Characterizing urban sprawl using multi-stage remote sensing images and landscape metrics

Computers, Environment and Urban Systems 30 (2006) 861–879 www.elsevier.com/locate/compenvurbsys

Characterizing urban sprawl using multi-stage remote sensing images and landscape metrics Wei Ji *, Jia Ma, Rima Wahab Twibell, Karen Underhill Laboratory for GIS and Remote Sensing, Department of Geosciences, The University of Missouri-Kansas City, Kansas City, MO 64110, USA Received 2 August 2004; accepted in revised form 31 August 2005

Abstract This study intends to explore the spatial analytical methods to identify both general trends and more subtle patterns of urban land changes. Landsat imagery of metropolitan Kansas City, USA was used to generate time series of land cover data over the past three decades. Based on remotely sensed land cover data, landscape metrics were calculated. Both the remotely sensed data and landscape metrics were used to characterize long-term trends and patterns of urban sprawl. Land cover change analyses at the metropolitan, county, and city levels reveal that over the past three decades the significant increase of built-up land in the study area was mainly at the expense of non-forest vegetation cover. The spatial and temporal heterogeneity of the land cover changes allowed the identification of fast and slow sprawling areas. The landscape metrics were analyzed across jurisdictional levels to understand the effects of the built-up expansion on the forestland and non-forest vegetation cover. The results of the analysis suggest that at the metropolitan level both the areas of non-forest vegetation and the forestland became more fragmented due to development while large forest patches were less affected. Metrics statistics show that this landscape effect occurred moderately at the county level, while it could be only weakly identified at the city level, suggesting a scale effect that the landscape response of urbanization can be better revealed within larger spatial units (e.g., a metropolitan area or a county as compared to a city). The interpretation of the built-up patch density metrics helped identify different stages of urbanization in two major urban sprawl directions of the metropolitan area. Land consumption indices (LCI) were devised to relate the remotely sensed built-up growth to changes in housing and commercial constructions as major driving factors, providing an effective measure to compare and characterize urban sprawl across jurisdictional boundaries and time periods.  2005 Elsevier Ltd. All rights reserved.

*

Corresponding author. Tel.: +1 816 235 2981. E-mail address: [email protected] (W. Ji).

0198-9715/$ - see front matter  2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.compenvurbsys.2005.09.002

862

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

Keywords: Remote sensing; Landscape metrics; Land cover; Urban sprawl; Kansas City

1. Introduction Urban sprawl, a consequence of socioeconomic development under certain circumstances, has increasingly become a major issue facing many metropolitan areas. Although a general consensus regarding the definition and impact of urban sprawl has not been achieved (Johnson, 2001), urban sprawl is often referred to as uncontrolled, scattered suburban development that increases traffic problems, depletes local resources, and destroys open space (Peiser, 2001). It is critically important to properly characterize urban sprawl in order to develop a comprehensive understanding of the causes and effects of urbanization processes. However, due to its association with poorly planned urban land use and economic activity (Pendall, 1999), urban sprawl is often evaluated and characterized exclusively based on major socioeconomic indicators such as population growth, commuting costs, employment shifts, city revenue change, and number of commercial establishments (Brueckner, 2000; Lucy & Phillips, 2001). This approach cannot effectively identify the impacts of urban sprawl in a spatial context. To fill this gap, remote sensing has been used to detect urban land cover changes in relation to urbanization (e.g., Chen, Zeng, & Xie, 2000; Epstein, Payne, & Kramer, 2002; Ji et al., 2001; Lo & Yang, 2002; Ward, Phinn, & Murray, 2000; Yeh & Li, 2001). Remote sensing techniques have advantages in characterizing the spatiotemporal trends of urban sprawl using multi-stage images, providing a basis for projecting future urbanization processes. Such information can support policymaking in urban planning and natural resource conservation. There are some issues facing urban remote sensing research which are addressed in our study: (1) As stated by Herold, Goldstein, and Clarke (2003), many urban remote sensing studies tend to focus on technical issues in data assembly and physical image classification rather than on the use of the mapped by-products in the spatiotemporal analysis of urban regions. In this regard, our study focuses on a comprehensive understanding of regional patterns of urban sprawl by innovatively analyzing the data generated with proved and operational geospatial methods such as conventional supervised classification techniques. (2) It has been a challenge to determine appropriate spatial detection boundaries of remote sensing in order to effectively relate trends and patterns of urban land cover changes to socioeconomic driving factors. In many of the previous studies, urban land cover change analyses were conducted either at the metropolitan scale as a whole, or at the single county or city level (e.g., Masek, Lindsay, & Govard, 2000; Ryznar & Wagner, 2001; Yang, 2002), or within natural landscape boundaries such as those of a watershed or habitat (e.g., Clapham, 2003; Phinn & Stanford, 2001; Wang & Moskovits, 2001). Some urban remote sensing projects generated data and statistics based on jurisdictions, but failed to justify how to effectively link the data to pattern analysis of urbanization across jurisdiction levels. Our study systematically explores and compares trends and patterns of urban sprawl across different levels of jurisdictions (i.e., cities, counties, and the metropolitan area). The adoption of this approach is based on three considerations: First, socioeconomic driving factors in relation to urban planning activities usually are more closely associated with jurisdictions rather than other spatial units like census units or watersheds. Secondly, for a regional study, jurisdictional units are sufficiently large to embrace diverse land cover types in order to effectively reveal trends and patterns of urban growth. This condition

