Analysis of the Land Use and Cover Changes in the Metropolitan Area of Tepic-Xalisco ( 1973 – 2015 ) through Landsat Images

Land use and cover changes (LUCC) have been identified as one of the main causes of biodiversity loss and deforestation in the world. Fundamentally, the urban land use has replaced agricultural and forest cover causing loss of environmental services. Monitoring and quantifying LUCC are essential to achieve a proper land management. The objective of this study was to analyze the LUCC in the metropolitan area of Tepic-Xalisco during the period 1973–2015. To find the best fit and obtain the different land use classes, supervised classification techniques were applied using Maximum Likelihood Classification (MLC), Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs). The results were validated with control points (ground truth) through cross tabulation. The best results were obtained from the SVMs method with kappa indices above 85%. The transition analysis infers that urban land has grown significantly during 42 years, increasing 62 km2 and replacing agricultural areas at a rate of 1.48 km2/year. Forest loss of 5.78 km2 annually was also identified. The results show the different land uses distribution and the dynamics developed in the past. This information may be used to simulate future LUCC and modeling different scenarios.


Introduction
Terrestrial ecosystems are important components of nature since they have biological and functional effects on climate regulation, the hydrologic cycle and as a source of natural resources to satisfy human needs.However, during the last 300 years, the planet has suffered big transformations [1].The ecosystems have been subject to accelerated processes of land use and cover changes (LUCC) [2], which have been identified as one of the main factors contributing to global environmental change [3][4][5], as a result of major current environmental problems [6] such as land loss and degradation, climate change, biodiversity loss, deforestation [7] and ecosystems fragmentation [4,8,9], which in turn cause loss of associated environmental services [10] to such a degree that more than half of the world's forest cover has been lost, and around 30% of these ecosystems face degradation processes.Anthropogenic activities are one of the main elements that contribute to land use changes [11].
Mexico is known as a megadiverse country, consisting of large diversity of organisms, landscapes and terrestrial ecosystems [12].Forests, jungles and other natural vegetation are distributed all over the country [13] covering 74% of the national territory [14], about 146 million hectares [12].The distribution of natural vegetation has been studied for monitoring LUCC, but most of the research conducted at a regional scale has focused on the analysis of losses in natural vegetation and deforestation [15,16].The deforestation rates recorded show a difference from 260,000 ha/year to 775,000 ha/year (i.e., 2600 km 2 /year to 7760 km 2 /year) [15].Since local scale studies on LUCC have been scarce [14], the present work was centered on monitoring LUCC locally, paying particular attention to the urban land use in the metropolitan area of Tepic-Xalisco as a starting point for further research on urban growth simulation and future scenarios design.
The advance in using Geographic Information Systems (GIS) and remote sensing techniques has proven to be very useful to get accurate and coherent information according to the spatial reality [16]; these tools are widely used to analyze the distribution, patterns and trends of the LUCC processes via different methods to obtain several land use classes in the territory, as well as diverse approaches to detect temporary differences, such as the traditional method of cross tabulation [17,18].
In this context, the objective of this study was to analyze the dynamics on urban LUCC at local scale.The methodology was developed through the analysis and processing of four Landsat satellite images corresponding to the years 1973, 1985, 2000 and 2015.To find the best fit and obtain the different land use classes, three supervised classification methods were applied: Maximum Likelihood Classification (MLC), Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs).The results were validated with control points (ground truth).Then, to identify the significant transitions between different land uses-especially in the urban land use changes-losses, gains, changes and interchanges were obtained through the cross-tabulation matrix and according to the methodology of Pontius [19].
The remainder of the paper is organized into four sections.Section 2 describes the study area and the satellite images that were used to obtain different classes of cover and land use.Section 3 outlins the procedure that was followed to classify the satellite images using two supervised classification methods and LUCC analysis.Section 4 describes and discusses the results obtained with respect to other similar works.Finally, the conclusions are presented in the Section 5.

