detecting and enumerating new building structures...

14
71 Geocarto International, Vol. 16, No. 1, March 2001 Published by Geocarto International Centre, G.P.O. Box 4122, Hong Kong. Detecting and Enumerating New Building Structures Utilizing Very-High Resolution Imaged Data and Image Processing DongMei Chen, Douglas Stow and Scott Daeschner Department of Geography San Diego State University San Diego, CA 92182, USA E-mail: [email protected] Linda Tucker Aerial Fotobank, Inc. 6181 Cornerstone Court, Suite 106 San Diego, CA 92121, USA Abstract Information on the number and type of new building structures is required by urban and transportation planners and the real estate industry. The goal of this paper is to explore the potential of high resolution imagery for meeting public and private sector demands for information on new buildings. The value of 1 m, 5 m, and 10 m panchromatic and 1 m color scanned aerial photography images acquired in 1997 and 1998 for a study area within the City of San Diego, California is assessed for general change detection and building enumeration. Both semi-automated and interactive change-detection approaches are evaluated. We demonstrate that interactive, visual-based approaches appear to be the most accurate (within 1% of actual count) and efficient approach for generating information on the number of new buildings associated with single family residential land use. More automated approaches to detecting and enumerating image microfeatures may be useful as enhancements for visual-based assessments and may be practical in areas composed mostly of large buildings associated with commercial and industrial land use. The highest accuracy for automated approaches was an undercounting of 11% for residential buildings and overcounting of 20% for those associated with commercial and industrial land use. Introduction Public sector policy decisions and private sector market analyses are often based upon information on the type and location of new buildings constructed each year. Potential users of such information include real estate, tax assessors, planning agencies, building contractors, demographers, banking institutions, planning agencies, and environmental organizations. The costs associated with traditional methods of extracting building information, however, need to be significantly reduced. Historically, land use changes have been manually detected by visually comparing recently acquired and historical imagery. An important component of land use change is the number of new buildings that are built yearly. Because of their relatively small size, buildings and other similar structures may be referred to as “microfeatures.” Manually identifying and then quantifying such features can be a time consuming and costly process. Aerial Fotobank Inc. (now called Landiscor, Inc.) is an aerial photography company providing both stock and custom aerial imagery in digital and analog formats. Each year, Aerial Fotobank acquires 1:48,000-scale, true color aerial photography for developed areas of southern California under stringent photogrammetric specifications. The annual image data base and Aerial Fotobank’s historical archives provide users with historical and current images for development, environmental, and urban planning analyses. Aerial Fotobank has been interested in utilizing their annual photography, or very high spatial resolution digital imagery from airborne or satellite sensors to generate information products on the number and location of buildings constructed each year. Aerial Fotobank contacted San Diego State University to conduct an assessment of the reliability and efficiency of image processing methods for deriving information on new building units. Many semi-automated and automated techniques have been established for deriving urban land use/cover change information (such as Ehlers et al. 1990, Forster 1983, Green et al. 1994, Howarth and Boasson 1983, Jensen and Toll 1982, and Jensen et al. 1994). These techniques include image overlay (Howarth and Wickware 1981), image differencing (Jensen and Toll 1982), image ratioing

Upload: lamdung

Post on 16-Apr-2018

223 views

Category:

Documents


2 download

TRANSCRIPT

71Geocarto International, Vol. 16, No. 1, March 2001Published by Geocarto International Centre, G.P.O. Box 4122, Hong Kong.

Detecting and Enumerating New Building Structures UtilizingVery-High Resolution Imaged Data and Image Processing

DongMei Chen, Douglas Stow and Scott DaeschnerDepartment of GeographySan Diego State UniversitySan Diego, CA 92182, USAE-mail: [email protected]

Linda TuckerAerial Fotobank, Inc.6181 Cornerstone Court, Suite 106San Diego, CA 92121, USA

Abstract

Information on the number and type of new building structures is required by urban and transportation plannersand the real estate industry. The goal of this paper is to explore the potential of high resolution imagery for meetingpublic and private sector demands for information on new buildings. The value of 1 m, 5 m, and 10 m panchromaticand 1 m color scanned aerial photography images acquired in 1997 and 1998 for a study area within the City of SanDiego, California is assessed for general change detection and building enumeration. Both semi-automated andinteractive change-detection approaches are evaluated. We demonstrate that interactive, visual-based approachesappear to be the most accurate (within 1% of actual count) and efficient approach for generating information onthe number of new buildings associated with single family residential land use. More automated approaches todetecting and enumerating image microfeatures may be useful as enhancements for visual-based assessments andmay be practical in areas composed mostly of large buildings associated with commercial and industrial land use.The highest accuracy for automated approaches was an undercounting of 11% for residential buildings andovercounting of 20% for those associated with commercial and industrial land use.

