Using RS Data-Based CA–Markov Model for Dynamic Simulation of Historical and Future LUCC in Vientiane, Laos

Land use/cover change (LUCC) is one of the causes of global climate and environmental change. Understanding rapid LUCC in urbanized areas is vital for natural resources management for sustainable development. This study primarily considered Vientiane, the capital of Laos, which experienced rapid LUCC due to both natural and anthropogenic factors. The study used geographical information system (GIS) combined with ERDAS and TerrSet technologies to objectively process the ground surveyed and remotely obtained data in order to investigate the historical LUCC as well as predict future LUCC in the study area during the periods of 1995–2018 and 2030–2050, respectively. A comprehensive list of assessment factors comprised of both natural and anthropogenic factors was used for analysis using the cellular automata–Markov (CA–Markov) model. The results show a historical loss of intact forest of 24.36% and of bare land of 1.01%. There were also tremendous increases in degraded forest (11.36%), agricultural land (8.91%), built-up areas (4.49%) and water bodies (1.16%). Finally, the LUCC prediction results indicate the conversion of land use from one type to another, particularly from natural to anthropogenic use, in the near future. These changes demonstrate that the losses associated with ecosystem services will destructively impact human wellbeing in the city and other areas of the country. The study results provide the basic scientific knowledge for LUCC planners, urban designers and natural resources managers. They serve as a decision-making support tool for the establishment of sustainable land resource utilization policies in Vientiane and other cities of similar conditions.


Introduction
Human socioeconomic activities, the associated human population increase and general urbanization trend have caused remarkable changes in the expansion of cities around the world, threatening the sustainability of land use/cover change (LUCC) of an area [1]. The United Nations predicts that the world's population will increase about 60% by 2050 [1,2]. To date, population growth in urban areas has been found in almost all countries around the world, and it has been among the major reasons for rapid changes of land use/land cover (LULC) [3]. Various natural land covers are being converted and replaced with urban construction for residential and other human use [3][4][5]. Land conversion from its original use into another land use type is called LUCC, which is a consequence approximately 820,940 in 2015 [31]. The city has a tropical monsoon climate with an average temperature of over 80 • F (27 • C) and an average annual precipitation of 1300-2100 mm, mainly during the five months from May to September. The dry season is from October to April [27]. The geography of Vientiane is mostly a mixture of mountainous and small-scale flat areas, with the elevation ranging from 70 to 950 m above sea level. Figure 1 presents the map of the study area.
Sustainability 2020, 12, x FOR PEER REVIEW 3 of 20 during the five months from May to September. The dry season is from October to April [27]. The geography of Vientiane is mostly a mixture of mountainous and small-scale flat areas, with the elevation ranging from 70 to 950 m above sea level. Figure 1 presents the map of the study area.

Data Sources
Satellite imagery acquisition was performed giving the special consideration of the cloud cover, seasonality, and the phonological effects. In this study, Landsat images from 1995 and 2004 were collected from Landsat Thematic Mapper TM 5, and images from 2013 and 2018 were obtained from Landsat 8-OLI with a ground resolution of 30 × 30 m [32]. The track numbers of 128/47, 128/48 were acquired as shown in Table 1. Using Google Earth at 15 m resolution, The LUCC classification was connected with ERDAS software, which enabled us to obtain the position where the images from remote sensing were visible. The road data system is a critical factor derived from OpenStreetMap [33], which is processed by QGIS software and used as a conversion format file for future LUCC prediction by the CA-Markov model in TerrSet software. Advanced Spaceborne Thermal Emission and Reflection Radiometer digital elevation model (ASTER DEM) with 30 m spatial resolution was obtained with the same coordinate system as the satellite images. Google Earth with the 15 m resolution was required for image comparison.

