GIS and Evidence-Based Policy Making - Chapter 2

2 Urban Neighborhood Pattern Recognition UsingHighSpatialResolutionRemotelySensed Data and Point-Based GIS Data Sources Victor Mesev and Paul McKenzie CONTENTS 2.1 Introduction................................................................................................. 19 2.1.1 Difficulties of Remotely Sensing Urban Land Use ................... 20 2.1.2 Aggregated Urban Spatial Patterns............................................. 21 2.1.3 Disaggregated Urban Spatial Patterns........................................ 22 2.2 Point-Based Geographies........................................................................... 23 2.2.1 Point-Based Geographies of Bristol and Belfast........................ 25 2.3 Nearest Neighbor Indices of Point Distributions.................................. 26 2.3.1 Belfast and Bristol Nearest Neighbor Indices............................ 27 2.3.2 Linear Nearest Neighbor Indices of Point Distributions ......... 33 2.4 Using Postal Point Data to Infer Land Use from IKONOS Imagery.............................................................................. 34 2.5 Conclusions and Further Research.......................................................... 39 Acknowledgments............................................................................................... 40 References ............................................................................................................. 40 2.1 Introduction The recent revitalized interest in satellite urban remote sensing is as much the result of strengthening links with GIS as it is of the breakthrough in super-high spatial resolution satellite sensor data (Mesev, 2003a). For the first time, the rationale for using satellite sensor data in urban planning can go beyond coarse approximations of land-use characterization and become more sympathetic to the needs for up-to-date precision maps of land-use delineation, which in turn are critical for government policy decisions on individual household behavior and interaction (Donnay, 1999). As such, 2007 by Taylor & Francis Group, LLC. there is now also a realistic possibility that urban remote sensing can begin to emulate environmental remote sensing and play an increasingly central role for supporting national and regional policies on the estimation of population change (Chen, 2002), the calculation of quality-of-life indices (Lo, 2003), and the evaluation of transport flow, as well as the precise delineation of urban features and measurement of size, density, and height of buildings (Barnsley et al., 2003). 2.1.1 Difficulties of Remotely Sensing Urban Land Use Wider adoption of data from Earth-orbiting satellite sensors by policy makers for urban mapping and monitoring is restricted essentially by two groups of factors: scale and land-use inference (Forster, 1985). Scale is commonly associated with the spatial resolution of sensor instrumentation, which determines the smallest discernible unit of measurement (the pixel) and is the direct consequence of the instantaneous field-of-view capabilities of the satellite sensor. In other words, the spatial resolution of a sensor determines the degree of spatial disaggregation and therefore the level of spatial clarity in the satellite image. Traditionally, satellite sensor data have been recorded at spatial resolutions considered too coarse for successful and routine identification and categorization of urban features (Welch, 1982). Until 1999, the most widely available satellite sensor data for optical urban observation have been obtained from Landsat multispectral scanning sys-tem (MSS) sensors at a spatial resolution of 79 m, the Landsat thematic mapper (TM) series of sensors at 30 m, and SPOT high-resolution visible (HRV) at 20 m (multispectral) and 10 m (panchromatic) scales. Data at these scales were never designed to represent precisely the intricate spatial vari-ations and heterogeneity of the physical layout of built structures, and as a consequence, pixels typically represent aggregations of disorganized mix-tures of urban land cover types (both built and natural) at inconsistent proportions. They, instead, are more suitable for citywide applications, focusing, for instance, on a more macro differentiation of built-up structures from natural biophysical land covers, or the categorization of residential density suitable only as input data for urban growth and density profile models. The second limitation to the more widespread adoption of satellite sensor data for urban monitoring, and hence policy making, is more con-ceptual rather than technical (Geoghegan et al., 1998). This limitation refers to the subjective interpretation of socioeconomic constructs and policy-driven administrative mixtures of urban land use from the biophysical configuration of land cover properties. This is where the advantage of remote