,RGB,R3,G3,B3] 0.9974 0.9568 1.85 1.02
[1,R,G,B,RG,RB,GB,R2,G2,B2,RGB,R2G,G2B,B2R,R2B,G2R,B2G,R3, G3,B3,R2GB,RG2B,RGB2] 0.9980 0.9695 1.75 0.97
[1,G,B,RG,R2,G2,RGB,B3] 0.9972 0.9407 2.03 1.08

Table 2
2 3
Coefficients of the polynomial statistical models /ðRGBÞ ¼ a þ bG þ cB þ dRG þ eR2 þ fG þ gRGB þ hB built to compute XYZ and HVC from RGB.

a b c d e f g h
X 1.3502 0.0759 0.0082 1.05E04 0.000381 0.000124 2.79E06 1.31E06
Y 0.7246 0.1146 0.0172 3.20E04 3.06E04 0.000604 3.08E06 1.54E06
Z 2.1291 0.1028 0.0879 4.80E04 1.04E04 0.000174 2.41 E06 2.13E06
H 10.5649 0.4622 0.2192 2.54E03 1.92E04 6.09E04 7.98E06 4.64E06
V 2.3252 0.0303 0.0069 5.36E05 4.33E05 1.61E05 2.40E07 2.96E08
C 1.8311 0.0157 0.0351 3.80E04 2.89E04 0.000163 3.00E07 2.25E07
204 L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208

100 35
Polynomial Polynomial
Colosol 30 Colosol
80
X- Computed 25

H - Computed
60 20

40 15

10
20
5

0 0
0 20 40 60 80 100 0 5 10 15 20 25 30 35
X- Spectroradiometer H - Munsell soil charts
100
Polynomial Polynomial
8
Colosol Colosol
80
Y - Computed

V - Computed
60

4
40

20 2

0 0
0 20 40 60 80 100 0 2 4 6 8
Y- Spectroradiometer V- Munsell soil charts
100 14
Polynomial Polynomial
Colosol 12 Colosol
80
10
Z - Computed

C - Computed

60
8

40 6

4
20
2

0 0
0 20 40 60 80 100 0 2 4 6 8 10 12 14
Z- Spectroradiometer C - Munsell soil charts

Fig. 3. Colour coordinates (XYZ and HVC) measured vs. computed from the RGB signals of HTC smartphone using empirical polynomial transformations and ColoSol software
(n = 238 chips of the Munsell soil-colour charts).

This phone has a dual core processor and runs the version Android
250 4.1.2. The built-in camera has also 8 megapixels and it was used
setting equals parameters that described before for the HTC smart-
200 phone except the resolution, which in this case was of
3264  2448 pixels, increasing the weight of each picture to
RGB mobile

2 MB. In this way, the same Android application was installed


150 and used by a device other than the one used in the design, in order
to compute again the soil colours under laboratory conditions. All
100 colour measurements were replicated three times to assess their
R reproducibility.
50 G
B 3. Results and discussion
0
0 50 100 150 200 250 3.1. Transformation equations from RGB to XYZ and HVC
RGB from XYZ
Once the image processing worked as expected in the design
Fig. 4. RGB coordinates calculated with ColoSol from XYZ-spectroradiometric (Fig. 1), the following major hurdle was to transform the RGB sig-
measurements vs. RGB coordinates measured with the HTC smartphone. nals of the mobile phone to CIE XYZ and Munsell HVC colour
L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208 205

Table 3
Some statistics of the relationships between colour coordinates measured with the spectroradiometer (XYZ) or annotated in the Munsell soil-colour charts (HVC) and computed
from the RGB values registered by the HTC smartphone, using the empirical polynomial transformation and ColoSol conversion program (n = 238 chips of the Munsell soil-colour
charts, r = Pearson’s correlation coefficient, GFC = goodness of fit coefficient, RMSE = root mean square error, STRESS = standardized residual sum of squares). The mean and
standard deviation of CIELAB (DEab) and Munsell (DEM) colour differences are also given.