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

863

may not be met with other spatial units like census blocks or tracts. Thirdly, there were insufficient research efforts using remote sensing imagery to identify spatiotemporal patterns of landscape effects of urban growth, especially as comparing these patterns at different jurisdiction scales. Applied to the characterization of urban sprawl in the Kansas City metropolitan area, our study had the following objectives: (1) detecting and comparing the variation of urban sprawl trends across the metropolitan, county, and city scales; (2) identifying spatial spread patterns of built-up land expansion from the urban core and associated landscape effects; and (3) formulating metrics to integrate remotely sensed land changes with relevant socioeconomic data to effectively characterize urban land consumption. To address these objectives, multi-stage remote sensing images and related landscape metrics (see below) are used to generate indicators and measurements of land cover changes. 2. Research issues on urban landscape metrics Landscape metrics (indices) are numeric measurements that quantify spatial patterning of land cover patches, land cover classes, or entire landscape mosaics of a geographic area (McGarigal & Marks, 1995). Its use has become a trend in urban change studies. Herold, Scepan, and Clarke (2002) and Herold et al. (2003) conducted a comprehensive study on measuring urban land cover dynamics using remote sensing and spatial (landscape) metrics that were analyzed and interpreted in conjunction with the spatial modeling of urban growth. Based on a 72-year time series data set compiled from interpreted historical aerial photographs and from IKONOS satellite imagery, their study demonstrated that an approach that combines remote sensing, landscape metrics, and urban modeling analysis may prove a promising new tool for understanding spatiotemporal patterns of urbanization. Other studies also explored various methods for applying landscape metrics in urban land cover research, such as urban gradient analysis (Luck & Wu, 2002) and the assessment of forest fragmentation (Civco, Hurd, Wilson, Arnold, & Prisloe, 2002). Generating landscape metrics from classified remote sensing data raises several fundamental research questions: how to evaluate the effects of spatial scale dependence of quantifying urban landscape heterogeneity with landscape metrics; how to interpret landscape metrics in relation to urban growth patterns; and how to generate effective landscape metrics to relate remotely sensed land changes to driving factors of urban growth. To understand spatial scale dependence of landscape metrics, Wu (2004) systematically studied scaling relations for landscape pattern measurement when comparing 17 class-level landscape metrics for four landscapes (land use and land cover maps), including the Phoenix urban area. The results of his study show two types of responses of landscape metrics to changing scale: simple scale functions (Type I metrics) and unpredictable behavior (Type II metrics). These results with real landscapes are generally supported by investigation with simulated landscapes (Shen, Jenerette, Wu, & Gardner, 2004). These findings may be used for reference in selecting appropriate landscape metrics to effectively characterize urban sprawl (see Section 3.3.2) and can help in understanding the effects of changing spatial resolution of the source data. Effectively interpreting landscape metrics has always been a challenge in many fields. As pointed out by McGarigal and Marks (1995), the authors of FRAGSTATS, a widely used tool for calculating landscape metrics, landscape metrics can easily become ‘‘golden’’ in the hands of less-informed users, in a way that the ‘‘garbage in–garbage out’’ axiom could

864

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

apply. This situation would be particularly problematic in urban sprawl characterization because of the heterogeneity of urban landscapes and the complexity of the factors driving urban growth. Thus, how landscape metrics are to be analyzed has become a research issue that needs to be fully addressed in urban applications. Several approaches may be used, individually or jointly, to effectively interpret the landscape metrics of urban landscapes: (1) analyzing the trends of landscape metrics (e.g., increase or decrease, etc.) as compared with the urban growth patterns identified by other methods like simulation models (e.g., Herold et al., 2003); (2) interpreting landscape patterns according to the computational expression of a particular metrics; in this case, certain landscape indices may show advantages for specific applications because of the simplicity of their computational algorithms (see Section 3.3.2); (3) exploring (geospatial) statistical relations between landscape metrics and the urban landscape patterns. In our study, approaches (2) and (3) are used jointly (see Section 4.2). There is a great interest in devising meaningful measures for urban sprawl. For this purpose, the measures ‘‘density’’, ‘‘urban density’’, ‘‘land use efficiency’’, ‘‘per capita land use consumption’’, etc. have been used (e.g., The Brookings Institution, 2002; US Environmental Protection Agency, 2001; Fulton, Pendall, Nguyen, & Harrison, 2001; Masek et al., 2000; Pendall, 1999; and Sutton, 2003). With individual focuses, many of these measures were devised to reflect the relationship between population change and land conversion to urban uses. In a recent effort, the concept ‘‘housing unit’’ was used as a proxy for population and combined with digital orthophoto data to generate urban sprawl metrics (Hasse & Lathrop, 2003). In this regard, we consider that since it is often difficult to distinguish population change in a given jurisdiction as either the cause or effect of urban development, the population factor should not be used as a sole indicator of urban sprawl. In most cases, an increasing (or diminishing) number of built-up activities like housing and commercial constructions would be more effective to indicate sprawl as consequences of land consumption because usually construction activities, as compared to population change, reflect directly economic opportunities as the major driving force of land alteration (Lambin et al., 2001). Thus, as a pilot method effort, we intend to formulate more effective metrics of urban sprawl by relating remotely sensed land change to construction-based land consumption. 3. Approaches 3.1. Study area The Kansas City metropolitan area is located along the eastern boundary of Kansas and the western boundary of Missouri in the central United States. It covers seven metropolitan counties with a total area of 8215.5 km2, and was ranked high among the sprawlthreatened metropolitan areas in the United States (Ewing, Pendall, & Chen, 2002; Sierra Club, 1998). Thirteen cities in this area were included in the study, with at least one city from each county (Fig. 1). The study area witnessed significant population and economic growth in recent decades. The total metropolitan population increased from 1,273,170 in 1970 to 1,656,014 in 2001, with an annual growth rate of 1.04% (US Census Bureau, 2002). As the Kansas City metropolitan area spreads over parts of two states, Kansas and Missouri, the two parts historically present different rates and patterns of socioeconomic development.