Introduction

Public sector policy decisions and private sector marketanalyses are often based upon information on the type andlocation of new buildings constructed each year. Potentialusers of such information include real estate, tax assessors,planning agencies, building contractors, demographers,banking institutions, planning agencies, and environmentalorganizations. The costs associated with traditional methodsof extracting building information, however, need to besignificantly reduced.

Historically, land use changes have been manuallydetected by visually comparing recently acquired andhistorical imagery. An important component of land usechange is the number of new buildings that are built yearly.Because of their relatively small size, buildings and othersimilar structures may be referred to as “microfeatures.”Manually identifying and then quantifying such featurescan be a time consuming and costly process.

Aerial Fotobank Inc. (now called Landiscor, Inc.) is anaerial photography company providing both stock andcustom aerial imagery in digital and analog formats. Each

year, Aerial Fotobank acquires 1:48,000-scale, true coloraerial photography for developed areas of southern Californiaunder stringent photogrammetric specifications. The annualimage data base and Aerial Fotobank’s historical archivesprovide users with historical and current images fordevelopment, environmental, and urban planning analyses.Aerial Fotobank has been interested in utilizing their annualphotography, or very high spatial resolution digital imageryfrom airborne or satellite sensors to generate informationproducts on the number and location of buildings constructedeach year. Aerial Fotobank contacted San Diego StateUniversity to conduct an assessment of the reliability andefficiency of image processing methods for derivinginformation on new building units.

Many semi-automated and automated techniques havebeen established for deriving urban land use/cover changeinformation (such as Ehlers et al. 1990, Forster 1983,Green et al. 1994, Howarth and Boasson 1983, Jensen andToll 1982, and Jensen et al. 1994). These techniques includeimage overlay (Howarth and Wickware 1981), imagedifferencing (Jensen and Toll 1982), image ratioing

72

Figure 1 The location of study area in San Diego, California, USA

(Howarth and Boasson 1983. Stow et al. 1990),principal component analysis (PCA) (Fung andLeDrew 1987), change vector analysis (Lambinand Strahler 1994), direct multidate classification(Estes et al. 1982), post-classification comparison(Gordon 1980, Howarth and Wickware 1981),and other advanced approaches such asknowledge-based system (Wang 1993). Forsimilar landscapes, various techniques may yielddifferent results. The type of change underinvestigation, data quality, availability, resolution,and the suitability of the algorithm for the changedetection required, all influence the choice of anoptimal technique to be used for image analysis(Green et al. 1994). Most of the previous studieshave focused on detecting land cover changefrom Landsat TM and MSS, and SPOT HRVdata. A few studies have evaluated thecomparative performance of various techniqueswith very high spatial resolution (1 to 5 m) digitalimages for urban-suburban areas (Green et al.1994, Singh 1989).

In the past two decades, different approacheshave been presented for automated buildingdetection and extraction (e.g. Huertas and Nevatia1988, Irvin and McKeown 1989, Lo 1995, Saharand Krupnik 1999, Shi and Shibasaki 1996, andWeidner and Forstner 1995). Most of these studiesare aimed at identifying or outlining the buildingsby using stereoscopic processing and/ortopographic analysis, which require digitalelevation models or a pair of stereo imagescontaining same buildings. Therefore, theseapproaches are often too complicated or expensiveand not feasible for extensive areas.

The detection and identification ofmicrofeatures, such as buildings, in digital imagesmay be very complex as their spectral and spatialcharacteristics may have great variability. Forexample, the contrast of buildings against theirbackground, their roof materials, and their shape,size and pattern may vary from one location toanother. The spacing between building units is akey determinant influencing the choice of imagespatial resolution and resultant error rates. Thesecomplicating factors make it difficult to developautomatic approaches that reliably detect andquantify new building units.

Our research focuses on pilot investigations totest which methods appear to be cost-effectiveand yield optimal results for detecting andcounting building changes. The methods includeboth visual interactive and semi-automatedapproaches. The overall objective of this study isto develop and test microfeature change-detectionapproaches to:1) detect areas with land use change by

comparing aerial images and/or high-resolution satellite imageryacquired in different years;

2) identify different categories of new building construction such asresidential houses, commercial/industrial buildings; and

3) quantify the annual change and count the number of new buildingsfor each category.

Study Area

The general study area is located in the northwest portion of theCity of San Diego, California, USA and is known as Sorrento Valley.This site is centred on latitude 117º10’W, longitude 32º54’N and hasan area of approximately 50 km2. The relative location in USA isshown in Figure 1. The study site contains numerous specific land usetypes within the broader categories of urban, agriculture andundeveloped land use, and therefore, a wide range of materials withdifferent spectral reflectance properties. It is a prosperous communitycharacterized by recent rapid growth especially in residential andlight industry. As a result, undeveloped land covered by semi-aridvegetation and agriculture of land are being converted to residential,commercial and industrial uses.