Polynomial ColoSol
X Y Z X Y Z
r 0.9971 0.9972 0.9973 0.9942 0.9945 0.9904
GFC 0.9992 0.9992 0.9990 0.9964 0.9948 0.9892
RMSE 1.4 1.3 1.2 2.9 4,3 5.5
STRESS 4.0 4.1 4.5 8.5 10.2 14.6
Mean DEab 2.0 14.0
SD DEab 1.1 4.5
H V C H V C
r 0.8313 0.9977 0.9708 0.8795 0.9965 0.9764
GFC 0.9882 0.9998 0.9919 0.9776 0.9964 0.9924
RMSE 2.75 0.12 0.52 5.4 0.6 2.0
STRESS 15.31 2.07 12.7 21.0 8.5 12.3
Mean DH/DV/DC 2.1 0.1 0.4 4.7 0.5 1.8
SD DH/DV/DC 1.7 0.1 0.3 2.6 0.3 0.9
Mean DEM 1.1 3.2
SD DEM 0.5 1.1

100 matrices were developed for different conditions (Wyszecki and


Stiles, 2000) from those used here, the RGB values of a same colour
might change because of their device dependence (Hong et al.,
80 2001; Westland and Ripamonti, 2004) or the influence of light
intensity (Viscarra Rossel et al., 2008), and the standard linear
transformations could not exactly match that of real cameras. Cer-
60
tainly, the RGB measured from the mobile phone and the RGB lin-
%

early transformed with ColoSol from XYZ of the Munsell chips


40 showed a similar relation (Fig. 4) to the typical curve for image de-
vices, indicating a non-linear response of the sensor (Westland and
Ripamonti, 2004). Whatever the reason, as in other studies dealing
20 Soil samples with the transformation between colour spaces (Johnson, 1996),
Munsell the mean colour differences between measured and estimated val-
NCS ues were lower using the polynomial transformations (Table 2), as
0
0 1 2 3 4 5 6 7 8 9 10 concluded from results shown in Table 3.
The prediction from the mobile phone of tristimulus XYZ proved
ΔE* ab<X
more accurate than did the Munsell parameters HVC with either
 transformation, according to the statistical coefficients listed in Ta-
Fig. 5. Percentage of samples with a CIELAB colour difference DEab between
measured (spectroradiometer and estimated (HTC smartphone) values lower than a ble 3. These coefficients are the Pearson’s linear correlation coeffi-
fixed X value in 238 Munsell chips, 60 soil samples, and 40 NCS samples. cient, the goodness of fit (GFC) coefficient (Romero et al., 1997), the
root mean square (RMSE) error, and the standardized residual sum
of squares (STRESS) index (García et al., 2007), which values for
coordinates. In addition to the simplest solution, i.e. standard col- perfect fit are 1, 1, 0 and 0, respectively. The estimates of H were
our-space transformations available in the colour literature (e.g. the worst. This indicates that using a spectroradiometer as a refer-
Viscarra Rossel et al., 2006), we also tested statistical transforma- ence made it possible to find better transformation equations, pre-
tion models developed from the data measured in the laboratory sumably because the mobile-phone camera and spectroradiometer
using the Munsell soil-colour charts. Table 1 shows some of these worked under the same illumination conditions (D65), including
statistical models, from a model of order one with four terms to a light quality and intensity, whereas Munsell colour space was de-
polynomial complex dependence of order three and 23 terms, fined with illuminant C. On the other hand, the RGB and XYZ colour
although we found that cubic models with eight coefficients (last spaces have similar geometries, whereas the cylindrical structure
row in Table 1) optimised the computational work to be made by of Munsell system, where H is the angular coordinate, could ham-
the smartphone providing a high correlation and relatively low col- per the transformation functions. It is known that when the Mun-
our differences between measured and predicted values. Table 2 sell radial coordinate C is low the uncertainty in the angular
lists the coefficients of the two cubic polynomial transformation coordinate H is higher, because H becomes indeterminate at null
equations forming matrices of dimension 3  8. For the calcula- chroma C. In our case, for example, of the 35 chips (16 in the chart
tions, we used the mean of three replicated measures on each sam- 5Y and 12 in the chart 10R) estimated with DH > 4 units, 72% had
ple, although its dispersion (standard deviation) was negligible C 6 2, 20% were extremely light and chromatic (7/8, 8/8), and 8%
both for RBG (<0.01) and XYZ (<0.03) because of the stability of had C = 3. Accordingly, our H model performed worst for extreme
the light source and homogeneity of the measured surface. colours located in the bottom left corner and upper right corner
The polynomial transformations proved slightly better than the of some Munsell charts.
standard transformations using the ColoSol software (Fig. 3). This With the polynomial transformation, the average of the colour
latter underestimated the colour coordinates, except Munsell C, differences between measured and predicted values in the 238
which was overestimated, while Z and H were the worst estimated chips of Munsell charts was 2.0 ± 1.1 CIELAB units. It is noteworthy
parameters. This may be because the standard transformation that more than 90% of the samples were measured by the mobile
206 L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208