sensing over GIS as a more objective data collection source quickly diminishes when human occupation is inferred from discrete multispectral radiometric values that only represent the reflective and emittance proper-ties of the physical landscape. Various combinations of coarse scales and loose inference between land cover and land use have historically restricted both scope and accuracy in 2007 by Taylor & Francis Group, LLC. urban applications using satellite sensor data. Rising to the challenge, some methodologies recognized the inherent heterogeneity of images of urban areas and developed techniques that manipulated pixels not in isolation but within groups using textural (Moller-Jensen, 1990; Myint, 2003), contextual, and spatial properties and arrangements. When classifying, the difficulty of representing highly heterogeneous urban pixels was also acknowledged by methodologies designed to incorporate information from outside the spec-tral domain (e.g., fuzzy sets, neural nets, and Bayesian modifications by Mesev, 1998, 2001). Other work includes three-dimensional urban remote sensing employing laser-induced detection and ranging (LIDAR; Barnsley et al., 2003) and interferometric synthetic aperture radar (SAR; Gamba et al., 2000; Grey et al., 2003), as well as promising advances in hyperspectral image interpretation and multiple endmember spectral mixture models (Rashed et al., 2003). The intention in all cases is to increase accuracy by reducing uncertainty in both the discrimination of urban land cover pixels and the inference of land-use pixels. However, success in improving urban land-use accuracy has been small to negligible, usually qualified by local site-specific and time-specific conditions. Instead, research on the spatial, rather than the spectral, characterization of urban land cover has gained momentum in recent years bringing about the development of a number of quantitative indices. Among the most widely applied have been the scale invariant properties of fractal geometry, which many proponents have argued are capable of measuring the struc-tural complexity and fragmentation of highly heterogeneous urban land cover (De Cola, 1989). These same objectives were also tested using syntactic pattern recognition systems employing graph-based methodolo-gies (Barnsley and Barr, 1997) and automated expert systems (Tullis and Jensen, 2003). More recently, landscape or spatial metrics have revitalized fractal geometry as part of a suite of indices, including the contagion and patch density measurements (Wu et al., 2000; Herold et al., 2002; Greenhill et al., 2003), as well as spatial metrics directly related to urban sprawl (Hasse and Lathrop, 2003). Central to the formulation of these metrics is the concept of photomorphic regions, or homogeneous urban patches, which are rou-tinely extracted from aerial photographs but are more of a challenge from satellite imagery (Aplin et al., 1999). The two principal criticisms of spatial metrics are that their functionality is completely dependent on an initial spectral characterization of the satellite imagery, and that they are conspicu-ously absent from the actual process of the characterization of homogeneous classes. In this sense, they are merely measuring the outcome of the classi-fication regardless of the accuracy. 2.1.2 Aggregated Urban Spatial Patterns The common factor in most of the methodologies outlined above that restricts improvements in classification and pattern recognition accuracy is undoubtedly the inability to measure urban land use at a scale fine enough 2007 by Taylor & Francis Group, LLC. to identify individual building characteristics and hence infer human behavioral processes. If the objective is to delineate the maximum extent of human settlement then traditional approaches using coarse spatial reso-lution imagery and aggregated government statistics may suffice. However, such citywide measures have yet to convince planners and decision makers of their importance and, as a consequence, play only peripheral roles within local government policies (Donnay, 1999). If proponents of remote sensing and GIS want to rebuild the reputation of their data, they need to seriously tackle the limitations of aggregated urban models and begin to embrace the challenges of disaggregated alternatives. Although such models may be more demanding theoretically and technically, they are nonetheless essen-tial pragmatically. Moreover, aggregate models are the standard vehicles for extraneous information commonly used in the augmentation of remotely sensed images representing urban land use (Chen, 2002). Most of these extraneous data are extracted from national population censuses. This is understandable given their ease of access, wide range of socioeconomic indicators, almost com-plete coverage, and reasonably reliable representation. However, census information is considered too confidential