After conducting a change detection enhancement of the generalstudy area (described in the methodology section below), two sub-areas were selected for detailed analysis based on the presence ofknown building developments for several types of land uses betweenthe time period of the study, 1997 to 1998. One area consistedexclusively of single-family residential dwellings and one exclusively

73

Visual interpretationa) 1 m PANb) 1 m colorc) 5 m PANd) 1 m color differencee) Masked single-date

images

General change detection

Building Identification and Enumerating

Semi-automatedApproachesa) Unsupervised

classificationb) Supervised

classificationc) Thresholding

1997 and 1998scanned aerial

photos

Preprocessing(Registration, transformation,

aggregation)

Image matching

Image differencing(1 m. 5 m. and l0 m)

Visually delineatingchange areas

Figure 2 Procedures used in this study

Image Type

l m scanned color

1 m simulatedPanchromatic

5 m simulatedPanchromatic

10 m simulatedPanchromatic

Table 1 Image types tested for land use/cover and microfeature changes

Spatial resolution1 m

1 m

5 m

10 m

Acquisition Date

Jan. 10, 1997March 7, 1998

Jan. 10, 1997March 7, 1998

Jan. 10, 1997March 7, 1998

Jan. 10, 1997March 7, 1998

Source

Scanned color aerialphotographs

Intensity band fromscanned color aerialphotographs

Aggregated from 1 mPanchromatic

Aggregated from 1 mPanchromatic

of new light industrial buildings. Both areas aretopographically flat, though the general study area istopographically complex. Image processing and analysistrials for testing building feature identification andenumeration approaches were conducted for these two sub-areas.

Methodology

In the process of identifying detailed change of buildings,two steps were involved besides data preprocessing. Thefirst step was to detect change areas within the generalstudy area. The second step was to detect and count newlyconstructed buildings within these change areas. Theprocedures followed in this study are diagrammed in Figure2. ERDAS IMAGINE image processing and ARC/INFOGIS software were used to develop and test advancedimage processing approaches.

Data and PreprocessingDigital high resolution multispectral image data were

generated by scanning 1:48,000-scale true-color aerialphotography acquired on January 10, 1997 and March 7,1998 (Figure 3). The solar zenith angles were approximately55 and 45 degree and the solar azimuth angles wereapproximately 175 and 215 degree for the 1997 and 1998aerial photographs, respectively. The color aerialphotographs from both years were scanned using an AGFAHorizon Ultra scanner at 2000 dpi. Both scanned imageswere then registered to 1994 U.S. Geological Survey DigitalOrthophotographic Quarter Quadrangles (DOQQ) imagesby visually selecting 100-150 evenly distributed groundcontrol points (GCPs) and performing geometric registrationby image warping. The RMS errors ranged from 2 to 5pixels for subsets of l m resolution images.

The image types tested in this study are listed in Table 1.The two dates of scanned aerial photographs were evaluatedas black and white panchromatic and true color formats andhad a nominal spatial resolution of 1 m. Panchromaticimages were derived by extracting the intensity bandfollowing an Intensity-Hue-Saturation transformation ofthe three-band color digital image. It represented a simple,one-dimensional spectral data set that was storage andprocessing efficient and also served to simulate future 1 mresolution panchromatic commercial satellite imagery, suchas Ikonos, Orbview, and Quick Bird panchromatic data.These 1 m panchromatic images were aggregated to simulate5 m and 10 m images for both 1997 and 1998. The 1 mdigital color image data were evaluated to determine ifthree spectral (visible) dimensional data set would providegreater discrimination and delineation of roof tops ofbuildings. Simulated 5 m panchromatic were used tosimulate the 5 m Indian Remote Satellite (IRS-C)panchromatic image and evaluated to determine if readilyavailable 5 m image data were sufficient for uniquelyidentifying and delineating individual buildings associatedwith several different urban land use types.

Change Detection ApproachOur first consideration for selecting a suitable change

detection approach was the amount of computation forprocessing an extensive area. Since the goal of the initialchange detection phase was to stratify large study areas to

74

Figure 3 Red band images derived by scanning color aerial photography acquired in1997 and 1998 depict the general study area. The black boxes A and Boutline the location of specific study sites for residential (top box) and lightindustrial/commercial (lower box) land use

detect subregions of land use/cover change for subsequent processingof more detailed features, a complex approach is not warranted.Stated differently, the objective of this phase is to determine largerparcels of likely land use/cover change, within which new buildingsmay have been constructed. Singh (1989) compared the performanceof various change detection techniques using MSS data and found thesimple techniques such as differencing and ratioing performed betterthan more sophisticated transformations such as PCA analysis andclassification. Therefore, a simple technique of image differencingwas used. The second consideration for change detection was todetermine the appropriate spatial resolution for effective changedetection and stratification. Image spatial resolutions of 1, 5, and 10m were evaluated.