phone with accuracy of 4 CIELAB units or better (Fig. 5). When it is although the ground samples exhibited a more homogeneous as-
taken into account that the theoretical colour visual threshold was pect, it did not lessen the measuring errors of mobile phone, prob-
around 1.0 CIELAB units, that the practical threshold for most col- ably because of the way in which the colour of all pixels in a
orimetric applications is around 3–4 units, and that only thresh- picture were managed in order to compute the soil colour as a
olds greater than 5–6 units should be considered erroneous whole. The mobile phone calculated the mode of the pixels present
(Melgosa et al., 1992; Huang et al., 2012), the current results sug- in an area of interest and with the use of this statistical parameter,
gest that the implementation of polynomial transformations from anomalous pixels in a picture of fine earth did not influence the fi-
RGB to XYZ or HVC in the mobile-phone application is the best op- nal colour coordinates.
tion to provide accurate colour information. Table 5, which lists the best and worst study cases with their
colour coordinates measured with the spectrophotometer (HVC),
3.2. Performance of the mobile-phone application in the validation spectroradiometer (XYZ), and mobile phone (RGB), confirmed the
samples good results for the soil samples: SOIL-va, SOIL-ck, SOIL-sn, SOIL-

The reliability of the mobile phone to detect colour was initially 100
checked in the NCS samples covering the colour gamut of the Mun- NCS
sell soil-colour charts. Due partly to the even surface and homoge- 80 Soils
nous colour of the samples, as well as the stability of illumination,
the replicated measurements in each sample had negligible disper-
sion (standard deviation <0.06 in RGB with the mobile phone and 60
<0.98 in XYZ with the spectroradiometer and <0.21 in HVC with

%
the spectrophotometer). The differences in each Munsell colour
40
coordinate (absolute values) between the determinations made
with the app in the HTC mobile phone and the ones provided by
our spectrophotometer are shown in Fig. 6 and the averages are 20
listed in Table 4. The hue differences (DH) ranged between 4
and 7, with 70% of the samples within the [2, 2] interval. A closest
0
agreement was found in Munsell value and chroma, with a range of
0 1 2 3 4 5 6 7 8 9 10
DV and DC differences [2, 2] and [0, 2] units, respectively, most of
ΔH
the colour samples being within the interval [0, 1]. Despite the low
mean value in DH (Table 4), its standard deviation was on the same 100
order, indicating that some colour samples had a major error.
NCS
According to the CIELAB differences found between the mobile
phone estimates and spectroradiometric measurements (CIE tri- 80 Soils

stimulus values XYZ), the NCS samples with the worst results were,
like the above-mentioned major error in Munsell chips, samples 60
yellower than 2.5Y (H > 22.5) and with low chroma or very light
and chromatic (samples NCS-25, NCS-22, NCS-9, and NCS-1 in Ta-
%

ble 5). Although the mean difference was 6.45 CIELAB units, it 40
reached up to 9.4 CIELAB units in some extreme samples (Table 5),
which produced the high standard deviation value of 2.05 CIELAB
20
units listed in Table 4. Around 50% of the NCS samples were mea-
sured using the HTC mobile phone with an error of less than 6 CIE-
LAB units (Fig. 5). 0
The measuring errors of the mobile phone dramatically de- 0 1 2 3 4 5 6 7 8 9 10
creased in soil samples with respect to those in NCS samples when ΔV
the spectroradiometer was considered as the reference (Fig. 5), and
consequently their mean DEab was lower in Table 4, but increased
100
somewhat when the spectrophotometer was considered (Fig. 6),
NCS
primarily DH. On the one hand, this different behaviour could be
Soils
attributed both to the illumination conditions, identical only for 80
the phone camera and spectroradiometer, and the problems once
again with the Munsell space geometry. On the other hand, the
60
best DEab results in fine-earth soil samples, despite their uneven
and textured surface with heterogeneous soil-pigment mixtures,
%