Data Sources
Satellite imagery acquisition was performed giving the special consideration of the cloud cover, seasonality, and the phonological effects. In this study, Landsat images from 1995 and 2004 were collected from Landsat Thematic Mapper TM 5, and images from 2013 and 2018 were obtained from Landsat 8-OLI with a ground resolution of 30 × 30 m [32]. The track numbers of 128/47, 128/48 were acquired as shown in Table 1. Using Google Earth at 15 m resolution, The LUCC classification was connected with ERDAS software, which enabled us to obtain the position where the images from remote sensing were visible. The road data system is a critical factor derived from OpenStreetMap [33], which is processed by QGIS software and used as a conversion format file for future LUCC prediction by the CA-Markov model in TerrSet software.

Image Processing and Data Analysis
To ensure the accuracy of the identification of spatiotemporal changes and the geometric compatibility with other sources' information, previous studies have used different pre-processing techniques such as ArcGIS, ENVI, ERDAS, QGIS, or other software capable of handling land use classification. Thus, in this study, the images from 1995, 2004, 2013 and 2018 were classified using ERDAS IMAGINE 2014 and ArcGIS 10.3 by combined classification. The combined classification included both supervised classification and unsupervised classification methods, which are broadly used in land cover classification. The maximum likelihood algorithm (MLC) is the most commonly used method in supervised classification with remotely sensed image data [3], which computes the posterior possibility of a pixel fitting into the corresponding class based on the Bayes theorem [34]. The Universal Transverse Mercator (UTM) map coordinate system, Zone 48 North, and Datum Arc 1984 (WGS_1984_UTM_ZONE_48N) were used in these images. Google Earth Pro was used to interpret images in ERDAS 2014, which can visualize the real visible images from different years. After both supervised and unsupervised classification, the LUCCs were successfully interpreted and classified in the study area [27]. This study classified LUCCs into six classes, water bodies (WB), built-up land (BUL), intact forest (IF), degraded forest (DF), agricultural land (AL), and bare land (BL) as shown in Table 2. Table 2. Definition of land use and land cover classification types.

Land Use and Cover Class Description
Water Bodies (WB) Reservoirs, fish ponds, or drainages.
Intact Forest (IF) Untapped forests not disturbed by human activity [36], high biodiversity and vegetation cover over 70% in the area, and trees higher than 10 m [37].

Degraded Forest (DF)
Forest that has been destroyed by human activity over a long time, resulting in a lack of biodiversity, loss of species, and vegetation cover of 10-15% [38,39].
Agricultural Land (AL) Land used for cultivation, including rice paddy, garden land, rubber plantation, and grassland.
Bare Land (BL) Beach, rock, and other empty lands.
The Kappa index and overall accuracy were assessed; if the accuracy of each image is over 85%, this is considered acceptable for LUCC prediction [25,40]. The algorithm of the Kappa coefficient (K) for the LUCC classification accuracy assessment is shown in Equation (1): where K is the kappa coefficient, N is the total number of sites in the matrix, r is the number of rows in the matrix, x ii is the number in rows i and columns i, x +i is the total for row i, and x i+ is the total for the column. The flowchart of the methodology (Figure 2) presents the study procedures from data sources/preparation to the final LULC classification and prediction.

Future Prediction of LUCC Dynamic
Markov chain analysis has recently been used in various studies of future land use prediction [41]. It can determine the probability of trends from the early images to the last images to identify the transition tendency of the future LUCC probability according to the specific years [42]. The main processing of the Markov chain generates a transfer matrix and probability transfer matrix for the prediction of future land use/cover change trends. The Markov chain model can be summarized as a set of state s s , s , s , … , s . In this study, the recent state is , and it transforms to state in the next stage, with the possibility indicated by transition probabilities . Therefore, state in the system could be identified by former stage in the Markov chain; see Equation (2) [24]: 0 P 1 and Pij 1, i, j 1, 2, 3 … n where P stands for the probability matrix in the Markov model, and is the probability of converting from current state i to state j in the next period. S is the land use status, and t; t+1 are the

