|
Abstract
Urban Land Cover Change Detection: a case study of Asmara, Eritrea Yikalo H. Araya, Pedro Cabral , Christian Hergarten 1. Introduction The history of urban growth and urbanisation reveals that urban areas belong to the most dynamic land cover types on earth. The trend of urban growth is usually towards the urban-rural fringe where there are less built areas, irrigation and other water management systems. Regardless of the regional economic importance, urban growth, particularly the expansion of residential and commercial land use towards the periphery of urban areas, has an impact on the ecosystem (Yuan et al. 2005). It is evident that such trend of urban growth has an impact on natural resources and on land cover dynamics at large. In Eritrea, land cover in general and urban land cover in particular has experienced a remarkable change in the last two decades. Hence, updated urban land cover mapping and change analysis are particularly useful for land use and environmental management in the country. Information from satellite remote sensing plays a useful role in understanding the nature of changes in land cover/land use and projecting possible future changes (Bauer et al. 2003). Such information is essential for future urban development plans. In spite of these facts, there are no extensive studies carried out in the country to analyse urban land cover changes using multitemporal and multiresolution remote sensing data. This paper presents an object-oriented image classification methodology to detect and analyse multitemporal satellite data. Furthermore, urban growth patterns were analyzed using selected class metrics. Finally, the paper also addresses driving forces for urban land cover change with reference to demographic and related factors. 2. Study area The study area is Asmara, capital of Eritrea, one of the famous historical and cultural cities in Africa (Figure 1). It is located on the highlands of East Africa (15o13´44´´ and 15o25´36´´ North and 38o 44´33´´ and 39o00´53´´ East) at about 90 km inland and with an elevation between 2100 to 2400 meters. The area has a moderate climate with temperature ranging from 0oC (winter night temperature) to +27oC (summer day temperature). The area has experienced remarkable land cover changes due to urban expansion, population pressure and various economic activities spreading particularly since the independence of the country in 1993.  Figure 1: Location of study area 3. Materials and methods 3.1 Data Analysis of the literature reviewed indicated that spectral responses of surface features, date of image acquisition, pre-processing stage and the area of interest are some of the considerations that must be taken into account in selecting images for land cover mapping (Yuan et al. 2005; Ju D and Weng 2007; Jensen and Im J 2007). The data used in this paper comprises one Landsat TM and one ETM+ scene of 1986 (18, April) and 2000 (27, January), respectively. Both images were acquired during the dry season which is generally cloud free for the study area. The resolution of the image is 30 meters. The Landsat images used in this study have been radiometically corrected by the Global Land Cover Facility (GLCF), University of Maryland (GLCF, 2007). Ancillary imagery data used in this study were acquired by IKONOS satellites of year 2000 (7, March) and ASTER (24, May). All data used in this study were projected to the Universal Transverse Mercator (UTM) projection system (zone 37, World Geodetic System 84). 3.2 Land cover classes The land cover classes applied in this study are adopted from the classification schemes used by the Ministry of Agriculture, Eritrea, which is based on the AFRICOVER land cover classification system (AFRICOVER, 2002). For the sake of simplicity, the analysts modified the descriptions of some of the land cover classes considering the land cover diversity of the study area (Table 1).  Table 1: Land cover classes (LCID is land cover ID used by the Ministry) 3.3 Image classification Image classification has been made using a pixel-based approach which included a variety of supervised (e.g. maximum-likelihood, neural networks, parallelepiped, etc), or unsupervised (e.g. clusters) classification methods. Recently, with the advent of very high resolution images (e.g. IKONOS, Quickbird), new object-oriented techniques have been developed in order to overcome the limitations of the pixel-based methods and their incapacity in dealing with texture. The object-oriented approach considers group of pixels and geometric properties of image objects instead of relying only on spectral characteristics of a single pixel. In this article, we adopted an object-oriented approach using a supervised maximum likelihood classifier. Object-oriented classification method The object-oriented method segments the images into homogenous regions based on neighbouring pixels’ spectral and spatial properties. Unlike many image analysis software