Panchromatic images derived from scanned aerial photographscaptured in 1998 were radiometrically normalized to 1997 scannedphotographs using an image-to-image regression approach. With thisapproach, image brightness values from one date were matched toimage brightness values from another date using coefficients of anordinary least-square regression line.

After subtracting 1997 images from adjusted 1998 images, the

resulting images yielded areas of brightnesschanges. Values closest to zero represented nochange while values further from zero, in boththe negative and positive directions, correlatedwith greater change. The resulting images fromimage differencing with 1 m, 5 m and 10 mresolution images are shown in Figure 4.

Based on visual interpretation of the changedetection images for the general study area, twosubareas were interactively delineated andextracted for subsequent microfeatureidentification. The following building featureidentification and enumeration approaches weretested for these two sub-areas.

Building-Feature Detection and EnumerationApproach

A variety of automatic or semiautomatictechniques for identifying and enhancing newbuilding features were selected from previousliterature and tested. These include per-pixelsupervised and unsupervised classification,image segmentation (Cross and Mason 1988),band ratioing (Jensen et al. 1994), edge andline fil tering (Wang 1987) and textureclassification (Gong and Howarth 1992). Forall of these approaches, a number of humandecisions were required to select training data,thresholds, and filters, and to reconnect andrebuild building object segments, renderingthe processes far from being “automated.” Thetime and computation costs associated withautomatic approaches tend to exceed those ofthe interactive approaches for the relativelysmall subareas. Thus, three types of imageprocessing and analysis approaches for trackingnew building developments were selected fortesting: 1) visual interpretation of multi-temporal images, 2) visual interpretation ofsingle-date images subjected to change masks,and 3) semi-automated microfeature detection.

Visual Interpretation – Multitemporal ImagesThe simplest approach of those tested was to

display each date of a multi-temporal image pairone at a time, or a multi-temporal differenceimage, and then to visually interpret and counteach microfeature that appeared to be anindividual building. An analyst interpreted andcounted from the images by viewing a colordisplay monitor. For the visual-based approaches,no further image processing is needed, so thecomputation time and cost is minimal mostlyassociated with interpretation time. Four typesof multi-temporal (1997 and 1998) images forthe two sub-areas were tested:

75

Figure 4 Change images derived from Panchromatic image differencing at (a) 10 meter, (b) 5 meter, and (c) 1 meterresolutions. The change images for the general study area are shown in the left while the two sub-areas A and B areshown in the right

l) 1 m simulated panchromatic black and white (singleband) image pair;

2) l m true color (three-band) image pair;

3) 5 m panchromatic black and white (single band) imagepair; and

4) l m true-color difference image (three band).

Visual Interpretation — Masked Single-Date imagesWhen identifying new buildings by interpretation of

multi-temporal image pairs, the analyst is required to maketwo passes at counting the number of buildings, which mayincrease the possibility of errors. When substantialmisregistration exists, color difference images may notclearly portray the boundaries of new buildings. An attemptat overcoming the above shortcomings led to thedevelopment and testing of masks for diverting interpretersaway from areas of existing buildings. As described inmore detail below, the masks are based on semi-automaticdetection of building features on the first date of the multi-

temporal image pair. Upon creation of the mask, the analystinterprets the unmasked (potential change) portions of thesecond (most recent date) of the multitemporal pair. Aswith the image difference product, the analyst makes abuilding count with a single interpretative pass.

The existing buildings masks were generated usingdifferent approaches for each of the two subareas. Forthe residential sub-area, a high-pass filter was applied tothe 1 m panchromatic 1997 image, which enhanced thehigh-frequency components. A texture image was createdby applying a 3x3 variance window to the high-passfiltered image. Low-pass filters were then applied to thetexture image to average and smooth the texture densitiesin residential area and non-residential area. Windowsizes of 5x5, 7x7, 9x9 to 11x11 were tested. A 7x7window size provided the best discrimination betweenresidential and non-residential area and was appliedsequentially in three passes. A binary image of residentialand non-residential areas was then created by thresholdingthe smoothed texture image. This threshold was selected

76

Figure 5 A series of image products used for multitemporal interpretation tests for residential subarea. Theground length of the sides of the square subarea is about 650 m

interactively such that only the residential areas weredisplayed. For the last step, 5x5 majority, minimum andmaximum filters were tested and applied to the binaryimage to remove small non-residential areas, and thin orfill the existing residential areas.

For the industrial/commercial sub-area, only twosimple processes were required to generate a non-changemask for subsequent identification of new buildings.First, a low-pass filter was applied to 1997 Pan image,which smoothed the surface of roofs of industrial/commercial buildings. A binary map was then created bythresholding the smoothed image. This threshold wasselected interactively so that only commercial/industrialroofs were displayed.