Study Area
The metropolitan area of Tepic-Xalisco is located in the central part of the state of Nayarit (Mexico), as presented in Figure 1.The study area comprises two of the main localities of the state that are linked by commercial and administrative activities through the Tepic-Xalisco highway, which in turn provoked a conurbation process that was formalized as metropolitan area in 2006 by the National Institute of Statistics and Geography (INEGI, for its abbreviation in Spanish), the National Population Council (CONAPO, for its abbreviation in Spanish) and the Ministry of Social Development (SEDESOL, for its abbreviation in Spanish).
The study area was delimited through a 900 km 2 quadrant (30 km × 30 km polygon) including the metropolitan area, a polygon wide enough to locally observe and analyze the processes of LUCC during a 42 year-period.The surrounding zone of the study area contains a diversity of land use where are predominant intensive agricultural activities-mainly devoted to sugarcane, maize, mango and jicama cropsas well as farming activities.To the West of the study area is located the Natural Protected Area (NPA) Sierra de San Juan with over 50% of cover consisting of pine trees and live oaks forest, a great variety of natural resources that provide diverse environmental services susceptible of exploitation.
The metropolitan area has suffered important changes in land use as a result from urban growth.The forest is endangered by the indiscriminate clearcutting activities and mining performed in the east of the mountain, at the boundaries of the locality of Xalisco.Therefore, a historical analysis is necessary to show the distribution of land use and the change processes during a period of 42 years.Such information may be used to simulate future land use changes, such as the urban growth and modeling several scenarios.

Data
To set a temporary standardized thematic nomenclature for the analysis of the different land use, the four Landsat satellite images described in Table 1 were used.The images were taken from the United States Geological Survey (USGS) official website (http://glovis.usgs.gov).

Methodology
Figure 2 shows the methodological process that was followed to analyze urban land use changes in the metropolitan area during a period of 42 years.First, supervised classification techniques were applied through Maximum Likelihood Classification (MLC), Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs); from the preparation of the images (pre-processing) to the application of three supervised classification methods (processing), and validation of classifications (post-processing).Then, to detect changes between different land uses, the periods 1973-1985, 1985-2000, 2000-2015 and 1973-2015 were analyzed using cross tabulation.Finally, to identify significant changes, transitions analysis was conducted using the method of Pontius [19].All processes were performed using GIS with ENVI 5.3, Arcgis 10.3 and Focus (PCI Geomatics, 2015) applications.

Pre-Processing
The scenes of Landsat images 1, 5, 7 and 8 were pre-processed to classify the different types of land use during the identification of changes in the urban areas for each analyzed period.Prior to images processing, the study area was cropped for each satellite image by means of a layer (30 km × 30 km polygon) wide enough to visualize urban land use changes in the metropolitan area during a 42-year period; the polygon was measured from the center of the urban land of the metropolitan area.
In the region of interest, it was necessary to verify and standardize the pixel size and the dimensions of each image by means of geometric correction in the WGS-1984 reference system, UTM projection in the zone 13 N.The validation of the geometric correction was obtained with the Mean Squared Error (MSE) for the control points, using one of the images as reference and comparing it (in pairs) to get the best geometric adjustment.To standardize a spatial resolution to 30 m, the four Landsat images were standardized by resampling the pixel size, especially the 1 and 2 Landsat images.Spectral bands were selected for each image, in particular the ones with optical spectrum, Near Infrared (NIR) and Short-Wave Infrared (SWIR); the panchromatic bands were omitted due to high atmospheric influence, as well as the thermal bands, especially in the 7 and 8 Landsat images.To easily visualize the several classes of covers and land uses, a RGB composition in false color was used, highlighting the strong green areas with the NIR, SWIR and R combination; in addition, 2% of linear contrast highlighting was applied to enhance the visualization and identification of the training sites.
The conversion to reflectance was performed to obtain the terrestrial area spectral reflectance values for the different covers and land uses, giving the spectral value to each pixel.The conversion to reflectance was conducted considering the method of Chavez [20] through the following equation (Equation ( 1)).
where ρ k is the reflectance for the k band; d is the factor that considers the solar variation from the Earth-Sun distance, calculated from the Julian Day; ∂ 1,k is the conversion to radiance multiplicative coefficient; θ e is the solar elevation angle; and E 0,k is the solar irradiance in the top of atmosphere for the k band.The data to make the conversion to reflectance were obtained from the header files of each satellite image.

