Digital change detection and expansion monitoring of urban areas using satellite images by means of classification methods Meysam Argany Remote Sensing Division Surveying and Geomatics Engineering Department Faculty of Engineering University of Tehran, Tehran, Iran margani@ut.ac.ir M. Reza Saradjian Remote Sensing Division Surveying and Geomatics Engineering Department Faculty of Engineering University of Tehran, Tehran, Iran ABSTRACT The present study indicates that remote sensing data may be used as a valuable tool to monitor urban expansion in some small towns turning to major cities. The highly detailed land-use classification of the agglomeration in Karaj region located at the west and nearby Tehran for instance, and its change extraction can be used as appropriate data input for preparing the necessary development and management plans. The increase in residential area or overall expansion of built-up land in the region and the according statistics extracted from this study may be considered as an alarm for the problem of increase in population, pollution and exploitation of resources in Karaj city. Based on this preliminary case, the expansion of the region has been studied for period of years. The method used is based on multi-temporal and multi-spectral satellite data. In this study, the possibility of using Landsat satellite images for documenting the expansion of built-up areas has been examined. At the initial stage, two TM images with 13 years time interval have been preprocessed. In the next stage, Maximum Likelihood Classification (MLC) technique has been applied on the images. The resulting two classification maps have been compared by applying post-classification comparison change detection technique in which its relevant change detection matrix has been used. Due to the six main classes identified, the relevant change detection matrix contains 30 off-diagonal components with corresponding brightness values. By selection of each component, one of the from-to classes of change information is provided. According to the study, it is concluded that considerable amount of from-to classes may be produced from which useful information necessary for management and development of the city may be extracted. Since the study has demonstrated potential as a means to detect, identify, map and monitor the changes, it is also concluded that the expansion of the city is about 12 percent compared to the first date image; and that the city expansion is mainly due to the lands changed from bare land or vegetation around the city to built-up areas. 1 INTRODUCTION It is important that dynamic land biophysical materials and man-made features that change rapidly over time, be monitored so that one can develop the capability to predict future change and to differentiate between the impact of human activities and natural activities on the environment. The overall aim in built-up area expansion monitoring is to assess the existing state compared to the past, find the change and its change trend, and predict changes in the future before they occur. Then, it would be possible to make sound decisions concerning the management and protection of the cities from over- population, over-pollution and over-exploitation of resources and resource shortages. The satellite image data used in this preliminary change detection and expansion of Karaj are a Landsat-5 TM acquired on 28/06/1987 and a Landsat-7 ETM+ acquired on 25/07/2000. The study area includes Karaj city, its suburbs, and agricultural areas around it. The longitude and latitude of top-left and bottom-down corner coordinates of the images are (50º 54' , 35º 52') and (51º 02' , 35º 46') respectively. In order to monitor the land-use and land-cover changes in Karaj and estimate the built-up area expansion, it is necessary to classify the TM images first, and then apply a change detection algorithm regarding the classes contributing to the change over time. The change classes mainly include built-up, vegetation, and bare land classes in the scene. The software package used to process images in this study was ENVI (version 3.6) and ILWIS (version 3.1 Academic). 2 CHANGE DETECTION METHODOLOGY The algorithm to be selected dictates both the classification type to be used and wheatear the from-to information is required to be extracted or not. Some methods do not provide quantitative information on the amount of area changed from one land cover class to another, but might help with the selection of one of the quantitative change detection techniques as implemented in this study prior to the main change detection algorithm. In a digital change detection method, accurate geometric correction and classification of each remotely sensed image to produce classification maps are required. These two maps are then compared on a pixel-by-pixel basis using a change detection matrix. Every error in the individual date classification map will also be present in the final change detection map (Rutchey and Velchek, 1994). Therefore, it is imperative that the individual classification maps used in the post-classification change detection method be as accurate as possible (Augenstein et al., 1991). This digital method applied in this study, is called Post-Classification Comparison Change Detection method and is the most commonly used quantitative method of change detection (Jensen et al., 1993a). 3 IMPLEMENTATION Two TM and ETM+ images with 13 years time interval have been geometrically corrected using a first order polynomial transformation. The second image from year 2000 was geometrically corrected when submitted to be applied in this study. The first image taken in year 1987, has been registered based on the second image using selection of GCPs among both images. The nearest-neighbor resampling has been applied so that the pixel values stay unchanged for more classification accuracy. The images used in this study using this method are shown in Figure 1.  