might be explained by the fact that, although covering a wide ga- 40


mut from 2.9YR to 2.3Y, these Mediterranean soil samples did
not have extreme colours in hue, lightness, or chromaticity, which
must have improved the performance of the mobile phone. Proba- 20
bly the relatively worse DEab results for the ground and homoge-
nized soil samples than for fine-earth soil samples (Table 4) was
0
also due to a somewhat more limiting colour gamut in the 15 soils 0 1 2 3 4 5 6 7 8 9 10
from which fine earth as well as ground samples were prepared.
For these 15 soils (on the average, 1.0Y and 3.0 in chroma), how- ΔC
ever, there was no statistically significant difference (P < 0.05) be- Fig. 6. Histograms for NCS (n = 40) and soil (n = 60) samples with the differences
tween the mobile-phone errors in fine earth (mean DEab ¼ 4:46) (absolute value) in Munsell hue (DH), value (DV), and chroma (DC) between data
and ground (mean DEab ¼ 5:04) samples. The results indicate that measured with the spectrophotometer and predicted with the HTC smartphone.
L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208 207

ll, and SOIL-sh even with errors close to the theoretical colour were measured by the mobile phone with an error lower than 5
threshold (1–2 CIELAB units). There was, however, an apparent CIELAB units (Fig. 5). It also bears mentioning that this error by

lack of consistency in the magnitudes of CIELAB DEab and Munsell the mobile phone in sensing colour is lower than those previously
(DH, DV, DC) differences between measured and predicted values. reported when the same soil samples were visually determined
Thus, for example, SOIL-sh had greater CIELAB difference (1.9 using Munsell soil-colour charts. Specifically, a group of 10 soil sci-
units) than SOIL-va (1.3 units), just the opposite behaviour that entists judging soil colour under controlled illumination had a
Munsell differences (0.3, 0.2, 0.1 against 0.6, 0.1, 0.2, for DH, DV, mean error of 4.4 CIELAB units in ground soil samples and 10.2
DC respectively). This fact, common in Table 5, has already been in aggregated soil samples with respect to instrumental measure-
statistically-discussed with the results of Table 4 and Figs. 5 and ment, and their inter-observer variability, i.e. disagreement among
6, indicating that there was no a good correlation between spectro- the ten Munsell notations on a given sample, was 5.1 CIELAB units
photometric and spectroradiometric measurements. In this way, (Sánchez-Marañón et al., 1995). Differences of up 5.5 CIELAB units
higher values in DEab did not necessarily correspond to higher were also measured by Sánchez-Marañón et al. (2005) in Munsell
errors in HVC and vice versa. soil-colour charts from different editions, manufacturers, and de-
Considering the entire dataset of soil samples (n = 60), we found gree of use, so inducing the same error in the soil colour determi-
that the discrepancies between the HTC smartphone and spectro- nation. In addition, the error made by the smartphone estimating
photometer ranged from 6 to 6 for Munsell hue, so that 60% of Munsell hue (59.0% of soils with DH > 2 units, Fig. 6) was lower
samples were determined with an error of less than 3.0 units than the variation in Munsell hue of soils observed under different
(Fig. 6). For Munsell value and chroma, most soil samples had an er- natural daylight conditions (78.6% of soils with DH > 2 units, Sán-
ror of less than 1.0 Munsell unit. The total colour difference be- chez-Marañón et al., 2011).
tween HTC smartphone and spectroradiometer was on the Finally, once the mobile-phone application developed with the
average 3.7 ± 1.8 CIELAB units, and more than 80% of soil samples HTC smartphone was installed in another device (Samsung smart-

Table 4

Differences in Munsell hue, value, and chroma (DH, DV, and DC in absolute value) and CIELAB colour differences DEab between data measured with a spectrophotometer (HVC)
and a spectroradiometer (XYZ) and estimated with two mobile phones (HTC and Samsung).

Group of samples Colour parameter Mean value Standard deviation


NCS samples DH 1.69 1.69
HTC smartphone (n = 40) DV 0.66 0.22
DC 0.36 0.34
DEab 6.45 2.05
Ground soil samples DH 2.05 1.23
HTC smartphone (n = 15) DV 0.57 0.31
DC 0.32 0.20
DEab 5.04 2.47
Fine-earth soil samples DH 2.72 1.70
HTC smartphone (n = 45) DV 0.57 0.36
DC 0.81 0.60
DEab 3.31 1.59
Soil samples (fine-earth plus ground samples) DH 3.38 1.55
Samsung smartphone (n = 60) DV 0.63 0.26
DC 0.25 0.17
DEab 5.46 2.46

Table 5

The five NCS- and SOIL-samples with the lowest (italics) and highest CIELAB colour differences DEab between the colour measured with a spectroradiometer (XYZ, mean of 3
replicates with SD < 1.38) and predicted from the RGB values (mean of 3 replicates with SD < 0.11) by the HTC smartphone. The Munsell HVC measured with a spectrophotometer
(mean of 3 replicates with SD < 0.24) and the differences (DH, DV, DC) with the predicted values by the mobile phone are also listed.