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

865

Fig. 1. Study area: the Kansas City metropolitan area. [The map depicts the seven counties selected for the analysis (Wyandotte, Johnson, and Miami in Kansas, and Cass, Jackson, Clay, and Platte in Missouri). The map also shows the cities selected for the study (numbered 1–13 they are: (1) Osawatomie, (2) Paolo, (3) Overland Park, (4) Shawnee, (5) Kansas City, KS, (6) Parkville, (7) Platte, (8) Kansas City, MO, (9) Smithville, (10) Kearny, (11) Liberty, (12) Lee’s Summit, and (13) Raymore).]

The topography of the area is generally characterized by rolling hills and open plains, with grasslands, croplands, forests, urban zones, and scattered water bodies as the predominant land cover types. Because of the population and economic growth, the region has experienced significant alteration of its natural landscapes as its urban built-up land increases (The Brookings Institution, 2002). However, there is a lack of knowledge of historical land cover dynamics at regional, county, and city scales, which hinders a comprehensive understanding of urban sprawl patterns and informed urban planning in the metropolitan area. Thus, the analysis of general trends and effects of urban sprawl as it relates to the heterogeneous mix of both landscapes and driving forces has become the central research need for this study area. 3.2. Remote sensing methods Image pre-processing, classification, and classification accuracy assessment were performed using ArcGIS (ESRI, Inc.) and ERDAS IMAGINE (Leica, Inc.). Six Landsat images (Path 28/Row 33 for MSS and Path 26/Row 33 for TM and ETM+) covering

866

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

the Kansas City metropolitan area were obtained for the years 1972, 1979, 1985, 1992, 1999, and 2001 from the US Geological Survey (USGS). The images from mid July to mid August were chosen for better land cover detection during the fast vegetation growth season. The data sources for 1972 and 1979 were Landsat-1 and Landsat-3 Multispectral Scanner (MSS), respectively. The data sources for 1985 and 1992 were Landsat-5 Thematic Mapper (TM), and for 1999 and 2001 images were Landsat-7 Enhanced Thematic Mapper (ETM+). The 1972 and 1979 images have a spatial resolution of 57 m · 57 m (nominal resolution), the 1985 image 28.5 m · 28.5 m (nominal resolution), and the 1992, 1999, and 2001 images 30 m · 30 m. The purchased images were rectified and georeferenced to the UTM projection (WGS 84 datum, Zone 15 North) by USGS prior to their distribution. We conducted an imageto-image registration of all images to the master scene, the 1999 image. The total root mean square error obtained for the Ground Control Points was 0.469 pixel (26.73 m) for 1972, 0.073 pixel (4.16 m) for 1979, 0.524 pixel (15.7 m) for 1985, 0.507 (15.21 m) for 1992, 0.486 pixel (14.6 m) for 2001. An ‘‘area of interest’’ (AOI) boundary representing the metropolitan study area was delineated for the remote sensing analysis. The city and county boundary data sets were obtained from the US Census Bureau (Census, 2000). A northwest portion of Platte County (approximately 30% or 335 km2), a small portion of northwest Johnson County (13% or 158 km2), and the northwest tip of Miami County (1.82% or 27 km2) were not included in the study area because of the swath limits of the satellite images. The omission of this small portion of the metropolitan area applied to all the analyzed images, thus did not affect the understanding of general trends of urban land cover changes through comparing these images. We conducted the land cover classification at the metropolitan level. The six Landsat images were classified using the supervised maximum likelihood classification method (Jensen, 2005). For the purpose of characterizing urban sprawl, we identified four major types of land cover: built-up area, forestland, non-forest vegetation, and water body (Fig. 2). The ‘‘built-up’’ class depicts residential areas of single houses and apartment buildings, shopping centers, industrial and commercial facilities, highways and major streets, and associated properties and parking lots. The ‘‘non-forest vegetation’’ class includes grasslands, brush land, and cropland. The 1972, 1979, and 1985 images were resampled in post-classification using the nearest neighbor interpolation to match the remaining images of 30 m · 30 m spatial resolution. During the supervised classification, a spectral signature file was generated and used for each of the six images. In this process, with information from field verifications, city maps, and online high-resolution images, the areas with known land cover types were identified on the images as class signature training sample sites. For different land cover types, the number of the training samples varied from approximately 10–100. For example, approximately 10 training samples were selected for water, which can be easily identified, while over 100 training samples were used for the non-forest vegetation that could be mistakenly classified into residential/commercial types. In the signature evaluation, individually selected training samples in each signature file were manipulated (deletion or merging with other samples) until the most satisfactory classification was achieved. An accuracy assessment was performed at the metropolitan level. For an unbiased assessment, the stratified random sample strategy (Jensen, 2005) was used to select 50 samples for each class totaling 200 points per image. To substantiate the results of the classi-

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

867

Fig. 2. Classified satellite images for the six detection years.

fication of the 1999 and 2001 images, the images representing recent or current land covers, fieldwork and map verifications were conducted using GPS to locate and compare the sampled pixels in question with the corresponding land covers on the ground. To assess the classification of the historical images of 1972, 1979, 1985, and 1992, the following two methods were employed: (1) field verification for the randomly generated reference points falling in the areas that had almost no changes over the period of the study, such as public parks and forests and historic business districts; additional verification of these sites was done with online high-resolution images of the metropolitan area (http://www.terraserver.com); (2) a variety of historical documents were used for ground

