Journal of Earth, Environment and Health Sciences

ORIGINAL ARTICLE
Year
: 2015  |  Volume : 1  |  Issue : 1  |  Page : 16--21

Modeling Land Use/Cover Changes by the Combination of Markov Chain and Cellular Automata Markov (CA-Markov) Models


Mozhgan Ahmadi Nadoushan1, Alireza Soffianian2, Alireza Alebrahim3,  
1 Department of Agriculture, Islamic Azad University, Khorasgan, Esfahan, Iran
2 Department of Natural Resources, Isfahan University of Technology, Isfahan, Iran
3 Department of Agriculture, Payame Noor University, Tehran, Iran

Correspondence Address:
Mozhgan Ahmadi Nadoushan
Department of Agriculture, Islamic Azad University, Khorasgan, Isfahan
Iran

Abstract

Land use/cover changes modeling is essential for land use planning and management. Arak is one of several cities in Iran which have undergone swift urban expansion during recent decades due to rapid industrialization and population growth. In this study, aerial photos and Landsat TM and IRS-P6 LISS-III images were used to predict land use/cover changes in Arak. Land use/cover maps were generated with four classes from visual interpretation of aerial photos and an artificial neural network classification method for satellite images. Both classification methods resulted in land use/cover maps with overall accuracy over 95 %. In order to predict changes, Markov chain and Cellular Automata Markov models were applied and a land use/cover map for 2025 was simulated. The results showed that the combination of satellite remote sensing, GIS and Markov models provides useful information on land use/cover dynamics in future which could be consequently used for land use planning.



How to cite this article:
Nadoushan MA, Soffianian A, Alebrahim A. Modeling Land Use/Cover Changes by the Combination of Markov Chain and Cellular Automata Markov (CA-Markov) Models.J Earth Environ Health Sci 2015;1:16-21


How to cite this URL:
Nadoushan MA, Soffianian A, Alebrahim A. Modeling Land Use/Cover Changes by the Combination of Markov Chain and Cellular Automata Markov (CA-Markov) Models. J Earth Environ Health Sci [serial online] 2015 [cited 2024 Mar 28 ];1:16-21
Available from: https://www.ijeehs.org/text.asp?2015/1/1/16/159922


Full Text

 Introduction



Land use/cover changes play a major role in the study of global change. Land use/cover and human and natural modifications have largely resulted in deforestation, biodiversity loss, global warming and increase of disaster. [1] Investigating the landscape structure and its change is a prerequisite to the study of ecosystem functions and processes, sustainable resource management and effective land use planning. [2] Land-cover and land-use change analysis and projection provide tools to assess ecosystem change and its environmental implications at various temporal and spatial scales. [3] Satellite remote sensing in conjunction with Geographic Information Systems (GIS) has been recognized as a powerful and effective tool in detecting land use and land cover changes. [4],[5] Satellite remote sensing is a potentially powerful means of monitoring land-use change at high temporal resolution and lower costs than those associated with the use of traditional methods and provides multi-spectral and multi-temporal data that can be used to quantify the type, amount and location of land use and land cover changes. [6]

In order to use Markov chain models it is necessary to accept several underlying assumptions. One basic assumption is to regard land use and land cover change as a stochastic process, and different categories as the states of a chain. Moreover, it is convenient to regard the change process as one which is discrete in time. [4]

Markov and CA-Markov models were processed for each pair of dates and transition probability and transition area matrices were produced. The transition probability matrix determines the likelihood of change in each pixel from one land use or land cover class to each other category between the two dates. The Markovian model outputs an image which reports the probability that each land use or land cover class would be found at each location in the next step. [7]

Markov Chain Analysis describes land use change from one period to another and uses this as the basis to project future changes. This is accomplished by developing a transition probability matrix of land use change from time one to time two, which will be the basis for projecting to a later time period.

One inherent problem with the Markov Analysis is that it provides no sense of geography. The transition probabilities may be accurate on a per category basis, but there is no knowledge of the spatial distribution of occurrences within each land use category; there is no spatial component in the modeling outcome. [8]

Wu et al., (2006) used satellite remote sensing and GIS to monitor and predict land use change in Beijing, China. A Maximum likelihood classifier was applied for land use classification and the overlay method was used for detecting land use change during a fifteen-year period stretching from 1986 to 2001. Finally, a projection of land use change was prepared for 2021 using Markov chains and regression analyses. Results showed that this integration of remote sensing and GIS technologies with Markov and regression models was useful for analyzing and predicting the process of land use change. [9]

Fan et al., (2008) explored land use and land cover changes in the Core corridor of the Pearl River Delta (China) from 1998 to 2003 using TM and ETM + images and the post-classification method. They also used Markov chain and Cellular Automata models to predict urban expansion in 2008 and 2013 and concluded that Markov chain modeling is an effective way to monitor and predict land use change and urban expansion. [10]