Sample H V C X Y Z R G B DH DV DC DEab
NCS-38 16.9 7.7 0.5 55.3 57.6 63.5 222.0 202.0 206.0 4.6 0.4 0.3 2.3
NCS-28 12.9 7.5 1.1 57.2 58.9 62.2 231.0 202.0 198.0 1.9 0.7 0.3 2.6
NCS-21 10.4 3.7 6.6 14.8 11.3 5.5 165.0 41.0 41.0 1.1 0.4 0.1 2.7
NCS-30 13.7 6.5 7.0 48.3 43.1 22.7 255.0 145.0 107.0 2.3 0.9 0.6 3.9
NCS-35 15.1 6.3 8.0 44.7 39.8 16.1 255.0 136.0 74.0 3.6 0.9 1.3 4.2
SOIL-va 21.7 4.4 1.9 12.7 13.0 9.5 105.6 75.2 56.3 0.6 0.1 0.2 1.3
SOIL-ck 21.4 6.0 2.7 30.7 31.6 21.9 166.4 135.6 94.2 3.5 0.0 0.1 1.7
SOIL-sn 19.9 4.9 3.6 21.9 21.2 12.7 149.4 98.0 56.3 2.3 0.2 0.0 1.8
SOIL-ll 19.9 4.0 2.6 20.6 20.9 13.3 140.8 98.0 65.5 2.1 1.0 0.3 1.8
SOIL-sh 19.5 3.8 2.6 11.9 11.2 7.3 105.6 59.1 39.9 0.3 0.2 0.1 1.9
NCS-4 18.2 6.2 5.0 35.2 34.2 18.2 231.0 156.0 99.0 0.3 1.0 0.6 9.2
NCS-25 23.1 3.3 2.7 6.7 6.9 3.8 99.0 65.0 41.0 1.6 0.8 0.3 9.4
NCS-22 23.3 7.6 2.1 41.3 43.6 34.4 222.0 193.0 156.0 1.3 0.2 0.2 9.4
NCS-9 25.5 4.3 2.5 14.7 15.7 10.2 140.0 112.0 74.0 0.4 1.0 0.0 9.4
NCS-1 22.5 7.1 6.3 40.0 40.9 16.1 239.0 179.0 82.0 0.1 0.6 0.0 9.4
SOIL-vg 22.6 5.1 2.3 23.4 24.3 19.2 157.9 124.8 85.0 1.8 0.6 0.4 7.6
SOIL-dd 19.8 5.1 3.7 26.9 27.0 15.1 181.0 110.4 74.0 1.6 0.6 0.8 7.8
SOIL-ps 15.9 4.0 5.0 20.2 18.3 6.8 167.7 65.4 33.0 0.5 0.7 1.3 7.9
SOIL-mo 22.2 5.4 2.1 23.5 24.6 17.1 178.3 129.1 109.7 4.2 0.5 0.6 9.8
SOIL-bc 20.5 6.1 3.9 32.8 33.2 19.6 221.3 162.4 103.4 1.2 1.1 0.8 11.1
208 L. Gómez-Robledo et al. / Computers and Electronics in Agriculture 99 (2013) 200–208