868

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

truthing. More specifically, we used more than 50 USGS topographic maps to verify the historical land cover types over the study period. Historical street maps and aerial photos for the same period were also collected and analyzed to provide further verification of past land cover types. Many of those historical documents were obtained from online sources such as the USGS repository (http://terraserver.microsoft.com). The results of the accuracy assessments are reported in Table 1, including the overall, producer’s and user’s accuracies as well as Kappa indexes (Jensen, 2005) of land cover classifications of the six images. Generally, the accuracy of the classifications is satisfactory, with the lowest values for the 1972 image (an overall accuracy of 85.50% and a Kappa index of 0.867). Except for the 1972 classification, the non-forest vegetation type had lower producer’s accuracies than the other land cover types for the remaining five images, ranging from 72.13% to 79.59%. This class type includes grasslands, brush, and cropland which are mixed with residential and commercial land cover types in many locations, potentially resulting in misclassifications. 3.3. Landscape information extraction 3.3.1. Across-scale land-cover data generation Land cover data at the metropolitan level is critical to understand the general long-term patterns of urban landscape change. Directly derived from the satellite image classifications, this data set includes area statistics for the four types of land cover (built-up area, forestland, non-forest vegetation, and water body) in the six detection years. To obtain land cover statistics for all the counties and cities in the metropolitan area, we subset the classified images with the corresponding county and city boundary AOI files. 3.3.2. Landscape metrics calculation Based on the remotely sensed land cover data, selected landscape metrics were calculated using the FRAGSTATS program (McGarigal & Marks, 1995) for each of the selected cities and counties, at the landscape (metropolitan) level, and for each of four directional sectors of the metropolitan area (see Section 3.3.3). The metrics include the patch density (PD) index of the built-up land cover as well as the patch density, the largest patch index (LPI), and the aggregation index (AI) of both forestland and non-forest vegetation. The PD equals the number of patches of the corresponding land cover type Table 1 Accuracy assessment of the classified images for the six detection years 1972 P (%) Built-up Forest Non-forest Water Overall accuracy Kappa indexes

1979

1985

1992

1999

U (%) P (%) U (%) P (%) U (%) P (%) U (%) P (%)

82.69 86.00 81.13 86.00 86.05 74.00 92.31 96.00 85.50

86.54 90.00 82.69 86.00 79.59 78.00 93.62 88.00 85.50

93.02 80.00 95.56 86.00 72.13 88.00 94.12 96.00 87.50

91.11 82.00 84.31 86.00 78.95 90.00 95.74 90.00 87.00

0.8067

0.8067

0.8333

0.8267

P stands for producer’s accuracy and U stands for user’s accuracy.

2001 U (%)

P (%)

U (%)

90.91 80.00 89.36 84.00 91.49 86.00 95.65 88.00 74.58 88.00 75.86 88.00 100.00 100.00 100.00 98.00 88.50 89.50 0.8467

0.8600

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

869

divided by total area of interest. It is a fundamental measure of land cover pattern of a class type or a landscape. The LPI at the class level quantifies the percentage of total land area comprised by the largest patch. It is a simple measure of the dominance of a land cover type. The AI is calculated from a patch adjacency matrix, which shows the frequency with which different pairs of patch types appear side-by-side on the map. It quantifies the degree of fragmentation of a land cover type or a landscape (He, DeZonia, & Mladenoff, 2000). According to Wu (2004), the PD and LPI are the Type I landscape indices that show the consistency of scaling relations between different landscapes and the similarity of the scale relations between different patch types within the same landscape. In addition, the simplicity of their computational expressions allows the identification of urban growth patterns by directly interpreting landscape metrics values. For example, increased urban built-up land cover may eventually reach a stage that has a reduced number of built-up land patches having increased sizes in a given area, resulting in low values of the built-up patch density. It is worthy noting that a per-pixel classification may generate the ‘‘salt-and-pepper’’ effect. It can result in scattered patches of very small sizes, potentially reducing the effectiveness of patch metrics to analyze landscape patterns. In some situations, the minimum patch size can be the size of a single pixel. We generated classified land cover maps intended to reduce the impact of the ‘‘salt-andpepper’’ effect. First, we used a small number of land cover types (four only). Secondly, as a key measure of classification quality, we tried to produce land cover maps that were as ‘‘clean’’ as possible by carefully evaluating and manipulating the spectral signatures of each land cover type until satisfactory results were reached. Finally, the neighbor rule was applied to link certain adjacent pixels to form a patch, further reducing the occurrences of patches with a single pixel or a small number of pixels. 3.3.3. Built-up spread pattern analysis Effective forecasting of urban sprawl dynamics depends largely on the understanding of subtle spatial and temporal patterns of the built-up land. For this purpose, we used selected landscape metrics to characterize the geographic distribution of built-up land cover as it has expanded away from the urban core and towards the periphery of the metropolitan area (Antrop & Van Eetvelde, 2000; Herold et al., 2002). To highlight long-term trends in spread patterns over the three decades, we compared the classified 1972 and 2001 images. A circle of 7.5 km in radius was drawn to delineate the boundary of a metropolitan urban core as we estimated it to be in 1972. Ten concentric buffer rings (4 km wide) were added to the urban core. To quantify the magnitude of urban sprawl over the different sectors of the metropolitan area in the last three decades, we delineated North to South (N–S) and East to West (E–W) transects cutting across the center of the depicted urban core and with lengths of 108 km and 69 km, respectively. Their axes divide the metropolitan area into four sectors (NW, NE, SW, and SE). Built-up land cover data for each buffer ring in each sector was then generated from the classified images for the years 1972 and 2001. Finally, to describe built-up spread effects, the built-up patch density and the forest aggregation indices for each of the 10 buffer rings were generated using the FRAGSTATS program (McGarigal & Marks, 1995). As the patch density provides information about the number of patches per unit area, it facilitates comparisons among landscapes of varying sizes at different distances from the urban core.