by most governments to report at the individual household level. Instead, households are aggregated by areal tracts of assumed uniformity. Standard dasymetric techniques (Langford, 2003), those using surface models (Martin et al., 2000), and statis-tical relationships (Chen, 2002) can help redress this homogeneity, especially at the urban fringe, but the majority of tracts within the city remain summa-tive. The aggregated scale used in reporting census records may not be considered a major hindrance to integration with satellite imagery if the areal resolution is of comparable size, typically 900 and 400 m2 for the Landsat TM and SPOT HRV sensors, respectively. Besides, these scales are adequate for many citywide applications, in particular monitoring urban sprawl and modeling density attenuation (Longley and Mesev, 2002). More-over, accuracy in image interpretation using these relatively coarse sensor datawith aggregatedcensus tractsishighlyvariable,dueinlargeparttosite-specific inconsistencies in the imagery and the effects of the ecological fallacy and modifiable areal unit problem in the aggregated census tracts. 2.1.3 Disaggregated Urban Spatial Patterns Instead of relying on aggregated census tracts, we will consider an alterna-tive scale of measurement; disaggregated GIS data at the point level of abstraction (Harris and Longley, 2000). Point data representing individual buildings are grouped into localized two-dimensional spatial patterns char-acterizing various types of both residential and commercial developments (Mesev, 2003b). Suffice to say, the level of characterization available from point patterns is highly limited with only two properties distinctively and consistently measurable. These are density and linearity, both calculated by standard nearest neighbor and linear-adjusted nearest neighbor indices, 2007 by Taylor & Francis Group, LLC. resp ectively. Altho ugh point -based spatial pattern s correspo nd with the location of buildings , they are neve rtheless dim ensionl ess and as suc bear an incompl ete relation with the actual physica l size of the buildin (Harris and Ch en, 2004). For a mo re comp lete repre sentation, the relianc on high spat ial re solution imagery to deli neate the phys ical extent of b ing size and shape. Point -based patte rns can then be used to iden tify conf iguration of vario us typ es of residen tial and co mmerci al land use second -order, classified sa tellite imagery. Thei r spatia l properti es will fu ther be explor ed with in an image pat tern reco gnition system designed iden tify and charact erize urban st ructural pat terns. The spatial res olution the satellite sensor data has to matc h the disaggr egated scale of these poin based patte rns. As suc h traditiona l Earth-obs ervation imag ery are far coarse to identi fy indiv idual buil dings. Instea d, the advent of the lat gene ration of hig h spatia l resolu tion imag ery at un preceden ted scal such as 4 m (multispectral) and 1 m (panchromatic) from the IKONOS sensor (Space Imaging), has restored enthusiasm in more precise urban mapping by satellite imagery. Rather than interpret coarse approximations of residential or commercial groups of buildings, it is now possible to identify individual properties with a reasonable degree of spatial delinea-tion. Of course, aerial photography has a long history in exact urban mapping, but unlike satellite imagery dynamic urban monitoring is ham-pered by the high cost of photographic acquisition, limited spatial coverage, and a lack of spectral information. The long-term goal of the research outlined in this chapter is to reaffirm the inextricable links between satellite imagery and GIS data by identifying a role for point-based building patterns within image processing that would improve both the accuracy and precision of urban land-use interpretation from this very fine scale satellite imagery. 2.2 Point-Based Geographies Point -based da ta a re the ultimate in disaggr egation . In Euclide an geom terms, points are dim ensionl ess; inste ad of size and shape, they repres e the pre cise locations of indiv idual enti ties. This chapte r will introduce databas es that hol d poi nt-base d inform ation represe nting the locat of every postal delive ry a ddress in the United Kin gdom . Creat ed by Ordnan ce Surve ys of the United Kin gdom and known as AD DRESS-PO IN in Great Britain and COMP AS (COM puteriz ed Point Addr ess Service ) Northern Ireland both databases hold spatial and attribute information on the location of delivery addresses as well as whether they are residential or co mmer cial properti es. Tables 2.1 and 2.2 list the comp lete arr a information from both databases. The planimetric coordinates of the point data representing postal delivery buildings are claimed to be precise to within 0.1 m (50 m in some rural 2007 by Taylor & Francis Group, LLC. ... - tailieumienphi.vn