phone), we measured the soil samples again under the same exper- García, A., Erenas, M.M., Marinetto, E.D., Abad, C.A., de Orbe-Paya, I., Palma, A.J.,
Capitán-Vallvey, L.F., 2011. Mobile phone platform as portable chemical
imental conditions (Fig. 2). The results with the Samsung smart-
analyzer. Sensor. Actuat. B – Chem. 156, 350–359.
phone were only slightly worse than using the HTC (Table 4) but Godlove, I.H., 1951. Improved color-difference formula, with applications to the
still acceptable. This indicates similarities in the performance of perceptibility and acceptability of fadings. J. Opt. Soc. Am. 41, 760–772.
both mobile-phone cameras to capture the RGB signals and facili- Hong, G., Luo, M.R., Rhodes, P.A., 2001. A study of digital camera colorimetric
characterization based on polynomial modeling. Color Res. Appl. 26, 76–84.
ties in the operating system to interchange the software. Huang, M., Liu, H., Cui, G., Luo, M.R., Melgosa, M., 2012. Evaluation of threshold color
differences using printed samples. J. Opt. Soc. Am. A 29, 883–891.
IUSS Working Group WRB, 2006. World reference base for soil resources 2006.
4. Conclusions World Soil Resources Reports 103.
Johnson, T., 1996. Methods for characterizing colour scanners and digital cameras.
Our results indicate that the technical resources of the current Displays 16, 183–191.
León, K., Mery, D., Pedreschi, F., León, J., 2006. Color measurement in L⁄a⁄b⁄ units
mobile phones can be exploited to use these electronic devices,
from RGB digital images. Food Res. Int. 39, 1084–1091.
which are prevalent worldwide and accessible to everyone, as Martinez, A.W., Phillips, S.T., Carrilho, E., Thomas 3rd., S.W., Sindi, H., Whitesides,
soil-colour sensors. The RGB signals captured by the camera and G.M., 2008. Simple telemedicine for developing regions: camera phones and
paper-based microfluidic devices for real-time, off-site diagnosis. Anal. Chem.
converted in colour coordinates by a software application running
80, 3699–3707.
inside the same smartphone, can thus achieve objective, easy, ra- Melgosa, M., Hita, E., Romero, J., del Barco, L.J., 1992. Some classical color differences
pid, and cheap colour measurements. Under controlled illumina- calculated with new formulas. J. Opt. Soc. Am. A 9, 1247–1254.
tion conditions, the measurements seem to be also more Meyer, G.E., Camargo Neto, J., Jones, D.D., Hindman, T.W., 2004. Intensified fuzzy
clusters for classifying plant, soil, and residue regions of interest from color
accurate than visual soil-colour determination using Munsell images. Comput. Electron. Agric. 42, 161–180.
charts. Munsell Color Company, 2000. Munsell Soil Color Charts. Munsell Color Co.,
Essential requirements of the mobile-phone application were a Baltimore, MD.
Nakamoto, T., Ikeda, T., Hirano, H., Arimoto, T., 2009. Humidity compensation by
moveable and visible frame on the smartphone screen to select the neural network for bad-smell sensing system using gas detector tube and built-
region of interest in the picture, a calibration or normalization tool in camera. 2009 IEEE Sensors, pp. 281–286.
for referencing colour to a standard white, and transformation NCS – Natural Colour SystemÒÓ, 2013. <http://www.ncscolour.com/en/ncs/>.
O’Donnell, T.K., Goyne, K.W., Miles, R.J., Baffaut, C., Anderson, S.H., Sudduth, K.A.,
equations to convert RGB values in Munsell or CIE colour 2011. Determination of representative elementary areas for soil redoximorphic
coordinates. features identified by digital image processing. Geoderma 161, 138–146.
In converting the image signals to colour, a polynomic-pro- Penrose, R., 1955. A generalized inverse for matrices. In: Mathematical Proceedings
of the Cambridge Philosophical Society, p. 51.
cess transform fitted to the colour gamut of the Munsell soil-col-
Puder, A., Antebi, O., 2013. Cross-compiling Android applications to iOS and
our charts worked better than standard equations from the Windows Phone 7. Mobile Netw. Appl. 18, 3–21.
colour literature, and also proved more efficient in reference to Rodriguez-Pulido, F.J., Gomez-Robledo, L., Melgosa, M., Gordillo, B., Gonzalez-Miret,
M.L., Heredia, F.J., 2012. Ripeness estimation of grape berries and seeds by
XYZ coordinates (very similar in geometry to RBG) measured
image analysis. Comput. Electron. Agric. 82, 128–133.
with a spectroradiometer under the same experimental condi- Romero, J., García-Beltran, A., Hernández-Andrés, J., 1997. Linear bases for
