**2. Methodology**

classified by country‐size class; that information was obtained from the World Urbanization Prospects reported by the United Nations (UN) [1]. In the five countries of the Indochinese Peninsula, there are 26 cities with populations of more than 300,000 and 16 cities with popu‐ lations of more than 500,000. From the table, we conclude that Thailand and Vietnam contain many more big urban agglomerations than the other countries, thus showing their high level of urbanization. Second, despite the presence of some metropolitan cities, urbanization in countries of the Indochinese Peninsula is broad‐based. Only 1 city with a population of more than 300,000 can be found in Cambodia and Laos; Thailand and Vietnam have only 1 supercity (5 million or more inhabitants) each. Increasingly, after a process of transition and transfor‐ mation, modernization and industrialization are emerging in Myanmar, Cambodia, Laos, Thailand, and Vietnam. These countries are gradually shifting from traditional farming to more diversified economies and to more open market‐based systems [7]. Parallel with this development are the growing economy links between the five countries and their neighbors, notably in terms of cross‐border trade, investment, and labor mobility [8]. Moreover, natural resources, particularly hydropower, are beginning to be developed and used in the region [9]. The Silk Road Economic Belt and the 21st Century Maritime Silk Road [One Belt and One Road (B&R)], which was proposed by China in 2014, comprise a development strategy and frame‐ work that aims to enhance the economic relationship among countries in Asia and Europe [10]. The China‐Indochinese Peninsula Economic Corridor, as an important international gateway of B&R, is supposed to develop a regional economic entity with common development that uses the railways and roads as a medium. The rich human and natural resource endowments of the Indochinese Peninsula region have made it a new frontier of Asian economic growth. The Indochinese Peninsula has tremendous potential to promote both regional economic growth and urban development. Thus, following the 2014 launch of B&R, the China‐Indochi‐ nese Peninsula Economic Corridor will witness a more rapid urbanization progress in the next

**Size class Cambodia Laos Myanmar Thailand Vietnam** 5 million or more – – – 1 1 1–5 million 1 – 3 1 3 500,000–1 million – 1 1 2 2 300,000–500,000 – – 2 5 3

**Table 1.** Number of agglomerations classified by country‐size class in the Indochinese Peninsula, 2015.

This chapter gives a general overview of the urbanization procession in countries in the Indochinese Peninsula region and presents the state‐of‐the‐art techniques for monitoring the spatiotemporal dynamics of the urban environment. An analysis of the forces driving urban expansion was also performed based on an integrated analysis of both natural and social

Sources: World Urbanization Prospects: The 2014 Revision, UN.

decade.

6 Sustainable Urbanization

economic factors.

### **2.1. Monitoring urban expansion in the Indochinese Peninsula**

Urban sprawl monitoring constitutes basic information for urban studies, and accurate information about the extent of urban growth is of great interest to researchers investigating urbanization progress. Long ago, conventional surveying and mapping techniques were primarily used to estimate urban sprawl. These methods were usually expensive and time‐ consuming, and some key information was unavailable for most cities, especially in develop‐ ing countries [11]. Fortunately, because remote sensing enjoys the advantages of being both cost‐effective and technologically sound, it is increasingly used in the analysis of spatiography and urban geography. Since the 1970s, a variety of studies have been conducted using remote sensing and Geographic Information System (GIS) technologies to examine land‐use change, to analyze large landscapes, to analyze farmland change and classification, and to analyze urban space structure and fractal shapes [12–15]. In recent years, with rapid economic development and population increases, rapid urbanization occurs and the city quickly expands; the strained relationship between the population and urban land‐use is attracting increasingly broad scholarly attention. Thus, extensive research studies have been performed to monitor urban sprawl using remotely sensed images through either an image‐to‐image comparison or a postclassification comparison [16–19]. In the countries of the Indochinese Peninsula, there has also been a great deal of research effort devoted to land‐use changes using remote sensing and GIS technologies. Kong et al. [20] investigated forestland changes attrib‐ uted to urbanization and agricultural land expansion in Naypyidaw (Myanmar's capital) using Landsat images. Similarly, based on an analysis of the pattern of urban growth from 1993 to 2011 in Siem Reap, Cambodia, Ourng and Rodrigues [21] reported that urban growth always came accompanied primary roads and the river. Using remote sensing images, Okamoto et al. [22] studied the urbanization of Vientiane (capital of Laos). Kimijiama and Nagai [23] also presented the relationship between urbanization and socioeconomic activities in Savannaket, Laos. Nevertheless, because of the limited availability of regional spatial data, those previous studies rarely focused on urban sprawl at the national level; moreover, there is no systematic study available on the multiple temporal phases and long time series of urban sprawl moni‐ toring by multiple‐sourced remote sensing data in the countries of the Indochinese Peninsula. Remote sensing and GIS technologies have broad application space in the estimation of urban sprawl in this region.

