Analysis of Air Temperature Distribution Inside a Cold Store by Means

of Geostatistical Methods
R. Gil and C.R. Bojac
Departamento de Ciencias Bsica
Universidad de Bogot
Jorge Tadeo Lozano
Bogot
Colombia

E. Schrevens
Department of Biosystems
Faculty of Bioscience Engineering
Katholieke Universiteit Leuven
Belgium

R. Suay
Centro de Agroingeniera
Instituto Valenciano de Investigacin Agrarias
Moncada, Valencia
Spain
Keywords: geostatistics, temperature distribution, cold store
Abstract
The objective of temperature control in cold storage systems is to maintain an
average temperature limited by a maximum temperature deviation within the
produce. Temperature is a key factor in postharvest quality that is usually
supported by the belief that it is well controlled and uniform in cold storage systems.
Cold stores show non-uniform air temperature distributions and produce quality
problems related to cold and hot spots are well-known. Air temperature control in
storage systems is usually related to one or a limited number of temperature sensors
in the store that are placed in locations easy accessible. The aim of this study was to
model and analyze spatio-temporal temperature distributions in vertical plans along
a cold store by means of a geostatistical approach. The use of this information to
determine the minimum number of air temperature sensors needed and their
location for better climate control is discussed. The spatio-temporal air temperature
variation was assessed inside an experimental cold store applying geostatistical
methods. The cold store (3.752.642.38 m) was equipped with an A/C unit. A
vertical grid of 36 regularly distributed air temperature sensors was installed. It was
shifted 0.35 m each day in order to cope with the 3.75 m length of the cold store in 10
vertical planes. Each plane was set to take measurements during 24 hours periods at
1 datum/min frequency. Empirical variograms were constructed for each vertical
plane. Various variogram models were evaluated to fit the dataset, and an ordinary
spatio-temporal kriging method was used to predict temperature values. The results
show temperature distribution gradients that have to be accounted for when
studying postharvest processes in cold storage systems. Geostatistical methods
proved to be a useful tool for the study of spatio-temporal air temperature
distributions that will contribute to better postharvest produce uniformity through
improved climate control performance.

INTRODUCTION
Cold storage is one of the oldest methods of food preservation; and nowadays, one
of the preferred methods to keep almost unchanged the original characteristics of
perishable foods for a relatively long time (Potter, 1992). However, in the case of
perishable products including fruit, it is important to start cooling within the first hour
after harvest to avoid loss of quality and reduction in quantity of marketable product
(Maxie et al., 1959). Inside a cold store temperature management is the most important
factor for minimizing the deterioration of perishable fruit and maximizing product
lifetime (Mitchell et al., 1996).
During cold storage, the products and the surrounding air continuously exchange
Proc. 4th International Conference
Postharvest Unlimited 2011
Eds.: P.M.A. Toivonen et al.
Acta Hort. 945, ISHS 2012