870

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

For time series analysis of the spreading of the built-up areas, the built up-areas generated for each buffer rings for 1972 were subtracted from the 2001 data. 3.3.4. Land consumption metrics For the jurisdiction-based land cover data we devised land consumption indices (LCI) (Eqs. (1) and (2)) that relate built-up area change to change in housing and commercial construction as major driving factors in urban land conversion. These construction-based metrics can be used to effectively compare land consumption across jurisdictions and to analyze the relative contribution of housing construction and commercial construction in urban land use. To calculate these metrics, we first obtained data on housing units and non-farm business establishments from the US Census Bureau (Census, 2000) for the study counties in the time periods corresponding or close to the satellite imagery dates. We then defined two indices, the ratios of the percent change of built-up land coverage to the percent change of the housing units (HU) or the business establishments (BE) in specific jurisdictions in a particular time period: LCIðHUÞ ¼ LCIðBEÞ ¼

%DBðt1 t0 Þ %D½HUðt1 t0 Þ

%DBðt1 t0 Þ %D½BEðt1 t0 Þ

ð1Þ ð2Þ

LCI land consumption index HU housing unit BE business establishment %DB percent change in built-up land %D[HU] percent change in housing unit %D[BE] percent change in business establishment beginning time; ending time t0; t1

4. Results and discussion 4.1. General trends General trends in land cover change were identified at the metropolitan, county, and city levels. As summarized in Table 2, the results of satellite image classification (Fig. 2) suggest that the metropolitan area has experienced significant land conversions mainly due to urban expansion. At the metropolitan scale, the built-up land cover increased over the past three decades at an average rate of 4.25% per year, and the non-forest vegetation cover bore the major burden of urbanization. The forestland cover remained relatively unchanged at the metropolitan level. The water body class, which was the smallest land cover among the four detected classes, doubled its extent, as major man-made reservoirs were created over the study period. Studying the spatial and temporal heterogeneity of the land cover changes at the county level allowed us to identify fast and slow sprawling areas that could not be detected at the metropolitan level. Fig. 3 suggests that all the counties experienced built-up land cover

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

871

Table 2 Proportion of land cover types obtained from classified satellite images in the Kansas City metropolitan area

Built-up Forest Non-forest Water

1972 (%)

1979 (%)

1985 (%)

1992 (%)

1999 (%)

2001 (%)

8.65 16.36 73.93 1.05

10.38 16.44 71.96 1.22

12.03 16.81 69.06 2.10

15.41 16.58 66.03 1.99

18.69 16.71 62.70 1.90

19.19 17.36 61.24 2.21

Fig. 3. Time series of built-up land cover (%) for the metropolitan Kansas City and study counties, 1972–2001.

increases in the six study years, but at significantly varying rates. Johnson and Clay Counties were the fastest growing areas over the past three decades, adding urbanized lands at an annual rate of 8.98% and 4.87%, respectively. Wyandotte County, on the other hand, was the slowest growing county, with an annual built-up growth rate of 1.88% over the same period. At a finer scale, the city-level data analysis was focused on identifying sprawl patterns that differ from those detected by the county-level analysis. Over the past three decades, all the study cities had gained built-up land at a greater rate than the corresponding county in which each city is located (Fig. 1). In terms of percentage of coverage, some cities (Shawnee and Overland Park) lost more forestland than the corresponding county. Furthermore, the time series data revealed that between 1972 and 2001, the selected suburban cities experienced an average growth rate of built-up area 3.22 times higher than that of the ‘‘urban-core’’ cities of Kansas City, Missouri and Kansas City, Kansas, showing a pattern of new urbanization centered in the suburbs (Fig. 4). More specifically, the suburban cities of Lee’s Summit, Shawnee, Overland Park, and Platte, witnessed high annual rates of built-up land cover increase of 12.28%, 9.4%, 9.09%, and 8.63%, respectively.

872

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

Detection Periods

1992-2001

1985-1992

1979-1985

Suburban cities Urban-core cities

1972-1979

0.00%

10.00%

20.00%

30.00%

40.00%

50.00%

60.00%

Rate of Built-up Growth

Fig. 4. Growth rates of urban-core cities versus suburban cities.

Landscape metrics were analyzed to reveal patterns of change in landscape characteristics at varying spatial scales. At the metropolitan level, the non-forest vegetation and the forestland became more fragmented as the result of the growing of the built-up area. This trend was demonstrated by the positive correlation between the built-up area (%) and the index of patch density of non-forest vegetation and forestland (NF-PD and F-PD) (Fig. 5a). It was also indicated by the negative correlation between built-up area and the aggregation index of both forestland and non-forest vegetation (F-AI and NF-AI) (Fig. 5b). Fig. 5c shows that as the built-up area has increased, the largest patch index of the nonforest vegetation (NF-LPI) decreased while the LPI of the forestland (F-LPI) remained relatively stable. This suggests that the non-forest vegetation experienced decrease and was fragmented in a larger range of patch sizes, while larger patches of forest were less affected at the metropolitan level. These trends in landscape change were also identified at both county and city levels (Table 3). However, at the city level, the correlations between landscape metrics and built-up area were the weakest at all the three spatial scales of analysis. This may suggest that the broader landscape effects of urbanization can be better revealed within larger spatial units like metropolitan areas and counties. 4.2. Built-up spread pattern analysis We used the built-up area patch density to characterize and compare the patterns of sprawl across the two most rapidly growing sectors (SW and SE) of the metropolitan area between 1972 and 2001 (Fig. 6). Interpretation of changes in patch density revealed that the two sectors were in different stages of urbanization. For the baseline year (1972), low patch density values were found, decreasing from the urban core (Fig. 6), suggesting that both sectors were in an early stage of urbanization. After three decades of growth, both sectors had higher built-up area patch densities (Fig. 6). However, the curves of 2001 suggest a different interpretation of the patch density change as compared to the early stage of urbanization: Lower patch density values of