Kaveh and Ebrahimi (2013) used CA Markov model in detection and simulating land use/cover change over 65 km of Aghbolagh river. Land use/cover map was prepared using 1956, 1969, 1998 aerial photos and 2006 satellite images. Then river, riparian area, agriculture and bare lands around it were illustrated and land use/cover maps of each year were drowned and overlaid through cross routine to evaluate model validation. Finally, the conditions of these classes were forecasted using CA-Markov model for 2016. Results of forecasting future changes based on 1969 and 1998 maps showed that in 2016, arable lands will have the most stability while river is the most vulnerable land cover to change. [11]

Nouri et al., (2014) used CA-Markov model for analysis of temporal changes and spatial distribution of urban land uses in Anzali located in Gilan province in the northwest of Iran. Changes and spatial distribution of land uses in the town were calculated using geographic information systems technology for a time span 1989-2011. In the next step the spatial distribution of urban land uses in 2021 was simulated using the transition matrix. [12]

The main objective of this study is to predict land use/cover changes of the city of Arak and its periphery for 2025 using land use/cover maps derived from aerial photographs and satellite images and Markov chain and CA-Markov model.

 Materials and Methods



Study Area

The study area is located between latitudes 34° 02' 40"-34 °08' 01" N and longitudes 49° 37' 31"-49 °47' 45" E in the center of Iran and has an area of about 15800 hectares [Figure 1]. This region has an elevation of about 1800 m above sea level and contains Arak, the capital city of Markazi Province, and its surrounding area. The average annual temperature of Arak is 13.8° C and average annual rainfall is 316 mm. Arak has experienced rapid expansion due to population growth and industrialization during recent decades. The population of Arak increased from just under 59000 in 1956 to 446760 in 2006.{Figure 1}

Data and data preprocessing

Remotely sensed data were used as the primary data source to generate input for change prediction. A time series of remote sensing data including aerial photos and satellite images was used to generate land use/cover maps.

Aerial photographs from 1972 at a scale of 1:10000 were employed for producing land use/cover maps. Landsat TM and IRS-P6 LISS III images (path 68, row 46) for the years 1990 and 2006 were also used in this study. A Digital Elevation Model (DEM) was produced from the digital topographic maps at a scale of 1:25.000. Finally, an IRS-1C PAN image was used to enhance the spatial resolution of IRS-P6 LISS-III image. To prepare data for mapping land use/cover, following procedures were performed.

Geometric correction

All images and aerial photographs were rectified to UTM zone 39 N with at least 25 well-distributed ground control points. At first geometric correction was carried out using topographic maps at a scale of 1:25000 to geocode the aerial photos and the 2006 IRS-1C PAN image. This rectified image was then employed to register the 2006 LISS-III image. Geometric correction of Landsat TM image of 1990 was carried out using the IRS-P6 LISS-III image. Finally, a first-order polynomial model was applied and all data were resampled to a 30 m pixel size using the nearest neighbor method. After geometric correction of aerial photos, all photos for each year were mosaicked to prepare one image for land use/cover mapping.

Topographic correction

A topographic correction was applied to all satellite images due to the mountainous condition of the study area. The solar azimuth and elevation were read from the satellite images' metadata files. Terrain attributes including slope and aspect were derived from the Digital Elevation Model (DEM). Topographic correction was carried out on images based on the Lambertian method.

Image enhancement

In this study, two false color composites (FCC) were produced for selecting training samples. Image fusion was also performed to increase spatial resolution of the LISS-III image. The LISS-III image was fused with IRS-1C PAN image to generate an image with a high spatial resolution.

Land use/cover mapping

For generating land use/cover maps from aerial photos and satellite images, four land use/cover classes were identified based on field work, false color composite images, and images derived from the fusion process. The area was classified into these main classes: Urban areas, vegetation cover, bare lands and rocky and mountainous areas. For the current study, urban areas include residential, commercial, industrial, educational, recreational establishments and transportation systems and vegetation cover encompasses all green spaces, farmlands and natural vegetation.

Aerial photos interpretation

The land use/cover pattern was interpreted visually on black and white aerial photographs and simultaneously digitized with the Arcmap software. Identifying features in aerial photos were based on tone, texture, pattern, size and shape.

Image classification and accuracy assessment

All land use/cover maps were generated through an artificial neural network classifier. In order to get precise classification results, the training samples were selected from FCC images and topographic maps. To prepare land use/cover maps, three-layer-perceptron neural networks were employed consisting of one input layer, one hidden layer and one output layer. The input layer included spectral bands and training samples and the output layer had four nodes. Experiments were conducted to select the optimum number of nodes in the hidden layer to maximize classification accuracy. The number of nodes in the hidden layer was selected equivalent to the number of nodes in the input layer. The parameters of momentum and learning rate were adjusted to 0.5 and 0.2 based on experimental results.

