Professional Documents
Culture Documents
Research article
a Federal University of Viçosa/UFV, Av. Peter Henry Rolfs; s/n, 36570-000, Viçosa, MG, Brazil
b Federal University of Espírito Santo/UFES, Center of Agrarian Sciences and Engineering, Alto Universitário; s/n, 29500-000, Alegre, ES, Brazil
Keywords: The foremost objective of this study was to analyze the performance of a Markov chain/cellular automata model for predicting
Prognosis land use/land cover changes in environments predisposed to desertification. The study area is the Vieira river basin, located in
Landscape Montes Claros (MG, Brazil). Land use/land cover prognosis was performed for the year 2005 so that this result could be
Cellular automata compared with the ranked image for the same year, taken as ground truth. Kappa indices were used to evaluate the change
Markov chain
level that occurred between these two cases. Results from cellular automata were evaluated from those of the Markov chain
Semiarid
model. The latter proved to be efficient in the quantitative prediction of changes in land use/land cover. Regarding the cellular
Degradation
automata, an average per-formance was noted in the spatial distribution of classes. Specifically, with regard to desertification,
the use of the CA-Markov model was effective at estimating the total area of the most susceptible class to this process, Bare
Soil; however, it was inefficient in its spatialization. Even with the caveats related to the performance of cellular automata, the
overall prediction capacity of CA-Markov models can be considered as good.
1. Introduction world's population lives in these areas (Matallo Jr, 2001). Despite these
alarming numbers, there is a surprising lack of consensus on the ap-propriate
Desertification is a process of land degradation, limited to dry re-gions. Its form of evaluating desertification (Verón et al., 2006).
possible incidence is restricted to arid, semiarid, or dry sub-humid areas. Owing to the intrinsically spatial nature, studies on desertification have
Among several factors considered as causes of desertifi-cation, climate benefited directly from recent advancements in the areas of re-mote sensing
change and anthropic activities are highlighted (Brasil, 1998). Recognized as and geographical information systems (GIS). Using these techniques, the CA-
one of the main causes of desertification, several human practices have Markov model combines the Markov chain and cellular automata (CA),
exceeded the productive land capacity, char-acterized by inappropriate soil where the probabilities of possible land use changes supplied by a Markov
management, including grazing, ero-sion, salinization, continuous use of fire, chain are spatialized by CA. Several authors studied desertification employing
extractivism, mining, and deforestation (Adamo and Crews-Meyer, 2006; the CA-Markov model and considered this model as viable and efficient in
D'Angelo et al., 2000; Huang et al., 2009; Mabbutt, 1992; McClure, 1998; predicting trends con-cerning the behavior of this land degradation process
Rhodes, 1991; Santini et al., 2010). (Huo-ping et al., 2009; Lu et al., 2009; Muller and Middleton, 1994; Wu et
al., 2010). Recent studies highlight that the association between Markov chain
Desertification affects over 100 countries and is present on all continents, and cellular automata makes it possible to predict the probability of trends in
especially in arid and semiarid regions that occupy ap-proximately one-third land use evolution, such as the phenomenon of desertification (Jiang and Lin,
of the land surface and are subject to the oc-currence of this process. It is 2018; Joseph et al., 2018; Varghese and Singh, 2016; Xu et al.,
worth mentioning that one-sixth of the
∗
Corresponding author. Federal University of Espírito Santo/UFES, Center of Agrarian Sciences and Engineering, Alto Universitário; s/n, 29500-000, Alegre, ES. Brazil.
E-mail addresses: kellyobarros@yahoo.com.br (K. de Oliveira Barros), cribeiro@ufv.br (C.A. Alvares Soares Ribeiro), gustavomarcatti@gmail.com (G.E. Marcatti),
alexandre.lorenzon@ufv.br (A.S. Lorenzon), nerolemos@yahoo.com.br (N.L. Martins de Castro), getulio.floresta@gmail.com (G.F. Domingues), jromario_carvalho@hotmail.com (J.
Romário de Carvalho), alexandre.santos@pq.cnpq.br (A. Rosa dos Santos).
https://doi.org/10.1016/j.jenvman.2018.07.064
Received 13 April 2018; Received in revised form 20 June 2018; Accepted 18 July 2018
0301-4797/ © 2018 Elsevier Ltd. All rights reserved.