Interpreter TestingAll of the multi-temporal difference and masked 1998

(second date) images were subjected to visual interpretertests to examine their relative utility for making counts ofnew buildings. Figures 5 and 6 show the series images oftwo sub-areas used for visual interpretation. Six types ofimage products for the residential sub-area and four forindustrial/commercial sub-area were displayed on colormonitors of different computers. Nine people with remotesensing and image interpretation backgrounds interpretedand counted building units on all of the image products.The ordering of the images varied for the interpreters.The results are recorded in Table 2 and 3 for residentialand commercial subsets, respectively.

77

Image Masked MaskedInterpreter l m Pan 5 m Pan 1 m Color Difference 1998 1998

Pan Color1997 1998 New 1997 1998 New 1997 1998 New

1 56 158 102 50 113 63 60 158 98 98 96 992 50 140 90 47 128 81 50 149 99 104 98 953 45 143 98 44 130 86 50 146 96 97 98 974 50 150 100 38 122 84 52 148 96 99 98 1025 44 147 103 50 148 98 50 147 97 98 98 986 48 148 100 50 140 90 49 149 100 99 96 987 50 148 98 48 156 108 50 148 98 99 98 988 54 137 83 39 110 71 49 146 97 98 103 999 51 140 89 42 128 86 50 148 98 98 99 98

Median 50 147 98 47 128 86 50 148 98 98 98 98Average 49.8 145.7 95.9 45.3 130.6 85.2 51.1 148.8 97.7 98.9 98.3 98.2Actual 50 148 98 50 148 98 50 148 98 98 98 98

Table 2 Results from visual interpretation and counting of buildings within the residential sub-area

Table 3 Results from visual interpretation and counting of buildingswithin the industrial/commercial sub-area

Interpreter 1 m Pan 5 m Pan Difference MaskedImage 1998 Pan

1997 1998 New 1997 1998 New1 14 22 8 14 21 7 8 72 12 16 4 12 18 6 7 53 12 16 4 12 17 5 5 54 12 17 5 12 22 10 18 55 12 18 6 12 17 5 6 66 10 15 5 12 16 4 7 57 12 17 5 12 18 6 10 58 13 17 4 12 15 4 5 59 7 14 13 6 11 5 11 11

Median 12 17 5 12 17 5 7 5Average 11.6 16.9 5.3 11.6 17.2 5.8 8.6 6Actual 12 17 5 12 17 5 5 5

Semi-automated ApproachesGiven the desire by commercial and government

organizations to automate the building identifications andenumeration processes, much effort was expended ondeveloping and testing more automated approaches.Approaches were developed and tested for the followingimages and sub-areas: a) 1 m color images of industrial/commercial sub-area; b) 1 m and 5 m Pan images of industrial/commercial sub-area; and c) 1 m color change-vector imagesof residential sub-area.

The complexity of a scene directly affects the selectionand success of an automatic procedure for detecting buildingfeatures. Three different approaches for detecting andenumerating buildings from the three types of images listedabove. More complete descriptions of the procedures andtheir advantages and limitations follow.

Unsupervised classificationFor the 1 m color images of industrial/commercial sub-

areas, a per-pixel unsupervised classification was appliedto both 1997 and 1998 images. First, unsupervisedclassification was applied to the 1 m 1997 and 1998 colorimages. Ten clusters were specified using the iterativeself-organizing (ISODATA) clustering routine. The outputcluster images from unsupervised classification weredisplayed and assigned to building roof or non-roof classes,and then recoded to 0 and 1, respectively (Figures 7a and7b). After applying a 5 x 5 majority filter to both 1997 andl 998 classified binary images, the 1997 classified roofimage was subtracted from the 1998 image. Finally, pixelclumping and sieving was applied to eliminate small pixelclumps (presumably noise) from the subtracted image.The apparent building objects received a unique identifierand the number was utilized as an estimator for the numberof new buildings. Figure 7c shows the final vector GISlayer depicting new building locations overlaid on 1998image.

Thresholding and edge detection methodA procedure involving thresholding and edge detection

process was applied to both 1 m and 5 m resolutionpanchromatic images of the industrial/commercial sub-area. For this approach, an edge detection using Sobelfilters was applied to both 1997 and 1998 images. Twobinary edge images of 1997 and 1998 were first createdusing thresholds to delineate edges. Thresholds wereselected interactively so that only distinct edges such asbuilding and road edges were enhanced. A 3 x 3 maximalfilter was applied to binary edge images to insure that theedges around buildings were connected. Then the binaryimages were vectorized to create building polygons. Theedge polygons of other features were eliminated by settingthe minimum and maximum polygon area thresholds basedon commercial/industrial building sizes. Finally, the

78

Figure 6 A series of image products used for multitemporal interpretation tests for commercial subarea. The ground length ofthe sides of the square subarea is about 860 m

Figure 7 Images obtained from unsupervised classification for 1 m color image of the industrial/commercial sub-area

number of buildings in 1997 and 1998 were compared todetermine the number of new buildings built over theintervening year.