Increasing Rate (%)

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

873

25 20 15 10 5 0 Metro KC 1972

Metro KC 1979

Metro KC 1985

Metro KC Built-up Percent (%)

Metro KC 1992 F-PD

Metro KC 2001 NF-PD

(a)

Increasing Rate (%)

110 90 70 50 30 10 -10

Metro KC 1972

Metro KC 1979

Metro KC 1985

Metro KC Built-up Percent (%)

Metro KC 1992 F-AI

Metro KC 2001 NF-AI

Increasing Rate (%)

(b) 45 35 25 15 5 -5

Metro KC 1972

Metro KC 1979

Metro KC 1985

Metro KC Built-up Percent (%)

Metro KC 1992 F-LPI

Metro KC 2001 NF-LPI

(c) Fig. 5. Correlation between the built-up (%) area and non-forest vegetation (NF-LPI) and forestland (F-LPI). (a) Correlation between the built-up area (%) and patch density of non-forest vegetation (NF-PD) and forestland (FPD). (b) Correlation between the built-up area (%) and the aggregation index of non-forest vegetation (NF-AI) and forestland (F-AI). (c) Correlation between the built-up (%) area and the largest patch index of non-forest vegetation (NF-LPI) and forestland (F-LPI).

built-up area reflect ‘‘fill-in’’ effects of urban development, which usually result in fewer and larger patches. Thus, the patterns of built-up area spread in the two sectors indicate different degrees of urbanization. Apparently, a higher level of and larger variation in built-up area patch density of the SE sector suggests that urbanization in this sector is

874

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

Table 3 Correlation coefficients for the relationships between the built-up area and the landscape metrics of forestland (F) and non-forest vegetation (NF) in 11 selected study cities and seven study counties for the period of study Jurisdiction level

Land cover type

Patch density (PD)

Largest patch density (LPI)

Aggregation index (AI)

City

Forest Non-forest

0.165 0.312*

0.269 0.300*

0.129 0.261

County

Forest Non-forest

0.384* 0.675**

0.197 0.514**

0.308 0.545**

* **

Correlation is significant at P < 0.05 level. Correlation is significant at P < 0.01 level.

0.9 0.8

Built-up Patch Density

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0-4

4-8

8-12

12-16

16-20

20-24

24-28

28-32

32-36

36-40

Distance from Urban Core (Km) 1972 SW

1972 SE

2001 SW

2001 SE

Fig. 6. Built-up patch density at different distances (km) from the urban core in the SW and SE sectors of the metropolitan.

proceeding more rapidly, with many distinct built-up clusters, while the SW sector has already become an area with large continuous urbanized lands. This contrast is clearly demonstrated in the classified satellite image for 2001 (Fig. 2). Differences in the built-up land patterns between the two sectors were further confirmed with a statistical analysis: In the SW sector, there was a strong negative correlation between the patch density values and the corresponding distances from the urban core (r = 0.626 with p < 0.01). By contrast, the SE sector shows a weak negative correlation (r = 0.361). In addition, the valleys of patch density curves revealed that significant ‘‘fillin’’ effects of built-up activity had occurred at the distance of 12–16 km and 24–28 km from the urban core for the SW sector, and of 4–8 km and 16–20 km for the SE sector, respectively. Fig. 7 shows the forest aggregation index (F-AI) relative to built-up areas at all the buffers for the same two sectors in 2001. A statistical analysis suggests that there is a signif-

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

875

90 80

Built-up Area (Km2)

70 60 50 40 30 20 10 0 Urban Core

0-4

4-8

8-12

12-16

16-20

20-24

24-28

28-32

32-36

36-40

Distance from Urban Core (Km) Buil t-up Area (SE-2001)

F-AI (SE-2001)

Buil t-up Area (SW-2001)

F-AI (SW-2001)

Fig. 7. The built-up area and forest aggregation indexes for the SE and SW sectors of the study area at various distances from the urban core.

icant positive correlation between the F-AI and the built-up area for the SW sector (r = 0.563, p < 0.01) while there is a weak negative relationship for the SE sector (r = 0.083). This reveals an interesting effect of urban expansion on forest lands: based on the values of F-AI, forest lands appeared to be more ‘‘aggregated’’ (less ‘‘fragmented’’) at locations having a high coverage of built-up lands (as is the case in the SW sector). These values seems to run against the common assumption that generally forestland in urbanized areas would become more fragmented (less aggregated) as urban built-up area increases. A possible explanation is that in highly developed urban areas small patches of forest lands may have been used for construction, thus larger remaining forest patches resulted in higher values of F-AI. Our field observations supported this explanation. 4.3. Land consumption metrics interpretation Due to the limited availability of housing and business construction data, the construction-based indices (B/HU and B/BE: Eqs. (1) and (2)) were calculated over two detection periods of 1979–1990 and 1990–2001 only at the county level (Table 4). Thus, this effort was focused on method development rather than regional application analysis. By replacing housing data in Eq. (1) with the population data from relevant US censuses, population-based indices (B/PP) were also calculated for the same periods for comparative analysis (Table 4). Identifying land consumption patterns with these indices requires the proper interpretation of index values. First, a negative value of the index usually indicates decreased housing or business construction activities because it would be reasonable to expect the builtup land to increase or remain nearly unchanged in an urbanizing area over a period of study, which results in a positive numerator (see Eqs. (1) and (2)).

876

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

Table 4 Land consumption indices for selected counties for the periods 1979–1990 and 1990–2001 1979–1990

Metropolitan KC Cass Co. Clay Co. Jackson Co. Johnson Co. Platte Co. Miami Co. Wyandotte Co.