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
161
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
0.81–1.00 is considered excellent. contiguity filter that applies a spatial weighting factor to each class, so that the
After ranking and evaluation of the accuracy, images were used as input spatial dependence between data is considered properly. This filter will
data for the projection of land use/land cover and, subsequently, to associate indicate higher or lower aptitude of the territory for each soil class in the
the behavior of each class with susceptibility to desertifi-cation. This stage study. The filter used was the 5 × 5 pattern of the CA-Markov model.
®
was performed in the Idrisi software.
The Kappa indices were created using ranked images of 1995 and 2000.
®
We used the CrossTab tool in ArcGIS , which supplies coefficients that
2.4. Markov chain and cellular automata indicate the land use change level between two dates.
After obtaining the Markov chain results, the CA-Markov model was
To employ the Markov chain, a pair of images with properly ranked dates applied. Faced with the absence of more significant information on the forces
is necessary. The year to which the scenery will be projected corresponds to driving the evolution of land use/land cover, an isotropic cost surface was
the period between the two images in the study added to the corresponding adopted, assigning a value of 1 to all cells of the Vieira river basin. The
year to the second image. Images of years 1995 and 2000 were selected to subjectivity and complexity naturally associated to the se-lection process of
apply the Markov chain. Since the year of pro-jection for this combination is factors and assignment of weights to describe their relative importance were
2005, it is possible to evaluate the ob-tained result by comparing it with the removed. The decisive rule of land use/land cover was based only on the
ranked image of 2005, taken as ground truth. transition probability produced by the Markov chain, considering uniform
properties irrespective of direction or land use.
Given that the Markov chain presents only the transition prob-abilities
among the different soil use classes, geographical reference was added to the
process using cellular automata (CA). The CA has a
162
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
2.5. CA-Markov model validation Area changing into Cropland/Pasture was 10%. In the ranking, it was possible
to notice a higher confusion of the rater regarding some shadow pixels or
To validate the result obtained in the CA-Markov process, cross even Bare Soil, which in the RGB composition pre-sented a gray shade,
®
tabulation was performed in ArcGIS software. Since this procedure indicates probably with similar spectral behavior to that of the pixels indicated by the
the coincidence between the classes of two images, it was possible to evaluate sample as Urban Area. Thus, this value can be justified by the confusion that
the result obtained by the Markov chain by cal-culating the Kappa index existed in the ranking.
produced between the two dates. The image obtained in this process in the For Bare Soil, it was observed that the value indicated as probability for
combination of years 1995 and 2000, in other words, the projected image for this class continuing unchanged is quite close to that indicated as this class
2005, and the image ranked with properly standardized numbers of each class changing into Cropland/Pasture. Therefore, it was possible to confirm the
with a date of 2005, was used. This procedure was repeated with the projected hypothesis that part of the areas identified as Bare Soil were simply
image of 2005 and the ranked image of 2000. momentarily without any agriculture and the identified Bare Soil is not
necessarily degraded soil; in other words, it could not be used for agriculture
To validate the result obtained by CA, a class-by-class comparison was or rangeland or even that vegetation could not re-generate. The change
performed between the ranked images of 2005 and the image ob-tained by the probability from Bare Soil to Forest was very small when compared with the
CA-Markov model, verifying the hits of the CA-Markov image for each case. other three possibilities, that is, only 5%. The Bare Soil and Forest classes can
It was possible to evaluate the CA performance in the spatial reference be considered opposite si-tuations when associated with degradation. This
determination for each class. low value of Bare Soil changing into Forest indicates the precarious capacity
of recovery of the area without anthropic intervention, and consequently, a
higher prob-ability of these areas, which are, in fact, Bare Soil, remaining as
2.6. Susceptibility to desertification relation with land use/land cover
such; in other words, a higher inclination for desertification. In the case of the
classes
transition possibility between Bare Soil and Urban Area, the value of 20%
was observed, which can be justified by the higher use of Bare Soil areas
A susceptibility relation was established between the defined classes and
close to the Urban Area.
occurrence of desertification in which the Bare Soil class was in-dicated as
that presenting highest susceptibility to this process, fol-lowed by the
Among the factors that may be acting directly in the transition from
Cropland/Pasture class, and then Forest and Urban Area classes.
pasture to degraded soils, we can highlight the trampling of the animals and
the lack of adequate pasture management, since in Brazil the an-imal
production system is extensive with free animals in the pasture (Vzzotto et al.,
3. Results and discussion 2000). These factors contribute to changes in soil phy-sical properties that
will lead to problems of penetration of the plant root system and infiltration of
3.1. Image ranking water into the soil, which can cause surface runoff and soil surface loss of
organic matter (Tollner et al., 1990; Vzzotto et al., 2000). This can lead to a
It was possible to notice in the three dates in this study that Forest is the shortage of nutrients in the soil and, consequently, affect the physiology of
most representative class in the Vieira river basin, followed by the plant, leading to death and causing bare soil (Tollner et al., 1990; Torres et
Cropland/Pasture, Urban Area, and lastly by Bare Soil (Table 1 and Fig. 3). al., 2012).
