Data sources
The data required for the study comprises the satellite data, toposheets, high-resolution Google earth images, City and taluk maps, transport network map, GPS survey points and ancillary non-spatial data, which were collected from various departmental offices and online websites. The satellite images used in the study for the year 2000, 2006, 2011, 2016 and 2020 were collected from multiple satellites (Table 1) due to the lack of temporal continuity of any single satellite mission imageries. The Landsat imageries used in the study were downloaded from the USGS earth explorer (https://earthexplorer.usgs.gov/). The LISS-III data of the year 2011 was obtained from the Bhuvan portal of ISRO (https://bhuvan.nrsc.gov.in/). The satellite data were collected at an approximate 5-year interval for the period from 2000 to 2020. The images were collected based on (i) availability, and (ii) cloud-free condition. Toposheets and high-resolution images from Google Earth were used as a reference for the image classification and accuracy assessment.
Table 1 Details of the satellite imageries used in the study
Mission - Sensor
|
Acquisition Date
|
Path/Row
|
Spatial Resolution
|
Landsat 7 - ETM +
|
20-12-2000
|
145/051
|
30 m
|
ResourceSat-1 / IRS-P6 - LISS IV
|
28-06-2006
|
(098/064, 097/064)
|
5.8 m
|
ResourceSat-1 / IRS-P6 - LISS III
|
18-11-2011, 23-11-2011
|
(098/064, 097/064)
|
23.5 m
|
Landsat 8 - OLI TIRS
|
15-02-16
|
145/051
|
30 m
|
Sentinel-2 - MSI
|
07-01-20
|
T43PDQ
|
10 m
|
The SRTM Digital Elevation Model (DEM) for elevation data has been downloaded from the USGS earth explorer. Similarly, the transport network of the study area for the different year has been downloaded from the open street map website and extracted from onscreen digitisation from toposheets. The processes in the study were performed using multiple software packages. The QGIS and ERDAS were used in pre-processing of Landsat data and classification of the satellite imageries. ArcGIS and QGIS software were used for the analysis of spatial data. The prediction was executed in the SLEUTH urban growth model in the Cygwin environment. SLEUTH Model (http://www.ncgia.ucsb.edu/projects/gig/index.html) (SLEUTH3.0_beta) is an open-source modelling package developed by Dr Keith C. Clarke under the Project Gigalopolis, USGS and is based on LINUX operating system. In the current study, the model was implemented using the Cygwin interface in the Windows environment.
Image processing, classification and accuracy assessment
The pre-processing of satellite imageries removes redundancy and inaccurate data while making it suitable for operational use (Jensen, 2005; Narumalani and Merani, 2016). In remote sensing, reflected energy received at the sensors can be different from the actual radiance reflected from the object due to atmospheric scattering, refraction and absorption. To get correct ground radiance, the radiometric error must be removed by converting digital number (DN) values to radiance, then converting at-sensor radiance to the top of atmosphere (ToA) reflectance, and then converting the ToA reflectance to surface reflectance (Hall, Strebel, Nickeson, and Goetz, 1991). The radiometric correction was performed using the Semi-Automatic Classification Plugin (SCP) of QGIS (Congedo, 2016; Leroux, Congedo, Bellón, Gaetano, and Bégué, 2018). Similarly, Resourcesat-1 images were subjected to geometric correction and reprojection by ground-based registration to remove distortions in the image (NRSA, 2004; Dave, Joshi, and Srivastava, 2015). Then the multispectral composites of satellite imageries were created by stacking the respective bands of Landsat, Resourcesat-1 and Sentinel - 2 images of the study area.
The classification is performed using the reference spectra or signatures derived from the images in the supervised classification technique. The signatures were used to classify using the maximum likelihood classifier algorithm, which is one of the most commonly used algorithms of supervised classification (Strahler, 1980). The LULC classification scheme included six LULC categories viz. (i) urban / built-up area; (ii) water bodies; (iii) vacant land / bare soil; (iv) cropland; (v) plantation; and (vi) natural vegetation. Since classification algorithms do not produce perfect classification results, post-classification measures are applied to minimize the errors using the pixels recording function in ERDAS IMAGINE. Afterwards, the accuracy assessment was carried out to estimate how well the classified LULC image classes are identified corresponding to its reference image. The most popular way to represent the classification accuracy of the remote sensing data is through an error matrix (Congalton, 1991). The accuracy is described by Users' accuracy (error of commission), Producers’ accuracy (error of omission), overall accuracy and kappa statistics (Congalton and Green, 2008). The raster images of LULC were converted into polygon to compute area statistics. Then using ArcGIS, urban polygon areas less than 4 Ha and non-urban polygon of less than 9 Ha was removed in the ‘Eliminate’ tool. As the usefulness of any classified map ultimately dependents on the production of output maps, tables and geospatial data (Lillesand, Kiefer, and Chipman, 2015), finally, maps and tables have been produced to show Spatio-temporal changes in land use and urban growth in Mangaluru during 2000-2020.