#### *2.1.1. Methods for extracting urban land with remote sensing*

Remote sensing image classification is a primary method of extracting land surface information at large scale. Different ground objects have different spectral characteristics, which are recorded in satellite remote sensing images, and pixels with similar spectral characters would be considered as one landscape class.

The process of landscape classification is complex. Its accuracy is always influenced by many factors such as the sources of the data, the image quality in remotely sensed images, and the method selected for the classification [24]. The classification method is important for landscape classification. Typically, the classification can be divided into two types: (1) pixels compared one by one, which involves monitoring the changes to each pixel by comparing the various temporal phases, and (2) first classifying and then comparing, which involves classifying the remote sensing images in different temporal phases separately and then comparing the classification results to monitor the land‐use change [25, 26]. The various methods all have their own advantages and disadvantages; therefore, there may be no single method that is suitable for all situations. For example, although the previous method is simple and easy to achieve, it only monitors the changes of the pixel rather than obtaining the changes of the objects; the second method is limited by classification accuracy in different temporal phases and accumulates calculated errors, thus making its accuracy dependent on the accuracy of the previous classification result. In reality, we should select a landscape classification method based on our needs. The common methods for extracting urban information are as follows:

## *2.1.1.1. Supervised classification*

Supervised classification is also known as the training classification method. In this method, we select training samples of different land‐use types in the image and analyze the sample information for each training area by the computer. Next, each pixel can be classified into the similar sample area based on the comparative result of the pixel and the training samples [27– 29]. The common methods for supervised classification are the single linkage method [30], the Fisher discriminant method [31, 32], the Bayes linear discriminant analysis [33], and the maximum likelihood method [34]. Supervised classification has the advantage of selectively determining the quantitative and categorical classification based on the study object and area while eliminating needless classifications. Additionally, supervised classification can control the selection of the training samples. Supervised classification is limited, however, by the subjective factors of humans to select the training samples and determine the classification system [35].

#### *2.1.1.2. Unsupervised classification*

Unsupervised classification primarily relies on the structural features of the image data and natural points for the object classification without the known training data and the number of the classification. Based on similar levels of the luminance value of the samples in the multi‐ dimensional spectrum space, the computer can automatically analyze the classifying param‐ eters and then classify the pixels accordingly [36–38].

Unsupervised classification depends on similar levels of the pixels' luminance value for object classification instead of relying on prior knowledge. The dependency on spectra quality replicates the biggest flaw of unsupervised classification; because of differences in location, shape, and character, the same ground objects may have different manifestations in the spectra image, inducing errors in classification. Compared with the method of supervised classifica‐ tion, the unsupervised classification is neither fast nor precise but does have the advantage of being highly objective [39]. According to the study by Xue and Ni [40], although unsupervised classification is unsuitable for the extraction of residential areas compared to the supervised classification, the selection of the training areas has a greater impact on classification accuracy determined by the supervised classification method. Moreover, Hu et al. [41] also extracted urban land‐use information using the two methods.

#### *2.1.1.3. Visual interpretation*

classification. Typically, the classification can be divided into two types: (1) pixels compared one by one, which involves monitoring the changes to each pixel by comparing the various temporal phases, and (2) first classifying and then comparing, which involves classifying the remote sensing images in different temporal phases separately and then comparing the classification results to monitor the land‐use change [25, 26]. The various methods all have their own advantages and disadvantages; therefore, there may be no single method that is suitable for all situations. For example, although the previous method is simple and easy to achieve, it only monitors the changes of the pixel rather than obtaining the changes of the objects; the second method is limited by classification accuracy in different temporal phases and accumulates calculated errors, thus making its accuracy dependent on the accuracy of the previous classification result. In reality, we should select a landscape classification method based on our needs. The common methods for extracting urban information are as follows:

Supervised classification is also known as the training classification method. In this method, we select training samples of different land‐use types in the image and analyze the sample information for each training area by the computer. Next, each pixel can be classified into the similar sample area based on the comparative result of the pixel and the training samples [27– 29]. The common methods for supervised classification are the single linkage method [30], the Fisher discriminant method [31, 32], the Bayes linear discriminant analysis [33], and the maximum likelihood method [34]. Supervised classification has the advantage of selectively determining the quantitative and categorical classification based on the study object and area while eliminating needless classifications. Additionally, supervised classification can control the selection of the training samples. Supervised classification is limited, however, by the subjective factors of humans to select the training samples and determine the classification

Unsupervised classification primarily relies on the structural features of the image data and natural points for the object classification without the known training data and the number of the classification. Based on similar levels of the luminance value of the samples in the multi‐ dimensional spectrum space, the computer can automatically analyze the classifying param‐

Unsupervised classification depends on similar levels of the pixels' luminance value for object classification instead of relying on prior knowledge. The dependency on spectra quality replicates the biggest flaw of unsupervised classification; because of differences in location, shape, and character, the same ground objects may have different manifestations in the spectra image, inducing errors in classification. Compared with the method of supervised classifica‐ tion, the unsupervised classification is neither fast nor precise but does have the advantage of being highly objective [39]. According to the study by Xue and Ni [40], although unsupervised classification is unsuitable for the extraction of residential areas compared to the supervised classification, the selection of the training areas has a greater impact on classification accuracy

*2.1.1.1. Supervised classification*

8 Sustainable Urbanization

*2.1.1.2. Unsupervised classification*

eters and then classify the pixels accordingly [36–38].

system [35].

Visual interpretation relates to extracting information about specified ground objects from the remote sensing image using either direct observations or assisted instruments [42]. Visual interpretation, which enjoys the advantages of easy operation and less equipment, is popular with geographers. The interpretation marks of visual interpretation can be categorized as direct and indirect. The ground object directly interpreted by the characteristic of the image is defined as the interpretation mark. The basic elements of the direct interpretation mark are the tone and the diagram: the tone reflects the image's physical property, whereas the diagram reflects its geometric properties. The tone is the analog recording of the grayscale, which shows the color code and chroma in the color image. In visual interpretation, although the tone of the recognized ground object is a quick mark, it is an uncertainty criterion because it suffers from various influence factors. Thus, the tone could merely be a relative reference for interpretation, and we cannot identify the ground object by relying exclusively on it [43]. The diagram can reflect the shape, size, location, and plane relation among the ground objects. In general, geomorphologic shape, vegetation distribution, water bodies, and bare land are all interpreted by the tone and diagram. Moreover, clouds, snow, urban land, open‐pit mines, and airports can be identified by the image [44]. Based on the studies of the phenomenon close to the interpreted objects, indirect interpretation is defined as inferring and distinguishing among the ground objects. Location, relative positions, and other things close to the interpreted objects can all be regarded as indirect marks. Location is both the reflection of the environment of the objects in the image and the relationship between the objects and the environment. Relative positions relate to the plane layout of the dependence relationship among the landscape elements and the objects in the image [43]. The common methods of visual interpretation include the direct identification method, the comparison method, and logical reasoning. The direct identification method can quickly interpret ground objects by their marks, whereas logical reasoning needs to identify objects' existence and properties by the appearance of the internal relations of the objects or natural phenomena. The comparison method first compares the ground objects and natural phenomenon in the image with a known remote sensing image and then identifies the properties of the objects. In any event, it is very important to analyze the comprehensive feature of interpretation objects for each visual interpretation method. To improve the precision of the interpretation, direct and indirect interpretation marks should be used conjunctively, and the image should be taken as a contrastive analysis of various bands and temporal phases [45].

#### *2.1.1.4. Automatic classification and change detection*

On a regional scale, moderate spatial resolution remote sensing images such as land resource satellite data are usually used for the data used in landscape classification and change detection. To rapidly achieve an accurate classification rapidly and to minimize human intervention, Jiang et al. propose an efficient, automatic landscape classification approach taking prior accurate land‐cover data as the background experience [46, 47]. By adopting the prior knowledge, this approach is distinguished from the previous semisupervised findings of landscape classifiers. This approach involves two steps. First, based on the historical image data, one detects changed landscape pixels from satellite images. Second, one classifies the changed pixels in the landscape based on pattern recognition and changed rules. This approach enjoys the advantages of multimethods in landscape classification, primarily described as follows: (1) the historical data for land‐use cover is high precision and can be better matched with the remote sensing data, and (2) based on the ecology view, this approach assumes that the junction of different land types is the fragile area, which is the main changed areas and the inner area is the relatively stable region. Furthermore, the big plaque will be more stable than the small one. This approach can be applied not only to microsatellite data but also to landscape classification for other spectral remote sensing images.