Future Prediction of LUCC Dynamic
Markov chain analysis has recently been used in various studies of future land use prediction [41]. It can determine the probability of trends from the early images to the last images to identify the transition tendency of the future LUCC probability according to the specific years [42]. The main processing of the Markov chain generates a transfer matrix and probability transfer matrix for the prediction of future land use/cover change trends. The Markov chain model can be summarized as a set of state s = {s 0 , s 1 , s 2 , . . . , s n }. In this study, the recent state is S t , and it transforms to state S j in the next stage, with the possibility indicated by transition probabilities P ij . Therefore, state S t+1 in the system could be identified by former stage S t . in the Markov chain; see Equation (2) [24]: 0 < P ij < 1 and n j=1 Pij = 1, i, j = 1, 2, 3 . . . n.
Sustainability 2020, 12, 8410 6 of 20 where P stands for the probability matrix in the Markov model, and P ij is the probability of converting from current state i to state j in the next period. S is the land use status, and t; t+1 are the time point-see Equation (3) [25]. A low transition probability will be near 0 and a higher transition probability will be near 1. In the Markov model, the land use/cover change prediction takes 2004 as the base (t 1 ) and the 2013 LUCC map as the later (t 2 ) image. Based on the transition matrix between the two periods from 2004 to 2013 and from 2013 to 2018, we can predict the LUCC in 2030, 2040 and 2050. Cellular automata (CA) is a bottom-up dynamic model that integrates the spatiotemporal dimension and thus adds a modeling direction. It simulates the time-space complexity even though space-time and state are discrete [20]. The CA ability is very important for land use and cover change to demonstrate and simulate the spatial and dynamic processes in LUCC prediction research. The cellular automata (CA) model mainly contains cells, cell space, neighbors, rules, and time. The neighbors are identified by the filter of the CA model. The closer the distance between the central cell and its neighbor, the greater the weight factor [24]: Here, S represents the set of states of the finite cells; t and t+1 are the early year and the later year; N is the neighborhood of cells, and f is the conversion rule of local space.
The CA-Markov model is a mixture between cellular automata and Markov chain, used to predict the transition possibility matrix in a cross-presentation of two different images so that it can stipulate a strong method for spatial-temporal dynamic modeling [43]. The CA-Markov model processes raster data in the land change modeler (LCM) by using TerrSet Clark lab software based on geotiff or tiff format data processing in ArcGIS. The transition probabilities matrix of the Markov chain model is the inputs of the CA model [44]. These are some of the procedures in the CA-Markov prediction: (1) creating a suitability atlas in multi-criteria evaluation MCE, (2) generating a state transition probability matrix and transfer matrix in the Markov model, and (3) future land use prediction using the CA model [45].
The kappa values of the CA-Markov model for LUCC simulation ranged from −1 to 1, where positive values are a sign of agreement, and negative values illustrate of a lack of agreement Kappa ≤ 0.5 shows high agreement, 0.5 ≤ Kappa ≤ 0.75 marks a moderate level of agreement, and 0.75 ≤ Kappa < 1 [14,46] indicates a high level of agreement. Equations (5) and (7) for the summary statistics are: where no information is N(n), medium grid cell-level information is M(m), and perfect grid cell-level information across the landscape is P(p).

Determined Driver Factors for LUCC Prediction in the CA-Markov Model
There are multiple factors that drive changes in land types in the study area. The digital elevation model (DEM), slope map, road data, the distance from the road, and the distance from built-up land have been taken into considered as driver factors, see Figure 3, and processed in ArcGIS and QGIS software for the LUCC prediction in the TerrSet Clark labs software.

Predicted LUCC Class Direction of Transition Potential
The system will process drivers in a run transition sub model (RTSM) using an multilayer perceptron (MLP) neural network in Terrset Clark lab software. In this study, 1358 samples per class were processed in the machine learning window. According to the results obtained, predictions of the transition potential in LUCC were made for four main districts: (1) Xaythany District, (2) Parkngum, (3) Hatsayfong, and (4) Sikhotabong; the current city expansion in 2018 is still located in three main districts, Chanthabouly (1), Sisaktanack (2), and Xaysetha (3), see Figure 1. The largest land use/cover type changes in future predictions are (1) agricultural land, (2) intact forest, and (3) degraded forest, and there is a slight change for bare land (see the future potential transition in