Supervised classificationFor residential areas and the l m color imagery, we

found that the unsupervised classification with 40 clusterswas not capable of distinguishing the spectral-brightnessvariability of the residential sub-area. Also, the difficultyof labeling more than 40 clusters led us to abandon theunsupervised approach and test a per-pixel supervisedclassification approach. Supervised classification wasapplied to two types of images:1) l m 1997 and 1998 color images; and

2) l m three-band change vector difference image. Theflowchart in Figure 8 shows the basic steps of thisprocedure.Six different classes were used in the training: two types

of roofs, two types of roads, vacant/cleared land, andvegetation. Groups of contiguous pixels were selected astraining samples to preserve spatial structures in the classsignatures. Figures 9a, 9b and 9d present the binary imagesof residential/nonresidential roofs after recoding theclassified images into ‘building’ and ‘non-building’categories from above two types of images. It is obviousthat the accuracy of the classification product derived fromthe change vector difference image is much lower thanindividual l997 and l998 color images. Since the

79

Figure 8 Flowchart for supervised routine

classification accuracy can be influenced by the selectionof training data, two sets of training data were used to testthe sensitivity of classification results to training data for1998 image. Figure 9c presents the 1998 binary imageusing another set of training data.

Upon visual comparison of the supervised classifiedimages of buildings, high commission and omission errorsbetween roofs, roads and vacant lands were evident. It isapparent that the classification accuracy from changevector difference image is much lower than individual1997 and 1998 color images. No further processing wasapplied in classified image generated from change vectorimage.

After applying a 3x3 majority filter to both 1997 and1998 roof images, and subtracting the binary roof images of1997 from that of 1998, images of new buildings wereobtained and are presented in Figure 9e. Several majority,minimum, and maximum filters were tested on the multi-temporal building change image to separate individualbuildings. The apparent roof features that were larger thanthe specified minimum size were vectorized and eachapparent building object received a unique identifier. The

maximum identifying polygon number was utilized as anestimate for the number of residential units.

Results and discussion

General change detectionIt can be seen from Figure 4 that visually multitemporal

difference images having 1 m, 5 m, and 10 m spatialresolutions are similar and many change features arehighlighted similarly in the general study area. However,detailed inspection of individual change building featuresreveals that boundaries of most buildings are not preciselydelineated at any of the three spatial resolutions. Thisproblem is largely associated with misregistration errors.The ability to distinguish individual buildings at eachresolution depends on the building size. It is impossible toidentify individual residential buildings within the studyarea at 10 m resolution image. Large commercial buildingscan be identified at all three resolution images, whereasindividual residential housing units were identified at 1 mand 5 m.

The initial aim was to generate an accurate changemask by applying change/no-change thresholds and

80

Figure 9 Residential features detected by supervised classification

spatial aggregation filters to extract building changeareas for the whole study area. Several difficulties wereencountered and potential difficulties were identified inachieving this aim. First, it can be difficult to overcomethe effect of shadow on change detection, especiallywhen high resolution images are used in an urban setting.Although in this study the images were acquired withsimilar solar illumination angles and thus, were minimallyinfluenced by differences in shadows, high-resolutionimage can be significantly affected by shadows if theimages are acquired with different solar geometrycharacteristics. The second difficulty was to determineoptimal or even acceptable threshold levels to delineatechange areas from multitemporal difference images.Another issue was misregistration between multi-temporal images, particularly those with 1 m spatialresolution. Although a more rigorous registrationapproach may provide greater accuracy, the excessivetime and associated computation costs makes itunwarranted for the change detection and stratificationprocess over extensive areas. For the general study area,visually identifying and manually delineating larger,contiguous areas of change based on differencing imageswas more efficient and reliable than the automatic change

detection and masking approach for the general studyarea.

Visual interpretationTables 2 and 3 summarize the visual interpretation results

obtained for the nine interpreters. Single pass countingusing the masked second-date Pan and color and differenceimages yielded equal or higher visual counting accuraciesthan counting for individual masked, multi-temporal images.When using the single-pass approach for the residentialsub-areas, all three of these images yielded building countsthat were essentially identical. For the industrial/commercialsub-area, an image change feature, the completedconstruction of a courtyard within a building complex, wasfrequently counted as a building for all approaches. However,other new construction features also caused confusion suchas new rectangular-shaped parking lots which have similarspectral-brightness changes and were falsely identified asnew buildings. Also, misregistration substantially impactedthe results of interpreting difference images. For residentialareas, visual counting accuracy was reduced significantlywhen using 5 m resolution images, where it is difficult toidentify individual building units. However, for largerindustrial/commercial buildings, similar accuracies were