1990–2001

B/HU

B/BE

B/PP

B/HU

B/BE

B/PP

4.37 2.8 1.62 7.84 2.01 0.08 4.95 25.24

1.23 0.92 0.53 3.35 0.92 0.09 1.24 6.31

4.06 2.44 1.74 39.2 2.09 0.39 2.54 3.64

1.31 0.32 1.55 3.40 1.63 0.93 0.65 3.09

1.73 0.33 2.78 3.31 1.69 0.91 0.87 2.33

2.14 0.52 2.01 4.93 2.10 1.23 1.15 15.06

B, BE, HU, and PP denote rates of change of built-up area, business establishments, housing units, and population, respectively.

Secondly, by comparing the two construction-based indices we can identify the relative significance of housing construction and business construction on land consumption in a jurisdiction (a county in this case) during a given period of time. For example, in Jackson County, for the built-up land change during the period 1979–1990, a smaller value of B/ BE (3.35) along a larger value of B/HU (7.84) suggests that business establishment might have played a more important role than housing construction on land consumption. Thirdly, for a given period of time, a higher index value for a sprawling jurisdiction would normally indicate a higher degree of urban sprawl (i.e., a lower density of land use). This situation would mean that relatively larger lands have been converted for housing and/or business construction, resulting in a larger ratio of the built-up land change to the construction change. According to our analyses, the B/HU indices for the period 1979–1990 reveals that Jackson, Wyandotte, and Miami counties were the most sprawling, and that Platte County sprawled the least. Between 1990 and 2001, Jackson, Johnson, and Clay counties experienced greater sprawling than the rest of the study counties according to B/HU index values. Lastly, construction-based indices can provide complementary understanding of land consumption, especially when using population-based indices alone could lead to skewed or invalid interpretations of land use. For instance, in the period 1979–1990, the population-based index (B/PP) for Wyandotte County was 3.64, indicating a population decrease. Using the population-based indices alone it would incorrectly assume that land consumption might have been decreased accordingly. However, the fact was that the county’s construction activity was still increasing during the period (Fig. 3), which was revealed by a low gain in housing units (B/HU = 25.24) and a relatively higher growth in business establishments (B/BE = 6.31). Only when the county’s population continued to decrease (B/PP = 15.06) in the following period 1990–2001, housing and commercial construction reversed these trends (B/HU = 3.09, B/BE = 2.33). This example suggests that the analysis of land consumption patterns may require the use of different types of metrics (construction-based vs. population-based). 5. Summary and conclusions This study demonstrated that regional patterns of urban sprawl were effectively identified and characterized using multi-stage satellite images and landscape metrics. This was

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

877

achieved by our approach that systematically explored and compared spatiotemporal patterns of urban expansion across different levels of jurisdictions with effective indicators and measurements of land cover changes. Over the past three decades, urbanization has significantly modified the land cover of metropolitan Kansas City. The built-up land markedly increased, mostly at the expense of the non-forest land, for all the study counties and cities as well as for the metropolitan area as a whole, while forestland remained relatively unchanged at the metropolitan level. County- and city-based analyses identified fast and slow sprawling spots of the metropolitan area. Urban sprawl predominantly expanded toward the southwest and southeast sectors of the metropolitan area. The results of the metropolitan-, county-, and city-level analyses indicate that sprawl patterns differed across jurisdictions and geographic scales. The study identified the landscape effects and spatial patterns of built-up land expansion. As a general trend at the metropolitan level, non-forest vegetation and forestland patches became more fragmented as a result of the increase of built-up area over the period of study. Landscape metrics analysis also indicates that at the metropolitan level large forest areas (patches) were less affected than areas of non-forest vegetation. The landscape metrics also suggest that effects of urbanization can be better identified at the metropolitan or county levels than at the city level. Comparative analysis of change in built-up patch density in relation to the spread pattern indicates that the SW and SE sectors of the metropolitan area are in different stages of urbanization. This analysis also reveals the ‘‘fill-in’’ effects of urban development as spreading away from the urban core. Statistical analysis suggests that forest aggregation indices have positive correlation with the area of built-up land in places that are highly developed, remaining fewer small forest patches. Compared to conventional population-based indices, the construction-based land consumption indices devised in this study provide a good measure that can help understanding the degree of urban sprawl across jurisdictions and identifying the relative impacts of housing construction and commercial construction on land consumption. These indices can provide complementary information on urban land conversion when solely using population-based indices may skew the interpretation. Acknowledgement This study was supported by the following grants awarded to W. Ji: US Environmental Protection Agency’s Regional Geographic Initiative Program (Grants MM98705401 and MM98735701) and the University of Missouri-Kansas City’s Faculty Research Grant. References Antrop, M., & Van Eetvelde, V. (2000). Holistic aspects of suburban landscapes: visual image interpretation and landscape metrics. Landscape and Urban Planning, 50, 43–58. Brueckner, J. K. (2000). Urban Sprawl: Diagnosis and remedies. International Regional Science Review, 23(2), 160–171. Clapham, W. B. Jr., (2003). Continuum-based classification of remotely sensed imagery to describe urban sprawl on a watershed scale. Remote Sensing of Environment, 86(3), 322–340. Chen, S., Zeng, S., & Xie, C. (2000). Remote Sensing and GIS for urban growth analysis in China. Photogrammetric Engineering and Remote Sensing, 66(5), 593–598.