Processing
For processing the images, spectral signatures were created from selecting training sites based on the identification of similar areas in different covers and land uses, combining the knowledge of the area for a proper selection of the regions of interest (ROI).To identify the different land use classes, some visual patterns such as tone, texture and the influence areas were used.
During the identification of the training sites, the separability of the spectral signatures was verified for the five land use classes described in Table 2.While selecting the training sites, control fields verified in situ were also set for validation of each classified image (post-processing).
To obtain the different land use classes, three supervised classification methods were used.First, the Maximum Likelihood Classification (MLC) method, as the most widely used in the scientific literature, is fast, easy to apply and enables a clear interpretation of the results [21].This algorithm can obtain a spectral image of each land use class through variance and covariance statistics of the set of training sites identified in the image and calculates the probability of belonging to each class according to the spectral signature; this method has been proven in works such as those of [22][23][24][25], with satisfactory results.
The Support Vector Machines (SVMs) algorithm was the second method applied.This automatic learning algorithm trains linear and non-linear learning functions by transforming the original data into a different space with a function (kernel) to obtain the hyperplane which maximizes the margin of separation between two or more classes to be classified [26].Currently, the SVMs algorithm is among the most reliable methods; therefore, it is used in many works [27][28][29] with satisfactory results.For the classification of images the Radial Basis Function (RBF) for not-linearly separable data was used.
Finally, the third method applied was the Artificial Neural Networks (ANNs), an automatic learning method that predicts a complex behavior from a sample of observed inputs and outputs.The network structure is based on a simplified model of the human brain consisting of three layers: input, hidden and output.This structure is trained to recognize the result from input values and classify the rest according to the given rules [30,31].Neural networks have been applied to classify satellite images with good results [32,33].
The ANNs classification was applied with a hidden layer of standard backpropagation for supervised learning by means of the logistic activation function for non-linear classification.

Post-Processing
To obtain a better representation of the land use mapping, each classified image was subject to a series of auxiliary processes: a process of majority filtering (3 × 3 pixels) and a method of generalization of polygons less than one hectare-as they are few representatives with respect to the minimum mapping unit-were applied, which reduced the image noise and eliminated the isolated polygons, resulting in the land use mapping for each year of analysis.Finally, to standardize and confirm the location of urban areas, a visual inspection of the mapped urban localities and the population census with historical data from INEGI for the same analyzed periods was carried out.
To validate the obtained results, the classified images were compared against the control fields through a cross-tabulation matrix for the different dates.At the same time the following were obtained: the kappa index, which shows the degree of similarity between a set of control fields and the classified image; the general accuracy, which indicates the percentage of pixels properly classified; the percentage of producer's accuracy, which sets the percentage of a kind of particular land use change correctly classified in the image; and the percentage of user's accuracy, which provides the percentage of a land use class in the image that matches with the class that corresponds in the land.
The model validated as the one with higher accuracy was used to represent the cover and land use mapping for the years 1973, 1985, 2000 and 2015.

Analysis of Land Use and Cover Changes
The analysis of LUCC was conducted through the cross-tabulation matrix to obtain losses, gains and interchanges between the different covers and land uses and after that, the significant transitions analysis was conducted according to the methodology proposed by [19].For each cover and land use the cross-tabulation matrix enables to obtain, through the diagonal values, the stable areas between two dates, as well as the losses (below the diagonal) and the gains (above the diagonal).Such significant transitions on each cover and land use were obtained by comparing the gains and/or real losses against the gains and/or expected losses randomly, divided by the gains and/or expected losses.This comparison gave as a result the transition index where the values less than one indicated non-significant changes among covers and land uses, while the positive values more than one, indicated significant transitions.