Predicted LUCC Class Direction of Transition Potential
The system will process drivers in a run transition sub model (RTSM) using an multilayer perceptron (MLP) neural network in Terrset Clark lab software. In this study, 1358 samples per class were processed in the machine learning window. According to the results obtained, predictions of the transition potential in LUCC were made for four main districts: (1) Xaythany District, (2) Parkngum, and there is a slight change for bare land (see the future potential transition in Figure 4). After we finalized the potential transition of the model processing, the LUCC prediction was able to continuously run the process to predict the specific time, provided by the software. After we finalized the potential transition of the model processing, the LUCC prediction was able to continuously run the process to predict the specific time, provided by the software.

Land Use/Cover Change Prediction and Validation
In this study, the LUCC classification in ArcGIS software from 2004 to 2013 and from 2013 to 2018 were processed using the CA-Markov model in Terrset software 18.31. The data processing in Terrset software based on the land change modeler (LCM). Firstly, input data in change analysis included the earlier land cover images of 2004 and the later land cover images of 2013. Then, the transition potential was used to run driver factors in the machine learning window. After we finished the machine learning processing, the land use prediction in 2018 was carried out via change prediction tools with validation data for a comparison between the actual land use in 2018 and the predicted land use in 2018. Finally, when the Kappa index was more than 80%, it meant that it was suitable for the LUCC predictions for 2030, 2040 and 2050 for 12, 10 and 10 year intervals, respectively, based on the data obtained in 2004 and 2018 that were used for the CA-Markov model integrated with Terrset software.

Classification Accuracy Verification
An important pre-requisite in the classification, detection, and prediction of LUCC studies is the accuracy and validation of the classification model [25]. The kappa statistic is a general metric applied for calculating the classification accuracy of both models as well as the user of the classification model [47]. The value of the kappa coefficient indicates the accuracy of the reference data and LUCC value in the image classification; the values of kappa are from −1 to +1 [48]. A kappa coefficient of < 0 means

Land Use/Cover Change Prediction and Validation
In this study, the LUCC classification in ArcGIS software from 2004 to 2013 and from 2013 to 2018 were processed using the CA-Markov model in Terrset software 18.31. The data processing in Terrset software based on the land change modeler (LCM). Firstly, input data in change analysis included the earlier land cover images of 2004 and the later land cover images of 2013. Then, the transition potential was used to run driver factors in the machine learning window. After we finished the machine learning processing, the land use prediction in 2018 was carried out via change prediction tools with validation data for a comparison between the actual land use in 2018 and the predicted land use in 2018. Finally, when the Kappa index was more than 80%, it meant that it was suitable for the LUCC predictions for 2030, 2040 and 2050 for 12, 10 and 10 year intervals, respectively, based on the data obtained in 2004 and 2018 that were used for the CA-Markov model integrated with Terrset software.