tions as the phone camera. Accordingly, although our initial representation of natural and artificial illuminants. J. Opt. Soc. Am. A 15,
objective was to reproduce the Munsell notation of soil samples, 1007–1014.
Salmeron, J.F., Gomez-Robledo, L., Carvajal, M.A., Huertas, R., Moyano, M.J., Gordillo,
the results suggest that CIE coordinates are more accurately pre- B., Palma, A.J., Heredia, F.J., Melgosa, M., 2012. Measuring the colour of virgin
dicted for this application. olive oils in a new colour scale using a low-cost portable electronic device. J.
Programming in Android allowed the application to be success- Food Eng. 111, 247–254.
Sánchez-Marañón, M., Delgado, G., Delgado, R., Pérez, M.M., Melgosa, M., 1995.
fully interchanged between different devices. However, to achieve Spectroradiometric and visual color measurements of disturbed and
an application working similarly in different models of mobile undisturbed soil samples. Soil Sci. 160, 291–303.
phones constitutes a challenge, because it depends mainly on the Sánchez-Marañón, M., Soriano, M., Melgosa, M., Delgado, G., Delgado, R., 2004.
Quantifying the effects of aggregation, particle size and components on the
quality of mobile-phone camera. The use of the mobile phone un- colour of Mediterranean soils. Eur. J. Soil Sci. 55, 551–565.
der non-controlled (variable) illumination conditions is another Sánchez-Marañón, M., Huertas, R., Melgosa, M., 2005. Colour variation in standard
step forward for our future research. soil-colour charts. Aust. J. Soil Res. 43, 827–837.
Sánchez-Marañón, M., García, P.A., Huertas, R., Hernandez-Andres, J., Melgosa, M.,
2011. Influence of natural daylight on soil color description: assessment using a
Acknowledgements color-appearance model. Soil Sci. Soc. Am. J. 75, 984–993.
Soil Survey Staff, 1993. Soil survey manual. USDA Handbook.
Soil Survey Staff, 1999. Soil Taxonomy: A Basic System of Soil Classification for
This work has been partially funded by Ministry of Economy
Making and Interpreting Soil Surveys, 2nd ed.
and Competitivity (Spain) under FIS2010-19839 Research Project, Torrent, J., Barrón, V., 1993. Laboratory measurement of soil color: theory and
and Junta de Andalucía (Spain) under Project PE10-TIC5997, with practice. In: Bigham, J.M., Ciolkosz, E.J. (Eds.), Soil Color, Special Publication No.
European Regional Development Fund (ERDF) support. 31, SSSA, Madison, WI, pp. 21–33.
Viscarra Rossel, R.A., 2006. ColoSol. A Colour Conversion Program for Soil Colour
v3.0. <http://www.usyd.edu.au/su/agric/acpa/people/rvrossel/>.
References Viscarra Rossel, R.A., Webster, R., 2011. Discrimination of Australian soil horizons
and classes from their visible-near infrared spectra. Eur. J. Soil Sci. 62, 637–647.
ASTM, 2008. Standard Practice for Specifying Color by the Munsell System. ASTM Viscarra Rossel, R.A., Minasny, B., Roudier, P., McBratney, A.B., 2006. Colour space
International D 1535-08, PA, USA. models for soil science. Geoderma 133, 320–337.
Aydemir, S., Keskin, S., Drees, L.R., 2004. Quantification of soil features using digital Viscarra Rossel, R.A., Fouad, Y., Walter, C., 2008. Using a digital camera to measure
image processing (DIP) techniques. Geoderma 119, 1–8. soil organic carbon and iron contents. Biosyst. Eng. 100, 149–159.
Berns, R., 2000. Billmeyer and Saltzman’s Principles of Color Technology. Wiley Viscarra Rossel, R.A., Adamchuck, V.I., Sudduth, K.A., McKenzie, N.J., Lobsey, C.,
Interscience. 2011. Proximal soil sensing: an effective approach for soil measurements in
Cadle, B.A., Rasmus, K.C., Varela, J.A., Leverich, L.S., O’Neill, C.E., Bachtell, R.K., space and time. Adv. Agron. 113, 243–291 (Chapter 5).
Cooper, D.C., 2010. Cellular phone-based image acquisition and quantitative Westland, S., Ripamonti, C., 2004. Computational Colour Science using Matlab, 1st
ratiometric method for detecting cocaine and benzoylecgonine for biological ed. Wiley, New York.
and forensic applications. Subst. Abuse 4, 21–33. Wyszecki, G., Stiles, W.S., 2000. Color Science: Concepts and Methods, Quantitative
CIE, 2004. Publication 15:2004. In: Bureau, C.C. (Ed.), Colorimetry, 3rd ed., Vienna. Data, and Formulae. John Wiley & Sons, New York.
García, P.A., Huertas, r., Melgosa, M., Cui, G., 2007. Measurement of the relationship
between perceived and computed color difference. J. Opt. Soc. Am. A 24, 1823–
1829.

View publication stats

Potrebbero piacerti anche