Satellite Images Classification
Figure 3 shows the mapping of covers and land uses obtained for each classified Landsat image.The urban land use has mainly replaced agricultural land use where the productive activities has been focused on sugarcane, maize, mango and jicama crops at rate of 1.48 km 2 /year.

Classification Validation
Table 3 shows the results from the process of classified images validation through the cross-tabulation matrix and the parameters obtained: kappa index, general accuracy, producer's accuracy and user's accuracy.Validation statistics show better results when using the SVMs classification method, from which general accuracy above 85% is recorded for the four classified Landsat images 1, 5, 7 and 8, unlike the maximum likelihood classification and the artificial neural networks methods.When selecting training fields, separability problems were identified between the secondary vegetation and the agricultural land use classes, which is reflected in the producer's accuracy percentage with values below 90%.The best fit was recorded when using the SVMs classification method, where the accuracy percentages average 96% for the user, and 95% for the producer; as shown in the Table 4.

Analysis of Land Use Changes
Table 5 shows the changes in area for each class of cover and land use occurred during 1973-2015.The urban land use has increased 7% from the overall analyzed area with an annual rate of 1.48 km 2 /year.At the same time, the forest cover has lost 28% of area in 42 years, with an annual rate of 5.78 km 2 /year.The annual rates of change for each class of cover and land use are shown in Figure 4. Differences were identified depending on the classification methods applied.When applying the SVMs and ANNs methods, the urban land use presented a similar trend of 1.48 km 2 /year of change.The three methods recorded differences regarding agricultural land use, where the ANNs method registered a rate of 4.55 km 2 /year, while the SVMs method obtained rates of 1.43 km 2 /year.The secondary vegetation cover also presented important differences; while the SVMs method registered a rate of 2.86 km 2 /year, the results of the MLC and ANNs methods presented a rate below 0.13 and 0.08 km 2 /year, respectively.With regard to the forest, the methods SVMs and ANNs obtained similar annual rates of 5.78 and 6.12 km 2 /year; the negative rate of change is due to the loss of surface.Since the SVMs method proved the best fit, the results from applying this method were used to analyze significant transitions and land use changes in the metropolitan area.
The changes occurred in this same period due to losses and gains are shown in Table 6.The main results show that the urban land use in the metropolitan area has increased 62 km 2 since 1973.The agricultural area and the secondary vegetation present most of the interchanges due to crop rotation, which is confirmed when observing the losses and gains of these same land use classes.The transitions analysis for each cover and land use is summarized in Table 7. Significant transitions for the analyzed periods are registered particularly in the agricultural land use change to urban land use and secondary vegetation.The increase of urban area and secondary vegetation is because the agricultural area is being replaced.During a 42-year period, 33.1 km 2 of agricultural use transformed into urban areas and 53.3 km 2 into secondary vegetation.The forest also has been affected with transitions, although non-significant.They recorded ecological changes of 49.1 km 2 to agricultural land and 195.3 km 2 to secondary vegetation.The methodology applied in this study is similar to the one used by Aguayo et al. [1], Lopez and Plata [25] and Antillón et al. [34], who applied the maximum likelihood algorithm to get the different land use classes of the study area and analyzed LUCC through cross-tabulation method or confusion matrix.The results of said investigations have had the same trend as obtained in the present study as regards to urban areas replacing agricultural lands, as well as a decrease in the forest area.
According to the scientific literature, the SVMs method has been used in several studies such as those developed by Mountrakis et al. [27]; Lu et al. [28]; and Xie et al. [29] with good results on image classification, similar to the results obtained in this work where the SVMs method registered the best fit.Some works have compared different supervised classification methods; for example, Pal and Mather [35] used the same three classification methods that were applied in this study (MLC, ANNs and SVMs), obtaining the best fit results through SVMs.On the other hand, Otukei and Blaschke [36] compared three methods of classification, MLC, SVMs and Decision Trees (DT), and affirmed that the best results were obtained when applying SVMs.In addition, Mondal et al. [37] compared the SVMs and MLC methods, and concluded that when preparing land use maps, the SVMs method is more appropriate than the MLC.Finally, in a more recent study, Wu et al. [38] compared the SVMs, ANNs and DT methods; although radial base and polynomial functions were used for SVMs, this method obtained better results with kappa indices (0.72 and 0.79 for each function, respectively).Both the results in the above mentioned research and in this work obtained the best fit applying the SVMs method; therefore, the results from applying this method were used for the LUCC analysis in the study area.
In Mexico, several studies on LUCC have been conducted to achieve a better understanding on the dynamics and processes of land use change (e.g., [2,8,14,24,25,34,[39][40][41][42][43]).These studies have been oriented to identify forest areas loss rates, unlike this research that was focused on analyzing the urban land use during three periods with the intention of identifying the historical dynamics of urban growth to be able to build urban growth simulation models and development of future scenarios.
Particularly, Cano et al. [44] analyzed urban land use in Hidalgo State (Mexico).From satellite images, they identified urban growth of 72.3 km 2 in a 14-year period, equivalent to an annual growth rate of 1.8%.On the other hand, by means of digital and visual techniques classification, Lopez and Plata [25] analyzed LUCC in the metropolitan area of Mexico City regarding urban expansion detecting an urban growth of 202 km 2 equivalent to 16% in a 10-year period.In comparison with the present work, in the metropolitan area of Tepic-Xalisco it was possible to quantify an urban growth of 62 km 2 during a 42-year period, with an annual rate of 1.48 km 2 , which means a relatively low growth regarding to the metropolitan area of Mexico City.
For the study area, the research on cover and land use analysis conducted by Nájera et al. [45] was identified for the Mololoa River watershed, which determined natural vegetation losses of 41.67 ha/year with deforestation rates of 0.1% and urban growth of 74.86 ha/year, which is nearly half that obtained by this study that registered an increase of urban land of 148 ha/year (1.48 km 2 /year).These results may be attributed to the difference in the boundaries set for the study areas, and to the methodology used to obtain the different land use classes.