operating at pixel level, object-oriented tools segment a multispectral image into spatially homogeneous areas. Although segmentation is not a new concept, classification using image segmentation has become increasingly important in recent times (Blaschke 2005). The segmentation process used here is based on region growing, an algorithm merging smallest objects containing single pixels into larger objects based on scale, colour and shape (Im et al. 2007). The segmentation process stops when the smallest growth of an object exceeds a user-defined threshold. The greater the scale parameter, (i.e. similarity unit and pixel area), the larger is the size of the resultant objects. In this study, a similarity unit and pixel area of 12 and 15, were selected, respectively to create the segmented image. In addition, knowledge- and resource-based sample collection was carried out. Image enhancement was performed prior for better discriminating the land cover classes. 200 sample objects were collected using the spectral information of the objects. The Bhattacharya classifier was then applied to produce the land cover map. The classification accuracy of the collected samples into their respective classes was assessed using SPRING’s “sample analysis” tool. Using this tool, classification accuracy is expected to be 100%. Finally, a membership function was used to label the generated classes. The overall classification process was carried out in open source software “SPRING” developed by the National Institute for Space Research (INPE), Brazil. The software is freely available on the internet (INPE, 2002). 3.4 Urban land cover change detection and analysis techniques Change detection analysis in remote sensing employs some detection algorithms to examine and view a set of multispectral images of the same area acquired in different period of time (Singh 1989). The most common technique for detecting changes is the post-classification technique (Bauer et al. 2003). The maps of 1986 2000, which were considered as “earlier image” and “later image”, respectively were put into the post-classification analysis. Both images were initially classified into several thematic classes followed by reclassification into built and non-built areas. A post classification differencing is then obtained by subtracting the corresponding reclassified maps of 2000 from 1986. Patterns of urban expansion or direction of urban development occurs in a variety of ways depending on the physiographical nature of the area and other factors. The rate of urban development also depends on many factors such as distance from roads or market areas, distance from city centres, slope, etc. Urban area (built areas) may occur spatially in isolated or fragmented mode in space. Therefore, it is imperative to quantify and analyze spatial and temporal dynamics of urban land cover and patterns of urban growth. The differences in representation of a space have led to a wide variety of spatial metrics for the description of spatial structure and pattern (Herold et al. 2003). Spatial metrics are algorithms used to quantify spatial characteristics of landscape or classes of patches (McGarigal et al. 2002). They were developed in the late 1980s based on information theory and fractal geometry (Herold et al. 2003). In this paper, spatial metrics are calculated based on thematic maps representing built and non-built spatial patches. The changes in urban landscape (e.g. development of discontinuous urban areas or urban fragmentation) are measured and analyzed using FRAGSTATS public domain software (McGarigal et al. 2002). In this paper, six spatial metrics (class area - CA, Number of patches – NP, Edge Density – ED, Largest Patch Index – LPI, Euclidian Mean Nearest Neighbour Distance – EMN, Area Weighted Mean Patch Fractal Dimension-FRAC_AM) which have already been used in different studies (Herold et al. 2003, Parker et al. 2001) are adopted and applied (Table 2).   Table 2: Class metrics adopted and used in this study (adapted from McGarigal et al. 2002) 3.5 Accuracy assessment The most common approach to assess accuracy of remotely sensed data uses an error matrix and is referred to as confusion matrix being recommended and adopted as the standard reporting convention (Congalton, 1991). In this paper, the assessment was carried out only for the land cover map obtained for 2000 using an IKONOS image acquired in 2000. Unfortunately, there was no ground–truth data available to assess the accuracy of the 1986 map. 500 samples were selected in ArcGIS using a random sample generator tool. The samples were then labelled into their respective classes and a confusion matrix was built. 4. Results 4.1 Urban land cover map and accuracy assessment In order to detect urban land cover changes and quantify the changes effectively, land cover maps of the study area were first derived (Figure 2) and then simplified into two broad classes to examine the spatial extent of built-up areas (Figure 3).  