**3. Application of machine learning in GIS**

Urban dispersal and expansion has become an important issue for municipalities, environmental scientists and urban planners. Especially, in megacities, this issue becomes more vital. Currently, more than %50 of the world's population lives in urban areas and then it is predicted to grow over the %65 by 2050, according to the United Nation report. For example, all population in the 500,000+ urban areas of Australia and New Zealand combines to equal that of Moscow or Bangkok, and only slightly larger than Los Angeles (16.4 million). It is known that developing countries have already begun a rapid urbanization [4]. The fact that the global population has increased rapidly since the industrial revolution of the 18th century, highlights the problems of urban planning and urbanization, because of the population gathering in certain centers [5]. This unnatural pace of urbanization has created significant social and environmental challenges for decision-makers [6]. In addition, modeling and simulation are effective tools for discovering the urban development mechanisms and for providing planning in growth management. Therefore, monitoring and modeling the urban sprawl of cities is a necessary key parameter to prevent precautions [7, 8].

As it has been illustrated in **Figure 2**, Asia remains the dominance in terms of megacities, with nearly 58 percent of the population in larger metropolitan areas. This is approximately five times as many greater urban area residents as in North America or Africa. Besides, Asia has more than five times as many larger urban area residents as Europe and eight times that of South America [3].

**3.1 Process description: impact and influence determination**

*A Review of the Machine Learning in GIS for Megacities Application*

*DOI: http://dx.doi.org/10.5772/intechopen.94033*

of dynamic metropolitan environments [36, 37].

(CNN) based approaches were used in [39, 41, 42].

tion [43, 47].

**35**

unreferenced functions [43].

ImageNet Large Scale Visual Recognition Competition [44].

Over the years, the fast development of map services [29] and volunteered geographic information (VGI) [30] has provided a huge number of geo-tagged images. This data source has given information on every corner of a city and has been enabling broader and more in-depth quantitative research in related fields. These data improve the understanding of the dynamic and physical features of the city by identifying landmark [31], detecting urban identities [32, 33], assessing the living environment inequality [34], and modeling human activities [35] and popular places [36]. Also, they provide information on the social and physical structures

The MIT Media Lab launched the "Place Pulse" program in 2013, which is a data

collection platform that enables volunteers to take part in the urban perception rating experiment. By the end of 2016, the MIT Place Pulse dataset had collected 1,170,000 pairwise comparisons from 81,630 online participants for 110,988 cityscape images. Given this dataset and advances in ML techniques, many studies have

Since previous approaches use low or mid-level image attributes, they have problem in extracting high-level information about the natural image. Some examples of these attributes are: Gist, SIFT- Fisher Vectors, DeCAF features [41], geometric classification map, color Histograms, HOG2x2, and Dense SIFT [40]. For instance, according to building models, SVM and Linear Regression were used in [41] to predict image labels. Support Vector Regression was used in [40], Ranking SVM it has been used in [42], and several convolutional neural network

Among the various image representations and models, approaches that uses deep convolutional neural network (DCNN), have outperformed conventional methods to a large extent. In [39], authors introduce a DCNN model that is based on

Recently, a shift-invariant and hierarchical model has emerged in the form of DCNN, because of the availability of large-scale annotated datasets and the rapid development of high-performance computing systems. In some research, DCNN was employed to conduct human perception modeling and prediction. Due to its powerful capability to learn and represent automatic image feature, this model has attracted a lot of attention and achieved great success in multiple fields, including speech recognition [45], natural language processing [46], and visual object detec-

A very deep convolutional neural network is hard to train and optimize because

In [49], a data-driven ML approach that measures how people perceive a place in a large-scale urban region, was proposed. In particular, a deep learning model, that had been trained with millions of human ratings of street-level imagery, was used to predict human perceptions of a street view image. The model achieved a high accuracy rate in predicting six human perceptual indicators. These indicators are: beautiful, boring, depressing, lively, safe and wealthy. It could help to map the distribution of the city-wide human perception for a new urban area. Besides, to determine the visual points that may cause a place to be perceived as various perceptions, a series of statistical analyses was performed. From the 150 object categories that had been segmented from the images of the street view, many

of disappearance gradients and the curse of dimensionality [43, 48]. ResNet is known as an acceptable attempt to address this problem. It was designed to learn the residual functions with regard to the layer inputs rather than learning the

the Deep Residual Network (ResNet) [43], which won the first place in the

tried to analyze human perceptions of urban appearance [34, 38–41].

#### **Figure 2.**

*Built-up urban area population in 2019.*

Urban expansion modeling became widespread in the 1960s [9] and have accelerated with the developments of the technologies like Remote Sensing (RS) and GIS. Today, RS has been widely recognized as an essential tool for urban planning, management and design due to allows to get spatiotemporal data that are necessary for modeling environmental impacts, urban expansion and population growth. Particularly, the benefits of satellite-based image data have attracted attention of the scientific studies on urban expansion and environment [10, 11]. RS enables the collection of spatial details data for large areas at different time intervals; therefore, it provides a unique perspective on revealing the spatial and temporal dynamics of the change process in the land use and urban expansion [12]. GIS technology is described as an effective tool for identifying and monitoring the land cover change at different scales [13–16].

The dynamic modeling via GIS as a tool for urban simulation has rapidly gained popularity [7, 17]. The application of ML models has increased noticeably in RS filed due to the increased availability of powerful and flexible ML software and improvements in computing performance [18, 19].

The useful application of artificial neural networks (ANNs) in interpreting spatial resource information have been proven i.e. one of the most common are backpropagation neural networks, which are widely used by spatial planners. However, to improve the usability of ANNs for map-based applications, a more efficient method for communicating between the GIS and a trained ANN, is critical. When ANNs and GIS are used together for many applications to improve decision-making quality. The ANN design will consist of numerous layers, all of which can have different weights. The training process of an ANN involves changing the weightings over time until as it is desired the network reaches the static or optimum firing state. After training, an ANN can be used for applications effectively and consistently. Through the application of an ANN, GIS professionals can add another dimension to their spatial capabilities. In some research, a combination of neural networks with remote sensing image data for mapping the urbanization dynamics, has been proposed [20–22].

As another helpful technic, recently evolutionary algorithms have been used to tackle a variety of complex computation and optimization problems, such as natural language processing [23], route finding [24] and image processing [25]. One of the most important applications of evolutionary algorithms is in the field of GIS [26–28]. It should be noted that generally these algorithms are not fast in comparison with other ML algorithms and their main usage is for optimization.

All in all, there are two main ways for satellite-image processing that each of them has its own advantages and disadvantages. Sometimes there is high-resolution data (satellite images) so processing this amount of data would take a lot of time, with high accuracy. Some other cases, depending on the issue the resolution is not good, but the measurement methods are highly qualified.