Figure 1: Geometrically corrected TM data of study area obtained in 1987 (left), and 2000 (right) 3.1 CLASSIFICATION To identify existing classes in the TM and ETM+ images, an unsupervised (i.e. K-means) classification technique has been applied first. The number of classes input to the algorithm at this stage was specified 17. Each similar groups of the classes found in this stage were then merged into main classes and therefore reduced the number of classes to 6. In the next stage, in order to increase the accuracy of the classification results, Maximum Likelihood Classification (MLC) technique has been applied to the images. Training site selections was made based on the classes found in the previous stage, and then the MLC was applied based on the statistics calculated for the training sites. The percentage of correctly classified pixels of the training sites together with error matrix has been used to measure the accuracy of classification maps. An error matrix produced for each classification map using training versus test reference information and considerations made on the total number of samples collected for each category and sampling scheme, also appropriate descriptive and multivariate statistics applied resulted in 82% overall accuracy for TM (1987) and 89% for ETM+ (2000). The final classification maps containing six main land-cover classes are shown in Figure 2. The classes are residential, industrial and commercial, bare land, vegetation, cultivation and gravel.  Figure 2: Classification of TM and ETM+ data of study area obtained in 1987 (left), and 2000 (right). 3.2 CHANGE DETECTION Post-Classification Comparison Change Detection technique has been applied in the stage of change detection. In this case, two images were preprocessed and geometrically corrected, then corrected images from date 1 and date 2 were classified and classification map of date 1 and date 2 were extracted. Implementing change detection was resulted in change map. Comparison of both classification maps on a pixel-by-pixel basis, resulted in creation of a change image map consisting of unique brightness values. In order to create the change map, each one of six classes was coded with its name. Therefore, the classes were coded as gravel, vegetation, industrial and commercial, bare land, cultivation and residential. The alteration of any of these names from one to other would produce a unique element of from-to classes. In order to color the change classes, all 30 possible colors (i.e. off-diagonal from-to land-cover change classes) have been used. By assigning a unique color look-up table value for each change class, all pixels that changed from 1987 to 2000 were uniquely colored and a from-to matrix with the size of 6×6 was created. In order to monitor the change both quantitatively and qualitatively, the growth rate of the urban area and the speed of change in the form of colors to indicate increment or decrement of classes has been established. In this case the easiest way to calculate the mean was used and population growth percentages per year were calculated. By selection of specific from-to classes for a particular change, a specific change from one class to other one is appeared in the change map. Only 30 possible selections of off-diagonal from-to land-cover change classes may be selected to produce the change detection map. Then, an image of particular changes based on built-up expansion was produced. To analyze the land use changes 1987-2000 attribute tables were created that provided information on the land use assigned to each pixel in the two years 1987 and 2000. The final step was visualizing the urban growth. The number of land use changes, as calculated in from-to matrix, is very high. For a good visual impression of the main trends of the land use changes it was necessary to group the land uses into a limited number of classes. In this case the classified maps of the base years 1987 and 2000 were reclassified from six into three classes. Using only three classes also made it possible to use easy to reproduce (black-and-white) maps using gray tones or hatching patterns. The new class urban consists of residential, industrial and commercial, and the new class non-urban consists of bare land, cultivation and vegetation. 4 RESULTS The final classification maps presented in Figure 2 consists of six different classes including two types of urban classes (i.e. Residential, Industrial and Commercial), and non-urban classes (i.e. Bare land, Vegetation, and Gravel. Then, particular changes based on the objective of this study that indicates study on built-up expansion was produced. Using the reclassification the land uses were grouped into a limited number of classes, Urban and Non-urban. Figure 3 shows an overlay of the two change classes.  Figure 3: Urban growth map. Table 1 indicates the change over land in Karaj in 1987 and 2000 (presented in ha and %). Based on this preliminary study, the residential expansion is about 10.48%, the industrial and commercial expansion is about 4.88%, and consequently the overall urban area expansion is 15.72%. As a result of the huge urban expansion in a relatively short period of time, has caused reductions on natural covers as follows: 4.13% reduction on the bare land, 1.73% reduction on the cultivation and 3.26% reduction on the vegetation (Table 3). The percentage of bare land changed to residential is about 10.14% and is mainly due to construction of new residential and townships around the city and along the roads connecting to the city. Also, a small percentage of bare land has changed to industrial. The change is mainly because of the development of new industrial complexes for moving polluting industry to outside of the city. The percentage of vegetation changed to residential is about 2.16%. Also, a small percentage of vegetation has changed to industrial and commercial. The percentage of cultivation changed to residential is about 0.64% (Table 4). These changes have been combined and shown in the urban growth map (Figure 3).     5 CONCLUSIONS The proposed method shows the applicability of remote sensing technology in Karaj region urban expansion monitoring. The detailed information extracted from change detection can be used as valuable data input for other studies. As discussed, the residential area has increased about 10.48%. The overall expansion of built-up land has been estimated about 15.72%. The statistics extracted from this preliminary study may be considered as an alarm for the problem of increase in population, pollution and exploitation of resources in Karaj. Urban areas which contain high spatial variation in land use forms often have difficulty in distinguishing mixed classes. It is suggested in future studies, by implementation of fuzzy logic classification schemes, the problem of mixed pixels to be alleviated. REFERENCES
|