SPATIOTEMPORAL LAND USE CHANGE ANALYSIS AND FUTURE URBAN GROWTH SIMULATION USING REMOTE SENSING: A CASE STUDY OF ANTALYA

The objectives of this study are: to create land-use maps by 5-year interval from 1995 to 2015, to analyse the land use change and urban development, and to estimate future land-use pattern and urban growth for the years: 2030, 2045 and 2060. Antalya, which is the 5 biggest city of Turkey, was selected as study area. In this study, there are basically three stages: (i) preprocessing and preparing additional bands, (ii) spatiotemporal land use detection using image classification and (iii) land use simulation using urban growth models. Firstly, atmospheric correction was applied to the Landsat 5 TM and Landsat 8 OLI images and land-cover indices, ASTER Global Digital Elevation Model (GDEM), and Nighttime data were prepared to use them as additional bands during the classification process. Secondly, Landsat images were classified using Random Forest (RF) machine-learning algorithm. Thirdly, urban simulations were performed for the years 2005, 2010, and 2015 and land-use pattern and urban growth was estimated for the years 2030, 2045 and 2060. The RF classification accuracies range from 84.44% to 92.82%. The urban areas increased from 49.56 km to 96.25 km from 1995 to 2015. The simulation accuracies were computed above 80%. According to the 2030, 2045 and 2060 simulation results, the urban areas were computed as 133.61 km, 148.27 km and 156.85 km, respectively. As a result, it was seen that the urban area of Antalya has almost doubled between the years 1995-2015 and the urban expansion is expected to continue increasing up to 1960. * Corresponding author


INTRODUCTION
People and therefore the populations are in the heart of sustainable development and urbanization, population growth, population ageing, migration trends are important parameters for sustainability and economic and social development (United Nations, Department of Economic and Social Affairs, Population Division, 2019). Globally, more people live in urban areas than in rural areas and by the year 2018, 55% of the world's population living in urban areas, while this rate is expected to be 68% by 2050 (United Nations, 2018). Urban population growth is increasing due to the total population growth and the increase in the percentage of urban residents, and these two factors are expected to add 2.5 billion to the world urban population by 2050 (United Nations, 2018). The rapidly increasing world population, the excessive and often uncontrolled urbanization lead many environmental, social and economic problems. Land-Use/Land-Cover (LULC) pattern is important for many urban research areas including urban growth prediction, environmental monitoring, land surface temperature calculation and climate change analysis. Hence, determining the LULC pattern, examining LULC change and estimating the future LULC pattern are the subjects that do not lose their popularity.
Urban growth models simulation and prediction has become imperative for ecosystem conservation and sustainable development (Yao et al., 2015). Urban growth models are tools that used to predict future urban expansion and/or change. Urban growth models such as Cellular Automata (Batty et al., 1999;Torrens, 2003), SLEUTH (Clarke et al., 1997), Logistic Regression (Landis, 1994) or Artificial Neural Network (Pijanowski et al., 2001) are the models used for urban development forecasting based on real situations and urban trends and providing graphical output (Newman et al., 2016). Cellular Automata (CA) is a mathematical model that consists of a lattice of cells that repeatedly update their status taking into account the transition rules and spatial neighborhood (Wolfram, 1984). According to the Aburas et al. (2016), the CA model is one of the most powerful models for urban development simulation and forecasting. MOLUSCE (Modules for Land Use Change Assessment) is a user-friendly and free plugin developed by Asia Air Survey, Ltd for Quantum GIS. This plugin is based on the Monte Carlo CA model approach (Gismondi, 2019). It randomly selects user-defined sample points to be used for model calibration and validation and supports four different modeling methods: (Artificial Neural Network (ANN), Logistic Regression (LR), Multiple Criteria Assessment (MCA) and Weight of Evidence (WoE)). It then computes calibration statistics, generates transition potential maps and simulates based on the Monte Carlo CA model approach (Gismondi, 2019). MOLUSCE was used in this study because it is free and based on CA-modeling.
The objectives of this study are: to create land use maps by 5year interval from 1995 to 2015, to analyse the land use change and urban development, and to estimate future land use pattern and urban growth for the years: 2030, 2045 and 2060. In this study, there are basically three stages: (i) preprocessing and preparing additional bands, (ii) spatiotemporal land use detection using image classification and (iii) land use simulation using urban growth models. In the first stage, initially atmospheric correction was applied to the Landsat 5 TM and Landsat 8 OLI images using DOS1 Atmospheric Correction Module of QGIS software. Then, Normalized Difference Vegetation Index (NDVI), Normalized Difference Build-up Index (NDBI) and Modified Normalized Difference Water Index (MNDWI) were generated for each date. In order to increase the classification accuracy, in addition to these indices, ASTER Global Digital Elevation Model (GDEM), Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS) and Visible Infrared Imaging Radiometer Suite Day/Night Band (VIIRS DNB) Nighttime Lights (NTL) data were used as additional bands during the classification process. In the second stage, Landsat images were classified using the Random Forest (RF) machine-learning algorithm. In the third stage of the study, urban simulations were performed for the years 2005, 2010, and 2015 for validating the simulation results by comparing them with classification results. Then, future land use pattern and urban growth were estimated for the years 2030, 2045 and 2060.