Conclusions
The validation results from the classifications developed suggest that the SVMs method gives the best fit and offers greater certainty on the distribution and quantification of the different classes of cover and land use obtained.
The urban land use in the metropolitan area of Tepic-Xalisco has experienced an important increase within a period of 42 years, exceeding ten times the urban area recorded in 1973, with a rate of 1.48 km 2 /year.This growth has produced significant changes in land use with transitions towards agricultural and secondary vegetation land use.The forest cover also has been affected, since it has experienced considerable losses of area with transformation trends towards secondary vegetation.In addition, the agricultural land use has been replaced as a result of urban growth.This situation has caused functional implications on ecosystems and to date losses of agricultural productive area are present, as well as deforestation processes.
The applied methodology enabled learning about the historical dynamics and quantifying the LUCC during a 42-year period, identifying the transitions between each land use.This information will help to establish land planning strategies, promote management and develop land use conservation policies.

Figure 1 .
Figure 1.Localization and delimitation of the study area.Source: Own elaboration based on INEGI's data.

Figure 4 .
Figure 4. Annual rates of change according to the classification methods applied.Source.Own elaboration.

Table 1 .
Landsat images used for mapping land uses in the study area.

Table 2 .
Description of identified covers and land uses.

Table 4 .
Percentages of producer's accuracy and user's accuracy.

Table 6 .
Land use losses, gains and interchanges between periods with SVMs.