Classification Accuracy Verification
An important pre-requisite in the classification, detection, and prediction of LUCC studies is the accuracy and validation of the classification model [25]. The kappa statistic is a general metric applied for calculating the classification accuracy of both models as well as the user of the classification model [47]. The value of the kappa coefficient indicates the accuracy of the reference data and LUCC value in the image classification; the values of kappa are from −1 to +1 [48]. A kappa coefficient of < 0 means no agreement, 0-0.2 means slight agreement, 0.2-0.41 means fair agreement, 0.41-0.60 means moderate agreement, 0.60-0.80 means substantial agreement, and 0.81-1.0 means perfect agreement [49].
According to the historical LUCC classification results, Land-sat TM and Land-sat OLI have randomly selected 108 points for each LULC class to assess the classification accuracy by manual checking in Google Earth Pro [24]. The overall LULC classification accuracy for the years 1995, 2004, 2013 and 2018 is 92.59%, 92.59%, 87.04% and 91.67%, respectively, with the overall kappa statistics of 0.9111, 0.9111, 0.8444, and 0.9000, respectively (Tables 3 and 4).  Rapid population growth, migration from rural to urban areas, the reclassification of rural areas as urban areas, lack of evaluation of ecological services, poverty, ignorance of biophysical limitations, and the use of ecologically incompatible technologies are the main reasons behind LUCC [50]. Land use/cover changes during 1995, 2004, 2013 and 2018 were analyzed using supervised and unsupervised classification based on ERDAS and GIS software. The total area and percentage of each LUCC class were compared. According to Table 5, there is strong evidence of landscape patterns changing throughout the last two and a half decades in Vientiane. In 1995, intact forest was the largest land cover class in the study area, with the initial coverage of 63.83% (which equated to an area of 2339.56 km 2 slowly decreasing to 39.47% in 2018, as shown in Table 5). Agricultural land was the next largest land use/cover change class, with an area of 811.40 km 2 (22.14%), while degraded forest was 260.36 km 2 (7.10%). Other LUCC classes accounted for smaller proportions of the study site: built-up land covered 101.60 km 2 (2.77%), water bodies covered 98.40 km 2 (2.68%), and bare land covered 53.77 km 2 (1.47%). Intact forest was located in higher elevation areas above about 277 m in the Phu-Phanang National biodiversity area, (see Figure 5). Due to human activity, the forest degraded very quickly, losing natural balance, and biodiversity and undergoing temperature changes. Since 1990, agricultural activity has become the dominant source of income in the study area, by producing 56% of the gross domestic product (GDP) [27]; this indicates that agriculture has taken over a large area since 1995.  [51]; meanwhile, the population of Laos was 4846 million people in 1995; and increased to 7062 million by 2018, such that the mean population increased rapidly over those 23 years. The results indicate that, from 1995 to 2018, according to the Landsat image classification, the built-up area increased quickly from 101.60 km 2 to 266.00 km 2 . Water bodies area increased, while bare land decreased. Most of the water was found in reservoirs, with the Mekong River in the southwest to northwest and Namngum River in the southeast to northeast of the study area. The decrease in bare land was due to a transition from built-up land and other land. At riversides, soil erosion was significant every year due to natural disasters, heavy rainfall, and heavy wind. However, sand excavation for construction must decrease.
According to the increases and decreases shown in Table 5           The significant conversions of intact forest directly increased the areas of agricultural land, degraded forest, built-up land, water, and bare land by 76.14, 49.79, 12.70, 2.66 and 0.91 km 2 , respectively. However, the total degraded forest at that point was 288.99 km 2 ; in contrast, the land converted in 2004 represented more than double the previous land use conversions of 1995, including 409.3 km 2 agricultural land, 215.3 km 2 intact forest, 24.14 km 2 built-up land, 3.13 km 2 bare land, and 1.95 km 2 water bodies. The land use types that were less converted were water bodies and bare land.

LUCC Detection Matrix from 1995 to 2004 and 2013 to 2018
Significant conversions of intact forest accounted for 2465.23 km 2 in 1995, converting 458.64 km 2 to degraded forest, 215.31 km 2 to agricultural land, 49.48 km 2 to built-up land, 5.58 km 2 to water bodies and 2.22 km 2 to bare land. Built-up land is the land use type that increased year after year; of the total land converted to built-up land in 2004, 32.94 km 2 was from agricultural land, 19.05 km 2 was from degraded forest, 49.48 km 2 was from intact forest, 14.83 km 2 was from built-up land, and other bare land and water contributed 0.73 km 2 and 0.83 km 2 , respectively, as shown in Table 6. The total area of each land use conversion in 1995 from water, built-up land, degraded forest, agricultural land, and bare land was 92.04, 76.06, 288.99, 715.58 and 25.77 km 2 , respectively. The total land use converted in 2004 was larger, accounting for 1876.2, 781.8, 759.82, 96 and 31.36 km 2 from intact forest, degraded forest, agricultural land, water and bare land, respectively, as shown in Table 7. According to Table 8