### *2.1.1.5. Normalized Difference Built‐up Index (NDBI) method*

Based on the deep analysis of the Normalized Difference Vegetation Index (NDVI), the NDBI was first proposed by Yang [48] and later improved by Zha and Ni [39]. For the Landsat TM image, the gray value of the objects will show only small changes, except for the urban land in the bands of TM4 and TM5. Based on the spectral characteristics, NDBI achieves urban land extraction using the following formula:

$$NDBII = \frac{(TM\,\mathsf{5} - TM\,\mathsf{4})}{(TM\,\mathsf{5} + TM\,\mathsf{4})} \tag{1}$$

where TM4 is band 4 and TM5 is band 5 of the TM image. Obviously, the value of NDBI should be between −1 and 1; after the two binary transform, the interval value of −1 and 0 is assigned as 0 and the others are assigned as 255, then obtaining the binary image for urban land extracting.

Based on the National Oceanic and Atmospheric Administration images, NDVI, which is used for vegetation‐information extraction on a regional scale, has tended to be mature. In recent years, this approach, which has been improved for urban land extracting, has been widely used in urban sprawl monitoring [41, 49].

#### *2.1.1.6. Artificial neural network (ANN) classification*

ANN is a complicated network system that is composed of abundant and simple processing elements; it contains engineering systems that simulate the operative mechanism and organ‐ izational structure of the human brain based on studies of human brain. ANN belongs to the nonparametric classifier. Since it was proposed in 1988, this approach has been attracting increasing attention to landscape classification and change detection. ANN is widely used in landscape classification such as land‐use classification, ground object identification in different temporal phases, fuzzy classification, remote sensing image classification, and the extraction of the shape structure of the image. With the development of the theory of ANN and techno‐ logical improvements, ANN has been an effective means for remote sensing classification. In recent years, numerous studies on ANN have been successfully performed for the geologic application. According to the characteristic of remote sensing, Dong et al. propose a landscape classification model based on the Hopfield ANN; this classifier proved to have better accuracy and higher efficiency than other methods [50]. Chen et al. have developed the Self Organizing Feature Map (SOFM) neural network, which is based on the weight of samples and data, to achieve the direct classification change detection for temporal and multispectral remote sensing data [51]. Common ANN models for landscape classification are the Multilayer Perceptron (MLP) classification model, the radial basis function (RBF) neural network classification model, the SOFM classification model, and the Adaptive Resonance Theory (ART) classification model. Aside from the above models, more other ANN models have been gradually applied to remote sensing classifications. In general, compared to other classification methods, ANN for land‐use extraction has the following advantages: (1) it has the properties of self‐learning, self‐organizing, and self‐adapting and can not only make maximum use of prior knowledge of the known samples type but also automatically extract the rules for multiclassification; (2) ANN need not make the assumptions of the probabilistic models; (3) with its capacity for fault tolerance, nonlinear decision boundaries can be developed in feature space in the ANN model; and (4) the ANN model has superior association power [52].

#### *2.1.2. Satellite data for extracting urban land*

#### *2.1.2.1. Resource satellite data*

Land resource satellites are primarily used for resource exploration and studying the natural ecoenvironment status of land surface; they are widely used for resource investigation, environmental monitoring, hazard monitoring, land‐use planning, and regional development. The most common resource satellites are NASA's Landsat, France's SPOT, the China Brazil Earth Resource Satellite, India's Cartosat‐1 (IRS‐P5), and the Moderate Resolution Imaging Spectroradiometer (MODIS). The following is a detailed introduction to NASA's Landsat.