Figure 2 a and b: Urban land cover maps of 1986 and 2000  Figure 3 a and b: Reclassified maps of 1986 and 2000 The classification carried out (for the reclassified map) produced an overall accuracy of. 93.64%. The error matrix also reveals the results of both user's and producer's accuracy. From this point of view, built up areas are classified as non-built areas to some extent. This is probably caused by the season of the image acquisition. During dry seasons it is expected that bare land and harvested land of the non-built areas would have similar spectral responses with built areas. In addition, the kappa coefficient, which expresses the proportionate reduction in error generated by a classification process compared with the error of a completely random classification (Congalton, 1991) was calculated. It is not uncommon that the Kappa coefficient appears to be low, giving the impression that the classification of remote sensing performed better than chance only by K point of proportion (Muzein, 2006). It was calculated to be 0.73 which is considered to be a good result.  Table 3: Error Matrix of the reclassified map of 2000 4.2 Urban land cover change detection and analysis A post-classification analysis procedure was carried out by overlaying the reclassified maps to generate a change map (Figure 4). Being capital of the country, most of the occupancies of the study area belong to urban or built up areas. Preliminary results from the multi-date visual change detection image indicate most of the changes in land cover have occurred in the peripheries of the urban areas in accordance with the land use policies adopted by the government for various development programs. The observed types of change identified in the study were urban expansion or densification. Besides, the results of the image classifications and change map could provide an estimate of the extent; pattern and direction of urban land cover changes in the study area. In spite of this, there has also been conversion of built areas into non-built areas (approximately 150 ha). It is less likely to have such kind of conversion and these are questionable results. These discrepancies or errors might have caused by differences in class definition, spectral responses of some features (e.g. built areas, bare land and harvested agricultural areas) and mapping inconsistencies, smoothing or generalization of the data. To alleviate such discrepancies in the change analysis, a new land cover map of 2000 (with classes of built and non-built) was generated by summing the reclassified land cover maps of 1986 and 2000.  Figure 4: Urbanisation between 1986 and 2000 Based on this time scale series analysis the urban growth has occurred in almost all part of the city, except for the north-western part. This area can be characterized as “Zone of Discard” as opposed to “Zone of Assimilation”, employing concepts from Urban Geography. This area is characterized by steep sided slopes and mountainous area, and hinders expansion of the city into its western part. With improvements of the infrastructures, some of the urban growth is also emerging from small semi-urban areas near Asmara (e.g. Adi-Guadad and Tsaeda-Christian). Table 4 contains summary of statistics from the spatial metrics of the changes. Analysis of the spatial metrics indicated that the urban land cover of Asmara has increased by approximately 1073.46 ha between 1986 and 2000. This reveals that there has been a tremendous growth in the built-up areas and spatial pattern of regional economic developments in the area. With improvements in infrastructures, the rate of urban growth has expanded into the rural-urban fringe. The result also shows an increase in the number of patches by 124.4%. This reflects the development of a number of isolated, fragmented or discontinuous built up features. ED has increased by 70.28% and this proves an increase in urban boundary. Furthermore, the LPI values increased by 70.28% and this could be associated with the contagion of small and isolated patches into the existing largest patch. FRAC_AM of built areas remained unchanged during the period of study. The ENN-MN decreased by 18.86% indicating that the built area patches are getting closer to each other during the studied period.  