29

heat, which plays an important role in the preservation and quality of the stored product.
The energy transport mechanism depends largely on the interaction between air currents
and the distribution of product stored within a room; changes in air currents cause
heterogeneity temperature. Thus, inadequate control of air currents and hence the
temperature, leads to quality decline of the product (Nahor et al., 2008). Thus, the main
purpose of the design and management of cold stores for storage of perishable products is
to achieve a uniform temperature distribution.
However, the airflow inside a cold store is a complex process. In recent years, it
has been shown that the use of Computational Fluid Dynamics (CFD) for solving
mathematical models of transport phenomena can reduce costs associated with the design
and optimization of food processing (Scott and Richardson, 1997; Xia and Sun, 2002),
and especially cold storage, equipment (Mirade et al., 2002; Nahor et al., 2008; Delele et
al., 2009). In this work we present a new approach based on studying the structure of air
temperature spatial dependence, which helps to understand the spatial variations of the
temperature inside a cold store.
The aim of this study was to analyze the spatio-temporal variation of air
temperature inside a cold store by means of a geostatistical modeling approach.
MATERIALS AND METHODS
A cold store (3.752.642.38 m) was equipped with an A/C unit. A vertical grid
of 36 regularly distributed air temperature sensors was installed. The grid was shifted 10
times (0.35 m each day) in order to cope with the 3.75 m length of the cold store. On each
vertical plane (P1 to P10) the sensors were fastened over one fine metallic system
support. The sensors were connected to an automatic temperature recorder that was set to
take measurements during 24 hour periods at 1 recording min-1 frequency. The cold store
was set to maintain a temperature of 12C. A geostatistical approach was used to analyze
the spatial distribution of the temperature and to predict the values at unsampled locations
inside the cold store. The work was divided into three main phases: exploratory data
analysis, structural analysis, and prediction at unsampled positions (Isaaks and Srivastava,
1989). Overall temperature averages were calculated for each plane and for the entire
measurement period to estimate the semi-variance. The first step was to carry out an
exploratory analysis focused on determining data trends and the distribution function of
the variable. The purpose was to identify the basic behavior of data through descriptive
statistics such as the average, standard deviation, skewness and kurtosis. Additionally, the
time series were drawn at each plane with the corresponding boxplot to determine the
temporal variation of temperature in each plane.
The structural analysis consisted of the determination of a valid spatial
dependence function for the empirical semivariance. The graphical representation of the
relationship between the semivariance h of a variable and the separation among points
(lag) is called the semivariogram. The semivariogram indicates the degree of spatial
dependence of the variable that decreases as distance increases. Isotropic classical
semivariograms was estimated using the following equation (Matheron, 1963):

(1)

where N(h) is the number of distinct pairs lagged by vector h, and Z(xi) is the observed
value at position xi. The shape of an empirical semivariogram can be fitted to a variety of
theoretical models; for the present work the performance of three models was evaluated.
The Cauchy, Gaussian and Exponential models defined by Cressie (1993) were tested to
fit the points of the semivariogram. Semivariogram models are defined by three
parameters: sill (c), nugget (c0), and range (a), where c is the ordinary sample variance,
equal to the sum of parameters nugget (c0) and partial sill (c1). Parameter c represents the
upper asymptote of the semivariogram, and it corresponds to the highest variability
explained by the semivariogram. The nugget is a consequence of both sampling error and
30

spatial variability at the minimum distance interval. The third parameter is the range (a)
that represents a measure of the maximum distance of influence.
Estimation of parameters a, c0 and c were done using the restricted maximum
likelihood method (Diggle and Ribeiro, 2007). With this method, semivariogram
parameter estimates are often less biased in comparison with the estimates made with the
maximum likelihood method, as the maximum likelihood is applied to the error contrasts
rather than the data themselves (Pleydell et al., 2004). The selection of the semivariogram
model to be used for the prediction of temperatures at unsampled locations was based on
three criteria: the estimated values of Akaike information criterion (AIC), Bayesian
information criterion (BIC) and the ratio between the nugget (c0) and sill (c), expressed as
a percentage. It is desirable that this ratio does not exceed 50% otherwise the prediction
error will be significant. The model with the lowest values for these three goodness-of-fit
measures was selected as the semivariogram model to represent each plane dataset.
Finally, the temperature values at unsampled locations were predicted by applying
the previously fitted model. Geostatistics offers many methods that provide estimates at
unsampled locations (Castrignano and Buttafuoco, 2004). Temperature profiles for each
plane were predicted based on a regular 11 cm grid, using the block ordinary kriging
technique. Ordinary kriging is considered the anchor algorithm of geostatistics because of
its robustness under different conditions (Deutsch and Journel, 1992). Maps obtained by
the kriging prediction provided a visual representation of the temperature distribution.
The ordinary Kriging prediction of (xa) at location xa is given by the following equation:

Z x
with

Z x

(2)

where i is the assigned weight, based on the fitted model to the empirical semivariogram,
for data Z(xi) from the n sampled locations. All statistical analysis were done with the
statistical software R (R Development Core Team, 2010) using the geoR package
(Ribeiro and Diggle, 2001).
RESULTS AND DISCUSSION
Average temperatures (standard deviation) for each plane are shown in Table 1.
At first sight there do not seem to be differences between the average temperatures for
each plane, only P1 showed a difference of 1C (problems related to the cold store
operation led to a decision to not considerate plane 1 data and results). All planes showed
average temperatures higher than the set point (12C). Kurtosis values around 3 and
skewness values close to 0 are associated to datasets with a distribution close to normality
(Lipschutz and Schiller, 2000). On the other hand, although the temperature has
fluctuations there is no time trend (Fig. 1). The similarity between the planes is that the
whole cold store operates under the same conditions, from the temporal point of view.
According to the results, temperature distribution for each plane was close to
normality, which is a desirable characteristic for geostatistical analysis. In order to select
the best semivariogram model, AIC, BIC, and c0/c ratio were estimated for each plane.
The exponential model showed the lowest values for the selection criteria (Table 2); it is
then expected that this model will optimally represent the temperature spatial variability
for most of the planes inside the cold store.
The semivariogram parameters for each plane, calculated according to the selected
model, are presented in Table 2. The parameters showed important differences in their
magnitudes depending on the plane. In particular, the nugget and sill presented very
similar values for P2 to P6. On the other hand, P7 to P10 showed an important difference
between these two parameters (nugget and sill), which is an indication of the presence of
temperature gradients for those zones of the cold store. The unexplained spatial
variability percentage was estimated from c0/c. In this study, temperature spatial
31

variability was satisfactorily explained by the semivariogram model for all planes, except
for P4 and P5 whose ratio c0/c were higher than 50%. Small values of c0/c usually
indicate that higher accuracy can be obtained by mapping the studied variable. P4 and P5
presented the highest c0/c ratio values with 57.93 and 75.32, respectively. The large c0/c
ratio effect for those planes can be partly explained by the accumulated uncertainty
related to the microscale variation (e.g., the distance between sampling points). For this
reason, the estimations made by the kriging procedure will be biased and will produce a
higher error for those planes. Percentages of c0/c lower than 25% indicate a strong spatial
dependence, while percentages higher than 75% point out a high unexplained variability
and in consequence a weak spatial dependence (Ruth and Lennartz, 2008).
Inside a cold store, the temperature is regulated by means of a ventilation flow
from a cooling unit. The temperature differences among vertical planes were related to
the position of the cooling unit inside the cold store. The temperature profile for each
plane was predicted using the exponential model, are presented in Figure 3. Temperature
distribution maps indicated that the vertical gradient was higher at planes P8, P9 and P10,
i.e., those closest to the refrigeration unit. This is in agreement with the models in
Figure 2, where the largest semivariances are associated with the same planes.
Planes P7, P8 and P9 show evidence of the presence of some cold spots as a
consequence A/C unit proximity. The airflow in the cold store, driven by the fan of the
A/C unit, moves in a circular way. Towards the near-wall region the turbulence is rapidly
augmented, as consequence a large gradient in mean velocity are generated (Xie et al.,
2006). Especially at P9, a cold spot is clearly identifiable near the central top area of the
cold store. The maps for the temperature spatial prediction made for other planes
indicated a minimal gradient, but the trend was clear, higher temperatures in the upper
areas and vice versa. An appropriate airflow may guarantee the uniformity of cold air
distribution, improve the heat transfer efficiency, reduce the weight loss, save the energy
and improve the post-storage quality of foods (Xie et al., 2006).
Based on the results of this study it seems that the installation of a single sensor
for temperature control is insufficient. In the current configuration the temperature sensor
is located in the upper zone of the cold store, close to the wall and near the refrigeration
unit. According to the results of the predictions this area recorded always higher
temperatures (Fig. 3, Plane 9). Based on this result, we would suggest the installation of at
least two additional temperature sensors, which must be located in the areas of greatest
variability. Under the specific conditions in which measurements were made, the
appropriate location for these sensors would be in the middle (140 cm above the floor)
and lower (20 cm above the floor) part of plane 9. These heights are suggested based on
the estimated range (a, 161.92 cm) for the plane 9 with the exponential model (Table 2).
The algorithm that defines the activation of the airflow inside the cold store should
integrate the readings from the three sensors.
CONCLUSIONS
The results showed that the presence of temperature variations have to be
accounted for when studying postharvest processes in cold storage systems. Geostatistical
methods proved to be a useful tool for the study of spatio-temporal air temperature
distributions that will contribute to better postharvest produce uniformity through
improved climate control performance.
Literature Cited
Castrignano, A. and Buttafuoco, G. 2004. Geostatistical stochastical simulation of soil
water content in forested area of south Italy. Biosyst. Eng. 87:257-266.
Chien, Y.L., Lee, D.Y., Guo, H.Y. and Houng, K.H. 1997. Geostatistical analysis of soil
properties of mid-west Taiwan soils. Soil Sci. 162:291-297.
Cressie, N. 1993. Statistics for Spatial Data. New York, N.Y.: Wiley Inter-Science.
Delele, M.A., Schenk, A., Tijskens, E., Ramon, B.M., Nicola, H. and Verboven, P. 2009.
Optimization of the humidification of cold stores by pressurized water atomizers
32