878

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

Civco, D. L., Hurd, J. D., Wilson, E. H., Arnold, C. L., & Prisloe, S. (2002). Quantifying and Describing Urbanizing Landscapes in the Northeast United States. Photogrammetric Engineering and Remote Sensing, 68(10), 1083–1090. Epstein, J., Payne, K., & Kramer, E. (2002). Techniques for mapping suburban sprawl. Photogrammetric Engineering and Remote Sensing, 68(9), 913–918. Ewing, R., Pendall, R., & Chen, D. (2002). Measuring sprawl and its impact. Smart Growth America. Available from . Last visited November 11, 2003. Fulton, W., Pendall, R., Nguyen, M., & Harrison, A. (2001). Who sprawls the most? How growth patterns differ across the US. The Brookings Institution Center on Urban and Metropolitan Policy. Hasse, J., & Lathrop, R. G. (2003). A housing-unit-level approach to characterizing residential sprawl. Photogrammetric Engineering and Remote Sensing, 69(9), 1021–1030. He, H. S., DeZonia, B., & Mladenoff, D. J. (2000). An aggregation index to quantify spatial patterns of landscapes. Landscape Ecology, 15, 590–602. Herold, M., Scepan, J., & Clarke, K. C. (2002). The use of remote sensing and landscape-metrics to describe structures and changes in urban land uses. Environment and Planning A, 34, 1443–1458. Herold, M., Goldstein, N. C., & Clarke, K. C. (2003). The spatiotemporal form of urban growth: measurement, analysis and modeling. Remote Sensing of the Environment, 86, 286–302. Jensen, J. R. (2005). Introductory to digital image processing: a remote sensing perspective (3rd ed.). Upper Saddle River, NJ, USA: Prentice Hall (pp. 526). Ji, C. Y., Lin, P., Li, X., Liu, Q., Sun, D., & Wang, S. (2001). Monitoring urban expansion with remote sensing in China. International Journal of Remote Sensing, 22(8), 1441–1455. Johnson, M. P. (2001). Environmental Impacts of urban sprawl: a survey of the literature and proposed research agenda. Environment and Planning A, 33(4), 717–735. Lambin, E. F., Turner, B. L., Geist, H. J., Agbola, S. B., Angelsen, A., Bruce, J. W., et al. (2001). The causes of land-use and land-cover change: moving beyond the myths. Global Environmental Change, 11, 261–269. Lo, C. P., & Yang, X. (2002). Drivers of land-use/ land-cover changes and dynamics modeling for the Atlanta. Georgia Metropolitan Area. Photogrammetric Engineering and Remote Sensing, 68(10), 1073–1082. Luck, M., & Wu, J. (2002). A gradient analysis of urban landscape pattern: A case study from the Phoenix metropolitan region, Arizona, USA. Landscape Ecology, 17(4), 327–339. Lucy, W. H., & Phillips, D. L. (2001). Suburbs and the Census: Patterns of Growth and Decline. Washington, DC: The Brookings Institute. McGarigal, K., & Marks, B. J. (1995). FRAGSTATS: spatial pattern analysis program for quantifying landscape structure. USDA Forest Service General Technical Report PNW-351. Masek, J. G., Lindsay, F. E., & Govard, S. N. (2000). Dynamics of urban growth in the Washington, DC metropolitan area, 1973–1996, from Landsat observations. International Journal of Remote Sensing, 21(18), 3473–3486. Peiser, R. (2001). Decomposing urban sprawl. Town Planning Review, 72(3), 275–298. Pendall, R. (1999). Do land-use controls cause sprawl? Environment and Planning B: Planning and Design, 26(4), 555–571. Phinn, S., & Stanford, M. (2001). Monitoring land-cover and land-use change in a rapidly urbanising coastal environment: The Maroochy and Mooloolah Rivers catchments, southeast Queensland, 1988–1997. Australian Geographical Studies, 39(2), 217–232. Ryznar, R. M., & Wagner, T. W. (2001). Using remotely sensed imagery to detect urban change. Viewing Detroit from Space. Journal of the American Planning Association, 67(3), 327–336. Shen, W., Jenerette, G. D., Wu, J., & Gardner, R. H. (2004). Evaluating empirical scaling relations of pattern metrics with simulated landscapes. Ecography, 27, 459–469. Sierra Club (1998). Sprawl Report, the dark side of the American dream: the costs and consequences of suburban sprawl. Available from . Last visited November 11, 2003. Sutton, P. C. (2003). A scale-adjusted measure of ‘‘urban sprawl’’ using nighttime satellite imagery. Remote sensing of Environment, 86, 353–369. The Brookings Institution Center on Urban and Metropolitan Policy (2002). The growth in the Heartland: challenges and opportunities for Missouri. US Census Bureau. Available from .

W. Ji et al. / Comput., Environ. and Urban Systems 30 (2006) 861–879

879

US Environmental Protection Agency (2001). Our Built and Natural Environments. A Technical Review of the Interactions between Land Use, Transportation, and Environmental Quality. 231-R-01-002. Available from . Last visited November 11, 2003. Wang, Y., & Moskovits, D. K. (2001). Tracking fragmentation of natural communities and changes in land cover: applications of Landsat data for conservation in an urban landscape (Chicago wilderness). Conservation Biology, 15(4), 835–843. Ward, D., Phinn, S. R., & Murray, A. T. (2000). Monitoring growth in rapidly urbanizing areas using remotely sensed data. Professional Geographer, 52(3), 371–386. Wu, J. (2004). Effects of changing scale on landscape pattern analysis: scaling relations. Landscape Ecology, 19, 125–138. Yang, X. (2002). Satellite monitoring of urban spatial growth in the Atlanta metropolitan area. Photogrammetric Engineering and Remote Sensing, 68(7), 725–734. Yeh, G. A., & Li, X. (2001). Measurement and monitoring of urban sprawl in a rapid growing region using entropy. Photogrammetric Engineering & Remote Sensing, 67(1), 83–90.