81

Figure 10 Demonstration of point coverage suggested for reducing errors andincreasing efficiency in visual counting. Each building receives a manuallydigitized point as the analyst interprets. An automatic routine counts thetotal number of points

Number ofnew buildingsautomaticallyenumerated

6

6

6

89

Table 4 Most successful semi-automated approaches and their results

Sub-area and imagetype

Industrial/commercialsub-area – 1 m color

Industrial/commercialsub-area – 1 mPanchromatic

Industrial/commercialsub-area – 5 mPanchromatic

Residential sub-area -1 m color

Automaticapproach

Unsupervisedclassification

Edge detectionand thresholding

Edge detectionand thresholding

Supervisedclacsification

Actual numberof new

buildings

5

5

5

98

achieved for both l m and 5 m resolution images.Visual counting seems to be a very simple and

cost-effective method for identifying buildingswithin relatively small study areas. However,when a large number of new buildings occur overa large areal extent, manual counting may beprone to increased errors. The analyst may losetrack, over-count, or fail to interpret and countcertain portions of a development. A techniquefor reducing errors in manual counting is to createa point layer by digitizing a point on each building(or other microfeature) as the analyst interpretsand counts (Figure 10). After each new buildinghas received a manually digitized point, a routineautomatically counts the total number of points,which should correspond to the number of newbuildings. In this way, the analyst is able tocontinue after an interruption or loss ofconcentration while in the middle of the processof interpreting and counting new buildings.

Semi-automated building detectionThe approaches, image types, and associated

building count results for the most successfulsemi-automated approaches are listed in Table4. Application of these semi-automated imageprocessing procedures were only partiallysuccessful in precisely delineating relativelylarge features such as commercial buildings.Compared to the results in Table 2 and 3, it isapparent that semi-automated approaches wereless accurate for generating information on thenumber of new building structures built in theprevious year than visual-based interpretationand enumeration, especially for residentialbuildings. The estimates for the number of newresidential buildings in the unsupervisedclassification procedure were found to besensitive to a) training data, b) the window sizeused in majority, minimum, and maximumfilters, c) the number of passes for applying thefilter, and d) the threshold size used to eliminatenoise and small units. The best estimate fromsemi-automatic approaches in our test was 89new residential buildings, with the actualnumber being 98. All of the semi-automatedapproaches resulted in an estimate of six newindustrial/commercial buildings in both 1 mand 5 m Pan images, which is one more thanthe actual number. One building was underconstruction in 1997 and completed by the1998 image date. For the visual counting, thisbuilding was generally not counted. But for theautomatic approach, a new building wascounted since a large portion of the buildinghad changed in surface material compositionand was identified as a new microfeature.

Summary and Conclusions

Providing information on the location, type and associated featuresof new building structures is important to policy and decision makersfrom both public and private sectors. This research focuses on pilottests of image-based methods for detecting and counting microfeaturechanges. Three semi-automated and two interactive image-basedapproaches to detecting and enumerating new building structures fromhigh spatial resolution, multi-temporal image data were performed fortwo study areas (one residential, one commercial) in the northwestportion of the City of San Diego. We investigated: a) the performanceof l m panchromatic versus l m color images, b) the performance ofimages with different resolutions ( l m versus 5 m, 10 m), c) techniquesfor detecting general land use change, and d) techniques for detecting

82

and counting new buildings.Based on this specific context, the following conclusions

were reached:• Prior to detecting and enumerating new microfeatures,

it may be useful to stratify images into apparentchange and no-change strata, using simple changedetection techniques. Simple image differencing waseffective to visually identify change areas. However,automated change masking techniques were notsufficiently reliable so as to warrant their use as thesole basis for stratification.

• For visually identifying change areas in support ofsubsequent detailed building enumeration, 10 mresolution image data provided similar results as 5 mresolution in the southern California context.However, 5 m and 10 m resolution image data are toocoarse for detecting and enumerating individualresidential buildings and 1 m resolution is sufficient.

• Interactive, visual-based interpretation andenumeration was the most accurate and efficientapproach for generating information on the numberof new building structures built in the previous year.

• Semi-automated approaches to detecting andenumerating microfeatures: 1) require multipleprocessing steps that end up being very costly unlessapplied over very large extents, 2) may be useful asenhancements for visual-based assessments ifsimplified processing steps are implemented, and 3)may be practical in areas composed mostly of largebuildings associated with commercial and industrialland use.

• Panchromatic images (single visible wavebandportrayed in black and white) with 1 m spatialresolution from airborne and future commercialsatellite sources should be suitable for delineating alltypes of buildings. Color imagery provides someadditional discrimination benefits for residentialbuildings, which are likely to be manifested in greaterefficiency of interpretation rather than as significantimprovements in accuracy.