based on a multiscale CFD model. J. Food Engin. 91:228-239.

Deutsch, C.V. and Journel, A.G. 1992. GSLIB: Geostatistical Software Library and
Users Guide. New York, N.Y.: Oxford University Press.
Diggle, P.J. and Ribeiro, P.J. 2007. Model-Based Geostatistics. New York. N.Y.:
Springer.
Isaaks, E.H. and Srivastava, R.M. 1989. An Introduction to Applied Geostatistics. New
York, N.Y.: Oxford University Press.
Lipschutz, S. and Schiller, J. (eds.). 2000. Schaums Outline of Theory and Problems of
Introduction to Probability and Statistics. McGraw-Hill, NY.
Matheron, G. 1963. Principles of geostatistics. Econ. Geol. 58(8):1246-1266.
Maxie, E.C., Mitchell, F.G. and Greathead, A. 1959. Quality study on strawberries:
Experiments with Shasta berries show harvested fruit should be protected against the
effects of high field temperatures. Calif. Agric. 13:6-6.
Mirade, P., Kondjoyan, A. and Daudin, J. 2002. Three-dimensional CFD calculations for
designing large food chillers. Comp. Electr. Agric. 34:67-88.
Mitchell, F.G., Mitcham, E.J., Thompson, J.F. and Welch, N. 1996. Handling strawberries
for fresh market. Univ. Calif., Div. Agric. Nat. Res., Pub. No. 2442.
Nahor, H.B., Hoang, M.L., Verboven, P., Baelmans, M. and Nicolai, B.M. 2008. CFD
model of the airflow, heat and mass transfer in cool stores. Internat. J. Refrig. 28:368380.
Pleydell, D.R.J., Raoul, F., Tourneux, F., Danson, F.M., Graham, A.J., Craig, P.S. and
Giraudoux, P. 2004. Modelling the spatial distribution of Echinococcus multilocularis
infection in foxes. Acta Trop. 91:253-265.
Potter, N.N. 1992. La Ciencia de los Alimentos. Editorial Andrmeda S.A. Mxico.
R Development Core Team. 2010. R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-90005107-0, URL http://www.R-project.org.
Ribeiro, P.J. and Diggle, P.J. 2001. geoR: a package for geostatistical analysis. R-NEWS,
1(2):15-18.
Ruth, B. and Lennartz, B. 2008. Spatial variability of soil properties and rice yield along
two catenas in southeast China. Pedosphere 18:409-420.
Scott, G. and Richardson, P. 1997. The application of computational fluid dynamics in the
food industry. Trends Food Sci. Technol. 8:119-124.
Xia, B. and Sun, D.W. 2002. Applications con computational fluid dynamics (CFD) in the
food industry: a review. Comp. Electr. Agric. 34:5-24.
Xie, J., Qu, X.H., Shi, J.Y. and Sun, D.W. 2006. Effects of design parameters on flow and
temperature fields of a cold store by CFD simulation. J. Food Engin. 77:355-363.