STUDY AREA AND USED DATA
Antalya is one of the most prominent cities of Turkey with its touristic and agricultural properties. Antalya city covers 20723km 2 (Antalya Provincial Culture and Tourism Directorate, 2020). The population of Antalya has increased from about 1.4 million to about 2.5 million from the years 2000 to 2019 (Turkish Statistical Institute, 2020). It is the 5 th biggest city of Turkey in terms of population with high urbanization rate and receives millions of tourists every year. The central (Aksu, Dosemealti, Kepez, Konyaalti, Muratpasa) and Serik districts, where the majority of the population live and urban development is observed most, are chosen as the study area. (Figure 1).   (Gallo et al., 1995) and can be used in urban research, such as urban expansion, energy consumption, population density, disaster assessment, economic assessment, and urban mapping (Elvidge et al. 2007;Ghosh et al. 2010;Chaturvedi et al. 2011, Zhang andLi, 2018;Goldblatt et al., 2018). In addition, slope, population, forest, distance to road, natural protection and agricultural land data were used in simulation processes.

METHODOLOGY
In this study, there are three stages: (i) preprocessing and preparing additional bands, (ii) spatiotemporal land use detection using image classification and (iii) land use simulation using urban growth models.
In the first stage, initially atmospheric correction was applied to the Landsat 5 TM and Landsat 8 OLI images. Then, NDVI, NDBI and MNDWI were generated for each date. In the second stage, Landsat images were classified using the RF classifier and additional bands. To increase the classification accuracy, ASTER GDEM, DMSP/OLS and VIIRS DNB NTL data were used as additional bands during the classification process. Finally, in the third stage urban simulations were performed for the years 2005, 2010, and 2015 for validating the simulation results by comparing them with classification results. Next, future land use patterns and urban growth models were simulated and estimated for the years 2030, 2045 and 2060. The flow-chart of this study is shown in Figure 2 and processing steps are explained in detail in the following sections.

Figure 2. Flow-chart of study
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2020, 2020 XXIV ISPRS Congress (2020 edition)

Pre-processing and Preparing Additional Bands and Data
Initially, Dark Object Subtraction (DOS) atmospheric correction method, DOS1 (Moran et al., 1992) was applied to the downloaded Landsat images with Semi-Automatic Classification Plugin (SCP) of the QGIS software. Later, in order to increase the classification accuracy, NDVI, NDBI and MNDWI indices were created for each date. These indices highlight certain objects in the image and help identify them more easily. While NDVI (Rouse et al., 1974) index is used for especially distinguishing the vegetation, NDBI (Zha et al., 2003;Liu and Zhang, 2011) and MNDWI (Xu, 2006) indices are used for built-up and water areas detection, respectively. NDVI, NDBI and MNDWI indices are calculated using the following equations (Eq.1-3). where

RED= pixel values of the red band GREEN= pixel values of the green band NIR= pixel values of the near infrared band SWIR= pixel values of the shortwave infrared band
Slope data is derived from ASTER GDEM data using the SLP module of the PCI Geomatics image-processing program. Distance to roads data was created in ArcGIS program using Euclidean Distance method. The population data were downloaded from the official website of Turkish Statistical Institute (Turkish Statistical Institute Central Distribution System, 2020). The population projections for the years 2030, 2045 and 2060 were estimated with the help of population growth rates data. Forest, agricultural land and natural protection data were provided from Antalya Metropolitan Municipality.