Table 4: Spatial metrics for land classification for years 1986 and 2000 and variation (%) 5. Conclusions and discussion In this study, an integrated approach of GIS, remote sensing and spatial analysis tools have been applied to detect and analyze the urban land cover change in Asmara, Eritrea. The availability of remote sensing satellite data from different time periods provided valuable information to study the dynamics. Although the purpose of the study was on urban (built-up) areas, land cover maps of the study area were first derived. The two land cover maps which were produced for 1986 and 2000 provided new information on spatial and temporal distributions of built-up (urban areas) in the study area. Based on the analysis of the metrics, it is reasonable to argue that there has been a relative change in the fragmentation and spatial patterns or structures of the built up features. In addition, the changes in urban environment reflect the development of both continuous and discontinuous (fragmented) urban features. The study area has experienced such development in almost all direction except in its western part. The area is part of the eastern escarpment of the country and is characterised by a very steep sided slope. This also hinders the development of the city to its eastern part. References Anderson J, Hardy E, Roach J, Witner R (1976) A land use and land cover classification system for use with remote sensor data. Geological Survey Professional Paper 964 Food and Agricultural organization (FAO) of the United Nation: AFRICOVER (2002) URL: www.africover.org, Accessed 1st November 2007 Bauer M., Yuan F, Saway K (2003) Multi-Temporal Landsat Image Classification and Change Analysis of Land cover in the Twin Cities (Minnesota) Metropolitan Area. Workshop on the Analysis of multi-temporal remote sensing images, Italy. Blaschke T (2005) Towards a framework for change detection based on image objects. Remote Sensing and GIS for Environmental Studies (Gottinger Geographische Abhandlunger vol.113). Cabral P, Geroyannis H, Gilg J, Painho M (2006) Analysis and modeling of land use and land cover change in Sintra-Cascais Area. URL: http://plone.itc.nl/agile/Conference/estoril/papers/06_Pedro%20Cabral.pdf, Accessed 15th October 2007 Civco D, Hurd J, Wilson E, Song M, Zhang Z (2002) A comparison of land use and land cover change detection methods. ASPRS-ACSM Annual conference and FIG-XXII congress Congalton R (1991) A review of assessing the accuracy of classifications of remotely sensed data. Int. Journal of Remote sensing 37:35-46 Elnazir R, Zhi F, Zheng C (2004) Satellite remote sensing for urban growth assessment in Shaozing City, Zhejiag Province. Journal of Zhejiang University Science 5(9):1095-1101 Global Land Cover Facility (GLCF), University of Maryland. (2007) URL: http://glcf.umiacs.umd.edu/index.shtml Herold M, Goldstein N, Clarke K (2003) The spatiotemporal form of urban growth: measurement, analysis and modeling. Remote sensing of Environment 86:286-302 IM J, Jensen J, Tulluis J (2007) Object-based change detection using correlation image analysis and image segmentation. International J. of Remote Sensing: 1-25 Jensen J, Im J (2007) Remote sensing change detection in urban environments: Geo-Spatial Technologies in Urban Environment, 2nd edit. Pages: 7-31 Ju D, Weng Q (2007) A survey of image classification methods and techniques for improving classification performance. Int. Journal of Remote sensing 28(5):823-870 McGarigal K, Cushman S, Neel M, Ene E, (2002) FRAGSTATS: Spatial pattern analysis program for categorical maps. URL: www.umass.edu/landeco/research/fragstats/fragstats.html, Accessed 15th November 2007 Moeller M, Stefanov W, Netzband M, (2004) Characterizing land cover changes in a rapidly growing metropolitan area using long term satellite imagery. ASPRS Annual Conference Proceedings, Denver, Colorado Muzein B (2006) Remote Sensing & GIS for Land Cover/ Land Use Change Detection and Analysis in the Semi-Natural Ecosystems and Agriculture Landscapes of the Central Ethiopian Rift Valley. PhD thesis-Technische Universität Dresden. National institute for Space Research (INPE), Brazil (2002). URL: http://www.dpi.inpe.br/spring/ Parker D, Evans T, Meretsky V (2001) Measuring emergent properties of agent-based land use/land cover models using spatial metrics. Seventh annual conference of the International Society for Computational Economics. Rowlands A, Lucas R (2004) Use of hyberspectral data for supporting the classification of agricultural land and semi-natural vegetation using multi-temporal satellite data. Proceeding of the Airborne imaging spectroscopy Workshop-Bruges. Santos T, Tenedorio J, Encarnacao S, Rocha J (2006) Comparison pixel vs. object-based classifiers for land cover mapping with Envista-Meris Data, 26th EARsel Symposium, Maio, Varsovia. Singh, A (1988) Digital change detection techniques using remotely sensed data. Int. Journal of remote sensing 10(6):989-1003 Tardie P, Congalton 2002 (2002) A change detection analysis using remotely sensed data to assess the progression of development in Essex County, Massachusetts from 1990 to 2001. URL: http://www.unh.edu/natural-resources/pdf/tardie-paper1.pdf Accessed 15th October 2007 Walter, V (2004) Object-based classification of remote sensing data for change detection. Int. Journal of remote sensing 58:225-238. Yuan F, Sawaya K, Loeffelholz B, Bauer, M (2005) Land cover classification and change analysis of the Twin Cities (Minnesota) metropolitan area by multitemporal landsat remote sensing. Int. Journal of remote sensing 98:317-328 |