The overall accuracy of land use/cover maps was calculated from error matrices. The Ground truth data were derived from GPS, topographic maps and false color composite images and an error matrix was generated for each land cover map.

Land Use/cover change prediction

Cellular Automata (CA) are used to add a spatial character to the model. Using the outputs from the Markov Chain Analysis, specifically the transition area matrices, a CA-Markov model will apply a contiguity filter to model land use change from time two to a later time period. [13]

To predict land use/cover changes in the study area, Markov chain analysis and CA-Markov model were employed. Transition probability matrices were derived from Markov chain analysis and predicted land use/cover maps were generated through CA-Markov model. Land use/cover maps from 1972 and 1990 were used for model calibration and then the models used to predict a land use/cover map for 2006. Finally, a land use/cover map for 2025 was simulated using models based on the land use/cover maps of 1990 and 2006. Real maps for 1990 and 2006 have been produced through neural network classification of satellite images, to be used as reference maps in the model validation process.

Model validation

Three methods were used to assess the overall performance of the model in predicting land cover changes:

Examine the goodness-of-fit of validation

Validation is the process of comparing the predicted map to a reference map to evaluate the performance of the model. [13] We used the VALIDATE module of the IDRISI software. This module examines the components of agreement and disagreement between the comparison and reference maps for the same date.

Real land cover maps derived from neural network classification were regarded as the reference maps and the comparison maps were the result of simulation. The validity of the predicted map was assessed using the reference map. The VALIDATE module also computes the Kappa index of agreement and its variants, each of which measures different characteristics of agreement: Kno gives the overall accuracy of a simulation run; Klocation indicates the level of agreement of location, given a specified quantity; and Kquantity indicates the level of agreement of quantity, given the model's ability to specify location. These variations complement the standard Kappa, which equals 1 when agreement is perfect and equals 0 when agreement is as expected by chance. [14]

Error matrices

Assessment of the model's predictive power also was performed by calculating the overall accuracy of predicted maps. For this purpose, Ground truth data obtained for classification accuracy assessment were used and the overall accuracy of simulated land cover maps was calculated.

 Results and Discussion



The root mean square errors (RMSE) for all aerial photographs were between 0.2 and 0.7 pixels. RMSE for PAN, LISS III and TM images were determined to be 0.42, 0.48 and 0.58 pixels, respectively.

Image fusion was carried out and resulted in an image with high spatial resolution of 5.8 meters. The fused image enhanced the capability of identifying features and selecting training samples.

Land use/cover mapping was done through visual interpretation of aerial photos and neural network classification of satellite images. Both methods generated land use/cover maps with high accuracies. The overall accuracy of the land use/cover maps for 1972, 1990 and 2006 were 99.05 %, 95.53 % and 95.53 %, respectively. The Kappa index for the 1972, 1990 and 2006 land use/cover maps were found to be 0.98, 0.92 and 0.93.

Change prediction

Real and predicted land use/cover map for 2006 is shown in [Figure 2] and simulated land use/cover map for 2025 is shown in [Figure 3]. Transition probability matrices were derived from Markov chain analysis of land use/cover maps for different time periods. According to the transition matrices of 1972-1990 and 1990-2006, the most stable classes were mountain and urban areas and the most unstable class for both periods was vegetation cover [Table 1] and [Table 2].{Figure 2}{Figure 3}{Table 1}{Table 2}

Transition probability of mountains, urban areas and vegetation cover from 1972 to 1990 were 1, 0.61 and 0.26 and for 1990-2006 were found to be 1, 0.77 and 0.39 as presented in [Table 2] and [Table 3] respectively. The low value of transition probability for bare lands mainly indicate loss of this land cover class in favor of urban areas and the low value of transition probability for vegetation cover show the conversion of these lands to urban areas and bare lands.

These results of Markov matrices represented mountain as the most constant class. This seems reasonable because converting mountainous areas to other land use/cover types is not a common and easy process except for some special conditions.{Table 3}

Model validation

Examine the goodness-of-fit of validation

The overall agreement is budgeted by three components: Agreement due to chance, agreement due to the predicted quantity of each land category and agreement due to the predicted location of each land category. The overall disagreement is budgeted by two components: Disagreement due to the predicted location of each land category and disagreement due to the predicted quantity of each land category. The interpretation of agreement is the proportion of pixels classified correctly, and the interpretation of disagreement is the proportion of pixels classified incorrectly. [15]

As shown in [Table 3], Klocation is estimated to be 0.71 and K quantity is 0.88 for predicted land use/cover map of 2006. Also the results of disagreement components indicate that the disagreement due to location is greater than the disagreement due to quantity [Figure 4]. From these values it is clear that the main part of error is associated with the simulation of location.{Figure 4}

These facts show that in our example, CA-Markov performed better at predicting the quantity of pixels rather than the location of pixels.