Spatiotemporal Land Use Detection using Image Classification
After the preparation of additional bands, Landsat images were classified using the RF classifier and LULC maps were created to understand the 1995-2015 LULC pattern change and to use them in the simulation process. To increase the classification accuracy, ASTER GDEM and NTL data were used as additional bands during the classification process. DMSP/OLS data is available until 2013. For this reason, DMSP/OLS data were used as an additional band in 1995 to 2010 classifications processes, and VIIRS DNB data was used as an additional band in 2015 classification process. The RF classifier was suggested by Breiman (2001) and it is a machine-learning algorithm. In this study, the RF machine-learning method was applied using Image RF IDL based tool (Aslan and Koc-San, 2015, Koc-San, 2013a, 2013bWaske et al., 2012). The RF IDL based tool is freely available and license and platform independent (Waske et al., 2012). For classification five LULC classes were determined: urban, vegetation, water, agriculture and other (Figure 3). In the classification process, 100 pixels per class were collected for training, while 500 pixels per class were collected for testing.

Land Use Simulation using Urban Growth Models
Different methods can be used to predict urban growth. Considering its structure, simplicity and possibility of evolution, the CA model is one of the most powerful models to simulate urban growth (Aburas et al., 2016). In this study, urban growth simulations were produced using the open source GIS software QGIS with plugin MOLUSCE, which use CA urban growth prediction.
The plugin contains different submodules: (i) Input Submodule can use different biophysical and socioeconomic driving factor data such as temporal land use maps, road network, rivers, topography, population.
(i) With the Area Change Submodule, land use changes can be calculated and LU change transition matrices and land use change maps can be created.
(iii) Sample Data Submodule randomly selects user-defined sample points to be used for model calibration and validation. It supports four modeling methods, which are ANN, LR, MCA and WoE modeling methods.
(iv) The Simulation Sub-Module calculates calibration statistics and creates transition potential maps and simulates based on the Monte Carlo CA model approach.
(v) The Validation Submodule calculates the kappa statistics (standard kappa, kappa histogram and kappa position) that will be used to validate the land use maps created by simulation. (Gismondi, 2019).
In this study, both ANN and LR methods were used.
Simulations are based on the produced LULC maps. However, the factors such as population, slope, roads, distance to the center also affect the urban development. Therefore, these factors should be included in the simulation. In this study, slope, forest, natural protection areas, distance to roads and agricultural areas layers were used for simulation processes.

RESULTS
The

CONLUSION
In this study, LULC maps were created by 5-year intervals from 1995 to 2015, to analyse the land use change and urban development, and then the future land use patterns and urban growth models for the years: 2030, 2045 and 2060 were predicted. Antalya, which is one of the most prominent cities of Turkey with its touristic and agricultural properties, was selected as study area. According to the obtained results, RF classification accuracies were computed as higher than 84%. In addition, the general kappa values indicate that simulations were quite successful with the values over than 80%. When the study results were analyzed, it is predicted that the urban areas, which have doubled from 49.56 km 2 to 96.25 km 2 in 2015, will continue to expand and will be 156.85 km 2 in 2060. In other words, the urban area of Antalya has almost doubled in 20 years  and according to the simulation results the urban expansion is expected to continue increasing in the future. For urban planning and sustainable development, revealing LULC changes and urban expansion trends using urban growth models and conserving protected areas, green areas and agricultural areas with correct plan decisions are important. The results of this study can be evaluated and considered by urban planners and decision makers.