SLEUTH model description
CA-based SLEUTH urban growth model is coupled with the Clark Urban Growth Model (UGM) and the Land Cover Deltatron (LCD) sub-models (Clarke K. C., 2008; Jat, Choudhary, and Saxena, 2017). SLEUTH requires input layers of the slope, land-use, urban (seed layer), transportation and hill shade. The urban growth in the model is influenced by four sub-steps or growth rules determining different growth forms such as, spontaneous, new spreading centre (diffusive), edge (organic), and road influenced (Clarke, Hoppen, and Gaydos, 1997; Mahiny and Clarke, 2012). These growth rules are in turn determined by five parameters namely; diffusion, bread, spread, slope resistance and road gravity with a coefficient value ranging between 0-100 (Jantz, Goetz, and Shelley, 2003; Bajracharya, Lippitt, and Sultana, 2020). Diffusion coefficient determines the possibility of random selection of pixcels for new urban development, the bread index determines the growth probabilities in isolated pixel and its potentiality to develop new urban centre, the spread index influences edge growth of pixels, slope limits growth on steeper slope and road gravity index determines urban growth along the roads (Dietzel and Clarke, 2007). Table 2 describes the relationship between the growth types and growth coefficients simulating urban growth. The cell state (urban) in the CA environment is determined by the growth rules.
The SLEUTH is developed on a series of rules in a nested loop iteration in which brute force calibration using Monte Carlo simulation is used to produce growth parameters coefficient ranging in values between 0-100 for each controlling parameters in coarse, fine and final calibration phases to finally predict urban growth (Liu, Sun, Yang, Su, and Qi, 2012; Abedini and Azizi, 2016). Further, the behaviour of the model is determined by the user-provided excluded layers and slope gradient, in which locations having slope gradient higher than 21% cannot be converted to urban (Clarke, Hoppen, and Gaydos, 1997; Silva and Clarke, 2002). Finally, the model is controlled by self-modifying growth rules initiated by the critical-low or critical-high growth rate of the model which produces an S-shaped growth curve and prevents the model from simulating linear urban growth (Silva and Clarke, 2002; Dadashpoor and Nateghi, 2017). The Leesalee metric generated during the final calibration of the model is used to choose the best-fit control parameters that best captures the pattern of urban changes and is used to predict urban growth (Chaudhuri and Clarke, 2019).
Table 2 Description of the relationship between growth types and growth coefficients for simulating urban growth
Growth rule
|
Growth coefficient
|
Description
|
Spontaneous growth
|
Diffusion, slope
|
Simulates random urban growth
|
New spreading centre (diffusive)
|
Bread, slope
|
New diffusion Centre
|
Edge (organic)
|
Spread, slope
|
Simulates edge growth of new or old urban centres
|
Road influenced growth
|
Bread, road gravity, spread, slope
|
Simulates new growth along roads
|
Model inputs
The SLEUTH model requires minimum spatial datasets as input layers in raster format such as slope, land-use, exclusion, urban, transportation and hillshade (Fig. 2). All the input data required for the model has been arranged using ARCMAP 10.3 software. The slope and hillshade layers were derived from SRTM DEM 30 m elevation data in ERDAS IMAGINE. The slope was calculated in percentage and the hillshade map of the study area was also prepared to be used as a background layer of the urban growth prediction map (Dietzel and Clarke, 2007). The land-use layer of the study area for the model is prepared using satellite imageries for the year 2000, 2011 and 2016. The exclusion layer was used to define areas not suitable for urban growth. The model requires a minimum of one excluded area layer, the excluded area layer was created by compiling all the water bodies Shapefile derived from the classified land use land cover map of the study area for 2000, 2011 and 2016 as there is less probability of water bodies converted for built-up purposes in future. The SLEUTH model requires a minimum of four historical urban seed layers (Liu, et al., 2019), the urban layers were extracted from satellite image classification as discussed in the image classification section, for the year 2000, 2006, 2011 and 2016. The urban extent layer is a critical input to the model which will be calibrated in phases for growth prediction (Dietzel and Clarke, 2007). Lastly, the transportation map was prepared for the year 2000, 2011 and 2016, which are digitised from toposheets and also downloaded from online sources (OpenStreetMap). The roads are categorised as national and state highways, then again as primary, secondary and tertiary roads, and residential roads. The entire layers used in the model were converted into Graphics Interchange Format (GIF) raster format with unsigned 8-bit pixel depth to use in the model. These layers were prepared with 30 meters and 100 meters spatial resolution to be used during coarse, fine and final phases of calibration. All the input layers were prepared with uniform arrays with the dimension of 30 m resolution data is 854×1388, whereas that of 100 m resolution data was 256×416. Then the implementation of the SLEUTH model follows three phases viz. test, calibration and prediction (Clarke, Hoppen, and Gaydos, 1997). The following sections briefly describe the calibration process implemented for urban growth prediction in Mangaluru.