The Kappa index was calculated, and the obtained value cor-responds to 0.89, The Forest class presented the highest probability of remaining unchanged
0.88, and 0.91 for years 1995, 2000, and 2005. Ac-cording to Landis and when compared with the other possible changes among classes. The highest
Koch (1977), such values reveal an excellent quality ranking, with great revealed tendency was for vegetation changing to Cropland/Pasture followed
coincidence between the ground truth and the ranked image. by Urban Area and Bare Soil, respectively. Such a fact reveals that not only
were humans a transforming agent of the landscape but also anthropic
pressure is, in great part, responsible for the suppression of land cover, which
3.2. Markov chain has intensified over the years mainly because of increased agriculture.
Transition probabilities among land use classes for 2005 are in listed in When comparing the chances of Urban Area and Forest classes suffering
Table 2. alterations, it is considered that Urban Area presents a higher probability for
It was noted that all classes presented a higher probability of re-maining smaller alterations, given that the land cover can suffer modifications faster
unchanged. When comparing the probability of the Urban Area class altering due to climatic events, fires, or any human actions. Despite Urban Area
to other classes, it was possible to notice that there was a higher trend of this having presented a high value of remaining un-changed in every year of
class turning into Forest (20%). City tree planting or even urban horizontal analysis, it was not superior to the value presented by Forest at any point. The
expansion of the municipality, with frequent land covered backyards, can also presented values, in these two cases, can possibly be explained by the
justify the probability found for this transformation between Urban Area and characteristics of these two classes in the images not ranked yet. In the case of
Forest being the second most representative (França and Soares, 2007). The Forest, it is possible to consider this class with lower degree of variability
probability of Urban among its pixel values. In contrast, in the case of Urban Area, the pixels
present con-siderable variation in values in its corresponding area. This pixel
Table 1 value variation is associated with variability itself of what composes an urban
Area of each land use class for dates of Sept. 9, 1995; Sept. 22, 2000, and Jul. area: construction, asphalt, vegetation, water, and bare soil. The pos-sibility of
18, 2005, in the Vieira River basin, Montes Claros (MG, Brazil). transition from Forest to Urban Area was only not lower than the change
Classes Dates
probability from Forest to Bare Soil. These low values are also justified by the
low occurrence of Forest in the surroundings of Urban Area, where Bare Soil
09/09/1995 09/22/2000 07/18/2005 and Cropland/Pasture classes prevail around the urban spot. Values of
2
possible transformations from the Cropland/Pasture class to Urban Area and
Area (km )
from Bare Soil to Urban Area were superior when the Forest probability was
Urban Area 58 65 80 changed to Urban Area. When compared with Cropland/Pasture, Bare Soil
Forest 280 308 295 had a higher prob-ability of changing into Urban Area, and such values are
Cropland/Pasture 194 155 146 justified by the
Bare Soil 48 52 59
163
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
Fig. 3. Supervised ranking of the Vieira river basin, Montes Claros (MG, Brazil) for dates: September 9, 1995 (A); September 22, 2000 (B); and July 18, 2005 (C).
Table 2 Comparing the values produced by the Markov chain projection and by
Transition probabilities (%) among land use classes for year 2005 in Vieira Rver basin, the ranking (Table 3), a higher proximity of Cropland/Pasture and Bare Soil
Montes Claros (MG, Brazil), based in years 1995 and 2000. 2
class values was observed, with a difference of 6 km for the first class and 8
2
2000
km for the second class. In the case of the Forest class, this difference
2
between two areas was higher at 26 km . Among classes, Urban Area
2
Urban Area Forest Cropland/Pasture Bare Soil
presented an intermediary difference of 12 km . However, despite these
differences, the Kappa between these two cases was 0.45, which in the
ranking of Landis and Koch (1977) indicates good quality. This Kappa
Projection for 2005 ranking confirms the good prediction performed by the Markov chain
1995 concerning probabilities of each land use/land cover class. Owing to the slight
difference between the result presented by the Markov chain and the ranking,
it is possible to confirm that change probabilities in land use were not, in fact,
Urban Area 61 20 10 8 constant along the years. Nevertheless, even when Markov chain determined
Forest 2 83 11 1 the stationarity of probabilities and this did not occur, the small difference
Cropland/Pasture 5 30 52 11 between its prediction and the ranking provides this part of the CA-Markov
model with good performance.