Model Validation of Predicted Land Use/Cover Change in 2018
The area transition matrix and area transition possibility matrix were created by using two land use types in 2004 and 2013; the results are shown in Table 5 [24]. Machine learning was run by the CA-Markov model in TerrSet software. Finally, a predicted map for 2018 was produced (see Figure 9). The predictions for 2018 had a relatively high Kappa coefficient for quantity and location [9]. The validation target, kappa index of agreement (KIA) was used for the 2018 LUCC predictions, which were acceptable according to both the actual 2018 LUCC and the predicted 2018 LUCC comparison. The kappa statistics were as follows: K no is 0.8873, K location is 0.8782, K strata is 0.8782, and K standard is 0.8430, as shown in Table 9. All the kappa results showed an acceptable standard greater than 80% which confirmed that the accuracy was reasonable for future land use prediction. The results obtained in terms of both actual 2018 LUCC and predicted 2018 LUCC showed different percentages for the increase in bare land (16.86%), intact forest (15.99%), and built-up land (7.81%), while other land uses degraded forest (15.74%), water bodies (13.93%), and agricultural land (10.71%), as shown in Table 10. The corrected percentage for each type of land use and land cover was over 85%, so the model was acceptable for making predictions for 2030, 2040 and 2050.

Model Validation of Predicted Land Use/Cover Change in 2018
The area transition matrix and area transition possibility matrix were created by using two land use types in 2004 and 2013; the results are shown in Table 5 [24]. Machine learning was run by the CA-Markov model in TerrSet software. Finally, a predicted map for 2018 was produced (see Figure  9). The predictions for 2018 had a relatively high Kappa coefficient for quantity and location [9]. The validation target, kappa index of agreement (KIA) was used for the 2018 LUCC predictions, which were acceptable according to both the actual 2018 LUCC and the predicted 2018 LUCC comparison. The kappa statistics were as follows: K is 0.8873, K is 0.8782, K is 0.8782, and K is 0.8430, as shown in Table 9. All the kappa results showed an acceptable standard greater than 80% which confirmed that the accuracy was reasonable for future land use prediction. The results obtained in terms of both actual 2018 LUCC and predicted 2018 LUCC showed different percentages for the increase in bare land (16.86%), intact forest (15.99%), and built-up land (7.81%), while other land uses degraded forest (15.74%), water bodies (13.93%), and agricultural land (10.71%), as shown in Table 10. The corrected percentage for each type of land use and land cover was over 85%, so the model was acceptable for making predictions for 2030, 2040 and 2050.    The future LUCC prediction for 2030, 2040 and 2050 in Vientiane were obtained, which followed by the same method used for the prediction of LUCC for 2018. The time surveyed of prediction was accounting for 32 years from 2018 to 2050, with time intervals for land use and cover change prediction of 12, 10, and 10 years, respectively. In this study, the future prediction focused on built-up land changes. According to Table 11 and Figure 10, built-up land greatly increased in 2030, 2040 and 2050, accounting for 457.41 km 2 (12.48%), 533.71 km 2 (14.56%), and 689.44 km 2 (18.81%), respectively. Due to the fast-growing population in Vientiane mainly caused by rural-urban migration for education, jobs, and good life, the future scenarios need to take into account economic development and technologies as reasons for the rapid growth of built-up land. However, other land use types were gradually reducing by being converted into built-up land. These are for instance intact forest which the simulation showed will increase slightly in 2040 but be reduced again in 2050 by 1764.95 km 2 (48.16%) and 1717.24 km 2 (46.85%), respectively, Likewise, the degraded forest and agricultural land decreased significantly. The final LUCC maps prediction of 2030, 2040, ad 2050 were presented in Figure 11, which obviously indicates that almost all agricultural land diminished by the city expansion; the expansion was noted to extend mostly along the southeast aspect, where the major street is located.