#### *2.1.2.1.1. Landsat*

taking prior accurate land‐cover data as the background experience [46, 47]. By adopting the prior knowledge, this approach is distinguished from the previous semisupervised findings of landscape classifiers. This approach involves two steps. First, based on the historical image data, one detects changed landscape pixels from satellite images. Second, one classifies the changed pixels in the landscape based on pattern recognition and changed rules. This approach enjoys the advantages of multimethods in landscape classification, primarily described as follows: (1) the historical data for land‐use cover is high precision and can be better matched with the remote sensing data, and (2) based on the ecology view, this approach assumes that the junction of different land types is the fragile area, which is the main changed areas and the inner area is the relatively stable region. Furthermore, the big plaque will be more stable than the small one. This approach can be applied not only to microsatellite data but also to landscape

Based on the deep analysis of the Normalized Difference Vegetation Index (NDVI), the NDBI was first proposed by Yang [48] and later improved by Zha and Ni [39]. For the Landsat TM image, the gray value of the objects will show only small changes, except for the urban land in the bands of TM4 and TM5. Based on the spectral characteristics, NDBI achieves urban land

> ( 5 4) ( 5 4) - <sup>=</sup> <sup>+</sup> *TM TM NDBI*

where TM4 is band 4 and TM5 is band 5 of the TM image. Obviously, the value of NDBI should be between −1 and 1; after the two binary transform, the interval value of −1 and 0 is assigned as 0 and the others are assigned as 255, then obtaining the binary image for urban land

Based on the National Oceanic and Atmospheric Administration images, NDVI, which is used for vegetation‐information extraction on a regional scale, has tended to be mature. In recent years, this approach, which has been improved for urban land extracting, has been widely

ANN is a complicated network system that is composed of abundant and simple processing elements; it contains engineering systems that simulate the operative mechanism and organ‐ izational structure of the human brain based on studies of human brain. ANN belongs to the nonparametric classifier. Since it was proposed in 1988, this approach has been attracting increasing attention to landscape classification and change detection. ANN is widely used in landscape classification such as land‐use classification, ground object identification in different temporal phases, fuzzy classification, remote sensing image classification, and the extraction of the shape structure of the image. With the development of the theory of ANN and techno‐

*TM TM* (1)

classification for other spectral remote sensing images.

*2.1.1.5. Normalized Difference Built‐up Index (NDBI) method*

extraction using the following formula:

used in urban sprawl monitoring [41, 49].

*2.1.1.6. Artificial neural network (ANN) classification*

extracting.

10 Sustainable Urbanization

Landsat‐1, which the United States developed in 1972, is the first resource satellite in the world. It orbits at 704 km high and an angle of 98.2°; it circles the Earth in 16 days. Landsat‐2 and Landsat‐3 were launched in 1975 and 1978, respectively. The three satellites were the first‐ generation test satellites and carried the same sensors [i.e., the Return Beam Vidicon (RBV) and Multispectral Scanner System (MSS)]. Landsat‐4 and Landsat‐5 were launched in 1982 and 1984, respectively. They were first‐generation practical satellites, carrying the MSS and Thematic Mapper (TM) sensors. Landsat‐7, launched in 1999, was the third‐generation satellite. The Enhanced TM Plus (ETM+) sensor, which had eight bands, was first carried in Landsat‐7, replacing the TM sensor [53]. The new sensor can work for the spectral region of visible light, near infrared, shortwave infrared, and thermal infrared, and it contains the following improvements over the TM: (1) it introduced the panchromatic band with a resolution of 15 m, (2) the resolution of thermal infrared band was improved to 60 m, and (3) the solar calibrators reduced the satellite's radiation‐calibration errors to less than 5%, which is five times that of a traditional satellite. Landsat‐8, which carried the Operational Land Imager (OLI) sensor and the Thermal Infrared Sensor (TIRS), was launched in 2013 by NASA [54]. OLI includes all of the bands in ETM+, improved to exclude water absorption characteristics. Comparing the previous sensors, OLI excluded the water absorption characteristics in 0.825 μm in band 5; the range of panchromatic band 8 was narrow, which can help in differentiating the vegetation and nonvegetation. Furthermore, two added bands were included in Landsat‐ 8, which were band 1 coastal (0.433–0.453 μm) for coastal zone monitoring and band 9 cirrus (1.360–1.390 μm) for clouds monitoring; the near‐ and short‐infrared bands were closer to the MODIS bands. **Tables 2** and **3** show the Landsat launched by the United States and the bands included in ETM+ and OLI.


**Table 2.** Landsat launched by the United States for remote sensing monitoring.


**Table 3.** Bands included in ETM+ and OLI.