33

Tables

Table 1. Descriptive statistics for the datasets measured on each individual vertical plane
(P1 to P10) used to analyze the spatio-temporal variation of temperature inside an
experimental cold store.
Plane
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10

34

Average Standard deviation

Skewness Kurtosis
(C)
(C)
14.11
0.73
0.81
3.24
13.24
0.60
0.95
3.10
13.28
0.61
1.00
3.27
13.25
0.59
0.99
3.68
13.30
0.55
0.84
3.36
13.35
0.60
0.97
3.84
13.28
0.62
0.70
3.33
13.17
0.64
0.70
3.20
13.24
0.80
0.25
4.04
13.12
0.59
0.96
3.18

Table 2. Estimated parameters: nugget (c0), sill (c), range (a), and nugget to sill ratio
(c0/c), for three semivariogram models fitted to each vertical plane inside the cold
store. Model selection criteria: Akaike information criterion (AIC) and Bayesian
information criterion (BIC) are included as well.
Model

Cauchy

Gaussian

Exponential

Plane
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10

Nugget
(c0)
0.24
0.29
0.27
0.28
0.27
0.24
0.03
0.17
0.25
0.26
0.25
0.24
0.28
0.24
0.27
0.25
0.19
0.19
0.29
0.27
0.14
0.24
0.23
0.25
0.26
0.18
0
0.01
0.04
0.23

Range
(a, cm)
1852.16
3718.42
3742.35
3718.42
3742.35
935.64
408.21
992.82
913.59
3742.35
168.98
92.37
324.27
86.11
13.04
111.56
97.28
111.53
103.99
324.27
561.26
557.67
561.26
557.67
561.26
151.84
105.30
153.46
161.92
561.26

Sill
(c)
0.98
0.39
0.46
0.31
0.16
0.21
0.48
0.55
0.84
0.47
0.54
0.13
0.20
0.12
0.03
0.14
0.24
0.33
0.49
0.21
0.93
0.25
0.27
0.18
0.08
0.22
0.43
0.52
0.81
0.25

AIC

BIC

81.27
63.21
68.69
60.66
66.97
71.74
71.42
76.31
93.21
69.48
82.05
64.72
69.04
62.16
68.30
71.59
70.68
75.67
92.9
69.71
81.13
63.62
69.65
61.18
67.37
71.79
71.20
76.25
93.16
70.47

87.93
69.19
75.13
66.52
73.52
78.18
77.75
82.86
99.76
76.03
88.70
70.71
75.48
68.02
74.85
78.03
77.01
82.22
99.45
76.26
87.78
69.61
76.10
67.04
73.92
78.24
77.54
82.80
99.71
77.02

c0/c
(%)
19.72
42.38
37.28
48.02
62.95
53.90
5.44
23.71
23.25
35.69
31.29
64.67
58.11
67.18
89.14
64.65
44.32
36.15
37.69
55.44
12.98
49.64
45.77
57.93
75.32
44.15
0
2.70
4.80
48.14

35

Figures
A.

B.

Fig. 1. Temperatures measured at each plane (P1 to P10): A) Time series at each plane in
a full day, B) Boxplot for temperature at each plane.

Fig. 2. Fitted semivariograms following an exponential model for the temperature

distribution on each plane (P1 to P10) of the experimental cold store.

36

Fig. 3. Temperature (C) distribution maps predicted by block ordinary kriging technique,
for each plane (P1 to P10) of the experimental cold store.

37

38