Model calibration
The model calibration begins with the test. The test mode checks the reaction of the model to the input datasets by loading the input file directory. Once the test step is executed successfully the calibration procedure was started. The calibration process adjusts the modelled data with the historical input datasets. The calibration is done in three scenarios; coarse, fine and final, which produce the best coefficients (diffusion, bread, spread, and slope resistance road gravity) which is used by the model to determine the growth rule to effectively simulate the urban growth (Silva and Clarke, 2002). The calibration of growth coefficients is done using the brute force calibration technique, which produces output for every possible combination of coefficient values (Bihamta, Soffianian, Fakheran, and Gholamalifard, 2015; Bajracharya, Lippitt, and Sultana, 2020). The calibration process makes the model to get adapted to the local settings of the study area (Clarke, Hoppen, and Gaydos, 1996). In the coarse calibration, the scenario file was created with a wide range of parameter values of START-STEP-STOP (0-20-100) (Table 3), and the start date and end date of the calibration were set at 2000 and 2016 respectively. The output from the coarse calibration was assessed from the control statistics file to narrow down the best fit values corresponding to the top ten Leesalee indexes for the next phase of the calibration. The Leesalee index measures spatial fit between the historical urban growths with the modelled growth (Silva and Clarke, 2002). In the fine calibration, the range of parameters was narrowed down and Monte Carlo iterations were set according to the input raster resolution. The final calibration output from the previous calibration was used to further narrow down the values and was executed with 50 Monte Carlo iterations. The fine calibration provides the best fit coefficients from running the model on the historical urban growth to simulate the future urban growth (Rafiee, Mahiny, Khorasani, Darvishsefat, and Danekar, 2009). The best-fit prediction parameters for the end year (2016) urban layer were derived from the final calibration is used to forecast urban area in Mangaluru for the year 2031.
Prediction
For the urban growth prediction the best-fit prediction coefficients was used to execute prediction mode in SLEUTH for Mangaluru (Fig. 3). The prediction was executed using 100 m resolution raster input layers with 1000 Monte Carlo iterations. The predicted output GIF images were converted to .tiff image to use in the GIS environment for further analysis of the forecasted urban growth results.
Urban growth prediction validation
To confirm the accuracy of modelled urban growth in SLEUTH for Mangaluru, model validation was performed. In the current study, the model validation was performed by employing statistical measurements and visual inspection techniques. The Kappa index of agreement and correlation coefficient of observed and predicted urban growth pixels were computed to statistically measure the validity of the model (Sakieh, et al., 2015; Ilyassov, Kantakumar, and Boyd, 2021). The Kappa statistics is a measure of agreement between the actual urban area and the modelled urban area (Congalton and Green, 2008; Foody, 2002). The statistic is computed for pixels in the two maps as a percentage of agreement as;
(1)
Where, Po is the proportion of the observed urban pixel and Pc is the expected proportion of correct urban pixels (Chaudhuri and Clarke, 2014). A Kappa value of 1 signifies that there is a perfect agreement in the data. However, the Kappa agreement is not the perfect method measurement of agreement (Foody, 2002). Then, the correlation coefficient between the historical observed and modeled urban pixels has been worked out for the year 2006, 2011, 2016 and 2020 (Sandamali, Kantakumar, and Sivanantharajah, 2018). Further, the visual association between the modeled and actual urban area was performed. For comparing the predicted growth with the observed growth of urban area classified LULC image of the year 2020 was used. The visual comparison of spatial association between the predicted and actual urban growth for the year 2020 was compared, in which areas of association, areas of over-predicted and under-predicted growth were spatially identified.
Landscape metrics
To quantify the Spatio-temporal dynamics of urban growth patterns spatial metrics have been used for both historical and predicted urban growth. The spatial metrics quantify spatial dynamics of urban patches and can be used as good indicators of urban form and morphology, planning scenario, ecology and socio-economic aspects (Herold, Goldstein, and Clarke, 2003; Alberti and Marzluff, 2004; Botequilha Leitão, Miller, Ahern, and McGarigal, 2006). The metrics quantify the physical dimensions of urban patches such as their shape, area, size, pattern and distances between the patches. The spatial metrics allows analysis of urban growth pattern. In the current study seven spatial indices measuring various aspects of urban patches were selected (Table 3). The urban patches for the year 2000, 2006, 2011, 2016 and 2020 of Mangaluru was derived from the classified maps by re-coding LULC classes into the urban and non-urban area and was compiled as binary maps in ArcMap. Later the same procedure was followed for the simulated 2031 urban map of Mangaluru. The metrics were calculated using the Fragstats 4.2 software (McGarigal, 2014).