Conclusions
This study applied satellite imagery from the period 1995-2004 from Landsat5 TM and that of 2013 and 2018 from Landsat8 OLI images to classify historical land use/cover changes in Vientiane, Laos. This used a geographic information system (ArcGIS 10.3) and ERDAS IMAGE 2014 technologies for LUCC classification. The application of machine learning in the CA-Markov model in TerrSet software predicted spatial and temporal changes in future LUCC by using the land change modeler (LCM) in TerrSet Clark Lab software. To process the CA-Markov model, a total of five drivers of urban sprawl arising from both natural and anthropogenic factors were collected. These are road network, digital elevation model (DEM), slope aspects, distance from road, and distance from urban areas. Another five independent variables were used to support the LCM analysis such as the conversion from all land uses to built-up land, conversion from bare land to built-up land, conversion from degraded forest to built-up land, conversion from agricultural lands to built-up land, and conversion from intact forest to built-up land.

Conclusions
This study applied satellite imagery from the period 1995-2004 from Landsat5 TM and that of 2013 and 2018 from Landsat8 OLI images to classify historical land use/cover changes in Vientiane, Laos. This used a geographic information system (ArcGIS 10.3) and ERDAS IMAGE 2014 technologies for LUCC classification. The application of machine learning in the CA-Markov model in TerrSet software predicted spatial and temporal changes in future LUCC by using the land change modeler (LCM) in TerrSet Clark Lab software. To process the CA-Markov model, a total of five drivers of urban sprawl arising from both natural and anthropogenic factors were collected. These are road network, digital elevation model (DEM), slope aspects, distance from road, and distance from urban areas. Another five independent variables were used to support the LCM analysis such as the conversion from all land uses to built-up land, conversion from bare land to built-up land, conversion from degraded forest to built-up land, conversion from agricultural lands to built-up land, and conversion from intact forest to built-up land.
The overall coefficients of historical LUCC classification for 1995, 2004, 2013 and 2018 were above 87%, as shown in Tables 3 and 4, which indicates a high accuracy rate. The classification results illustrated that built-up land was the type of land use increasing the fastest in the area, due to the growth in the population of the city to 528,109 in 1995; 698,318 in 2005, and 820,940 in 2015 [52]. Other land use type such as agricultural land, water bodies, and degraded forest also increased year by year; in contrast, intact forest experienced a reduction in area due to the lack of strict policies, which provide the guidelines to prohibit land expansion and deforestation in Vientiane city. The kappa index of agreement for 2018 predicting land use indicated an accuracy of above 84%; as shown in Table 9, which showed the suitability of its use for future predictions with the CA-Markov model. The statistical results of the future of the predicted LUCC illustrate that urban built-up land continued to increase rapidly, leading to the conversions of other land use types such as agricultural land, degraded forest, intact forest, etc. Other land uses like bare land and water bodies were stable, because water bodies (rivers, reservoirs and drainage) cannot be converted to urban built-up lands, while bare land is mainly located by the Mekong River and due to the need for sand for construction, is gradually but significantly reduced. The assessment of agricultural land and degraded forest led to concerns about these land use types disappearing in the near future due to farming activities and urban sprawl. Due to the urban sprawl in Vientiane, the hotspot areas may change. However, the scale of urban sprawl in the four major districts, Xaythany, Parkngum, Hatsayfong, and Sikhotabong is likely to be large, whereas the other five districts may be gradually reduced in scale over time.
From the analysis of previous and future land use and land cover change, the change has obviously been slow in past years, but the speed increased up to 2018, mainly due to the fast development of the city. Nevertheless, when compared to relative land use variation in other countries with more rapidly increasing populations, the impact in Vientiane is still under lower change.
The approach, technologies, and model used in this study obtained a high kappa coefficient, which demonstrates that the model can produce high-quality simulated results and can be used as a reference in other research and other study areas. The recommendations of this paper can be applied to any urbanized expansion analysis in a CA-Markov model based on the existing drivers in the selected study area.
Finally, this study recommends to urban designers and policymakers who take responsibility for city planning to consider adopting an ecosystem-based approach to socioeconomic and ecological problems in the city. Such approaches can be implemented in short-term and long-term development plans. On the top of that, the urban development plans must provide policies and financial support for implementing smart interventions such as ecosystem conservation, farming zones and water conservation. They must also manage settlement expansion, restore and protect water sources and watersheds, increase the forest cover area through the afforestation mechanism (green city approach) and pass regulations that restrict unplanned settlement for sustainable urban development.