Bare Soil 20 5 32 41
When comparing Fig. 3C, from 2005 taken as ground truth, with Fig. 4,
the result of the projection for the same year, it can be noticed that the rater
tendency of reduction of voids existing within and around city limits, making confusion regarding some pixels of Bare Soil or shadow, classifying them as
this area more densely occupied. This was observed visually on the classified Urban Area, was reduced, although not removed completely. This observation
images when compared with the most recent classification from 2005 (Fig. is associated with the role of the CA of the model that, due to the tendency to
3C). expand border spots, irrespective of class, ended up reducing this rater
The Forest class presented a lower change probability to Bare Soil. It is confusion since it was a matter of small groups or even individual pixels. This
necessary to consider that in this ranking, the type of land cover was not behavior of the CA is si-milar to the natural behavior of the environment, in
specified, and as observed in the land cover mapping prepared by the IEF and which there is a higher change probability in transition areas of classes than
UFLA, deciduous forest occurs in the Vieira river basin, in other words, forest when compared with change in the interior of a spot of a determined class
in which trees lose most of their leaves in periods of drought (Carvalho and distant from the border. Hence, in the visible comparison between the image
Felfili, 2011), providing less land cover to the soil during this period of produced by CA (Fig. 4) and the ranking image of 2005 (Fig. 3C), Bare Soil
climatic severity. Since a single class was determined, irrespective of areas, in parts already existing in the image of 2000 (Fig. 3B), became larger
vegetation type and its characteristics, the demonstrated probability is instead of being distributed in smaller areas, but more scattered in the Vieira
considered acceptable, which points to a great deal of soil protection by the river basin, as noticed in the ranking image of 2005 (Fig. 3C).
vegetation, regardless of specificity.
The second highest probability after Forest not changing in any other
class was the transition to Cropland/Pasture. Despite values showing the When the image produced by CA (Fig. 4) was compared with the ranked
pressure that these activities apply on land cover, the probability value of the image of 2000 (Fig. 3B), it was possible to notice higher simi-larity between
Cropland/Pasture class changing into Forest was higher than the opposite these images than between CA images (Fig. 4) and the 2005 ranking (Fig.
situation. This situation shows a higher probability that cropland and 3C). Thus, the Kappa was produced between the ranked image of 2000 and
rangeland areas develop vegetation, probably due to issues around areas being the projection for 2005. The obtained Kappa was 0.50, which is higher than
deserted than to the reverse situation, which would correspond to pressure on the Kappa produced between the CA result and the 2005 ranking (0.45).
this last class. These values show that, in fact, there is a greater similarity between the CA
result and the 2000 ranking than with the 2005 ranking. This fact can be
justified because the year
3.3. CA and CA-Markov model validation
Table 3 demonstrates values obtained for each class by the Markov chain
with projection for 2005 and for ranking in the same year. Fig. 4 shows
spatial distribution of land use/land cover classes of the Vieira river basin for
2005 by the projection performed by the CA-Markov model.
164
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
Fig. 4. Projection of land use/land cover classes for 2005 for the Vieira river basin, Montes Claros (MG, Brazil).
Table 3 3.4. Susceptibility to desertification and land use/land cover classes in the
Areas of land use/land cover classes obtained by Markov chain projected for year 2005 Vieira river basin
and for supervised image ranking of the same year.
Urban Area 80 45 56
Forest 295 246 83
Cropland/Pasture 146 71 49
Bare Soil 59 18 30
165
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
classes was involved in this process. Although for the Vieira river basin Applied to Global Environment (GAGEN), to the National Council for
urbanization presented only a small change in vegetation areas, this value Scientific and Technological Development (CNPq) and the Postgraduate
must not be neglected. Program in Forest Sciences of the Federal University of Viçosa for providing
the necessary resources for implementation and execution of this work.
3.5. Evaluation of analytical tools
166
K. de Oliveira Barros et al. Journal of Environmental Management 225 (2018) 160–167
167