Error matrices

Another method used to assess the capability of the models in predicting land cover change is to calculate the accuracy of simulated land cover map based on the ground truth data and to compare it with the accuracy of the real map. The overall accuracy of real and predicted land cover maps for 2006 are 95.5 and 81.5 % as it is shown in [Table 4] and [Table 5]. As is obvious, the overall accuracy of simulated land cover is not high and the model was not very successful in predicting land cover changes of 1990.{Table 4}{Table 5}

In general, it is better to use Markov models to predict land cover changes with two main classes such as built-up/non-built-up, forest/non-forest, etc. Pontius and Malanson (2005) noted that it might be advisable to simplify the base data in order to focus on the major signal of change, which frequently is the conversion from a single non-disturbed category to a single disturbed category. [13]

Another factor which might affect the result of modeling was time intervals. For stochastic models, short intervals result in better prediction accuracy. If more data points were used in the present study, the results would be more accurate. In this study, time intervals were quite long because of inaccessibility of images for a number of years.

 Conclusion



In the present study, Markov and CA- Markov models were helpful for predicting land use/cover changes in 1990 and 2006. The outcomes of this research indicate that Landsat TM and IRS-P6 LISS-III images can be effectively used for generating accurate land use/cover maps as the overall accuracies of all generated land cover map were found to be over 95 %.

According to the results of the CA-Markov model, urban expansion will occur in the future. The combination of satellite remote sensing, GIS and Markov models provides useful information on land use/cover dynamics and change trends into the future which could help policy makers to make better decision for the future of the study area. The provided future projection could be effectively used for land use planning, decision making and land management, especially if its use is confined more to general trends than to specific land-use locations, where accuracy was lower. The predictive power of the CA-Markov model, especially in predicting the location of pixels, was not very high in this research but in general Markov models have shown the capability for the prediction of land cover/land use change trends.

References

1Reis S. Analyzing land use/land cover changes using remote sensing and GIS in Rize, North-East Turkey. Sensors 2008;8:6188-202.
2Matsushita B, Xu M, Fukushima T. Characterizing the changes in landscape structure in the Lake Kasumigaura Basin, Japan using a high-quality GIS dataset. Landsc Urban Plan 2006;78:241-50.
3Lopez E, Bacco G, Mendoza M, Duhau E. Predicting land-cover and land-use change in the urban fringe: A case in Morelia city, Mexico. Landsc Urban Plan 2001;55:271-85.
4Weng Q. Land use change analysis in the Zhujiang Delta of China using satellite remote sensing, GIS and stochastic modelling. J Environ Manage 2002;64:273-84.
5Dewan AM, Yamaguchi Y. Using remote sensing and GIS to detect and monitor land use and land cover change in Dhaka Metropolitan of Bangladesh during 1960-2005. Environ Monit Assess 2009;150:237-49.
6Hegazy IR, Kaloop MR. Monitoring urban growth and land use change detection with GIS and remote sensing techniques in daqahlia governorate Egypt. Int J Sustain Built Environ 2015.
7Houet T, Hubert-Moy L. Modelling and projecting land-use and land-cover changes with a Cellular Automaton in considering landscape trajectories: An improvement for simulation of plausible future states. Earsel Proc 2006;5:63-76.
8Pontius RG, Malanson J. Comparison of the structure and accuracy of two land change models. Int J Geogr Inf Sci 2005;19:243-65.
9Wu Q, Li H, Wang R, Paulussen J, He Y, Wang M, et al. Monitoring and predicting land use change in Beijing using remote sensing and GIS. Landsc Urban Plan 2006;78:322-33.
10Fan F, Wang Y, Wang Z. Temporal and spatial change detecting (1998-2003) and predicting of land use and land cover in core corridor of Pearl River Delta (China) by using TM and ETM + images. Environ Monit Assess 2008;137:127-47.
11Kaveh N, Ebrahimi AA. Markov chain model for simulating land use/cover change (Case study: Aghbolagh river). J Appl RS GIS Tech Nat Resour Sci 2013;4:41-51.
12Nouri J, Gharagozlu A, Reza A, Shahrzad F, Mahsa A. Predicting urban land use changes using a ca-markov model. Arabian J Sci Eng 2014;39:5565-73.
13Eastman RJ. IDRISI Kilimanjaro Tutorial. Clark Labs. Worcester: Clark University; 2003.
14Pontius RG, Claessens L, Hopkinson C, Marzouk A, Rastetter E, Schneider L, et al. Scenarios of land-use change and nitrogen release in the Ipswich watershed, Massachusetts, USA. 4 th International Conference on Integrating GIS and Environmental Modelling, Alberta, Canada, 2000.
15Pontius RG, Spencer J. Uncertainty in extrapolations of predictive land-change models. Environ Plan 2005;32:211-30.