• The precision of geometric registration of multi-temporal images was critical to the success of bothvisual and automated identification of newmicrofeatures, when based on multitemporal layerstacks. If misregistration is substantial, then it maybe necessary to implement approaches that are basedon processing each date of imagery separately; theseapproaches tend to be less efficient. When imageregistration is precise, the most effective and efficientimage type for visually detecting and countingindividual structures of all building types was a multi-temporal difference image.

Acknowledgments

This research was funded by and conducted as part of the

NASA Affiliated Research Center program underCooperative Agreement NCCB-16. John Kaiser coordinatedthis project and helped to generate some figures. Specialthanks go to Oscar Weser, former president of AerialFotobank, Inc., for his suggestions and support for thisproject. The authors would like to thank the reviewers fortheir valuable comments and suggestions.

References

Cross, A.M, and D.C:. Mason, 1988. Segmentation of remotely sensedimages by a split-and-merge process. International Journal ofRemote Sensing, 9(8):1329-1345.

Ehlers, M., M. A. Jadkowski, R.R. Howard, and D.E. Brostuen, 1990.Application of SPOT data for regional growth analysis and localplanning. Photogrammetric Engineering and Remote Sensing,56(2): 175-180.

Estes, J.E., and Stow, D., and J.R. Jensen, 1982. Monitoring land useand land cover changes, In Remote Sensing for ResourceManagement (C.J. Johannsen and J.L. Sanders, editors), Iowa:Soil Conservation Society of American, pp: 100-110.

Forster. B.C., 1983. An examination of some problems and solutionsin monitoring urban areas from satellite platforms, InternationalJournal of Remote Sensing. 6(1): 139-151.

Fung, T., and E. LeDrew, 1987. Application of principal componentsanalysis for change detection, Photogrammetric Engineering andRemote Sensing, 53(12):1649-1658.

Gong, P. and P.J. Howarth, 1992. Frequency-based contextualclassification and gray-level vector reduction for land-useidentification. Photogrammetric Engineering and Remote Sensing,58:423-437.

Gordon, S.. 1980. Utilizing Landsat imagery to monitor land usechange: A case study of Ohio. Remote Sensing of Environment, 9:189- 196.

Green, K., D. Kempka, and L. Lackey, 1994. Using remote sensing todetect and monitor land-cover and land-use Change.Photogrammetric Engineering and Remote Sensing, 60(3): 331-337.

Howarth, J.P., and G.M. Wickware, 1981. Procedure for changedetection using Landsat digital data. International Journal ofRemote Sensing, 2:277-291.

Howarth. J.P., and E. Boasson, 1983. Landsat digital enhancement forchange detection in urban environments, Remote Sensing ofEnvironment, 13 (2): 149- 160.

Huertas, A., and R. Nevatia, 1988. Detecting buildings in aerialimages. Computer Vision, Graphics and Image Processing, 41:131-152.

Irvin, R.B., and D.M. McKeown, Jr., 1989. Methods for exploiting therelationship between buildings and their shadows in aerial imagery,IEEE Transactions on Systems, Man and Cybernetics, 19(6): 1564-1575.

Jensen, J.R., and D.L. Toll, 1982. Detecting residential land-usedevelopment at the urban fringe. Photogrammetric Engineeringand Remote Sensing, 48(4):629-643.

Jensen, J.R., D.J. Cowen, J. Halls, S. Narumalani, N.J. Schmidt, B.A.Davis, and B. Burgess, 1994. Improved urban infrastructure mappingand forecasting for BellSouth using remote sensing and GISTechnology, Photogrammetric Engineering and Remote Sensing,60(3): 339-349.

83

Lambin, E.F., and A.H. Strahler, 1994. Change-vector analysis inmultitemporal space: A tool to detect and categorize land-coverchange process using high temporal-resolution satellite data.Remote Sensing of Environment, 48:231 -244.

Lo, C.P., 1995. Automated population and dwelling unit estimationfrom high-resolution satellite images: a GIS approach. InternationalJournal of Remote Sensing, 16(1).17-34.

Sahar. L. and A. Krupnik, 1999. Semiautomatic extraction of buildingoutlines from large-scale aerial images. PhotogrammetricEngineering and Remote Sensing, 65(4):459-465.

Shi, Z., and R. Shibasaki, 1996. Towards automated house detectionfrom digital stereo imagery for GIS database revision. InternationalArchives of Photogrammetry and Remote Sensing, 31 (B4):780-785.

Singh, A., 1989. Digital change detection techniques using remotely-sensed Data. International Journal of Remote Sensing, 10(6): 989-1003.

Stow, D.A., D. Collins and D. McKinsey, 1990. Land use changedetection based on multi-date imagery from different satellitesensor systems. Geocarto International, (3): 3-12.

Wang, F., 1993. A knowledge-based vision system for detecting landchanges at urban fringes, IEEE Transactions on Geoscience &Remote Sensing, 31: 136- 145.

Wang, J., 1987. Sequential thinning algorithms for remote sensingapplications. Proceedings, IEEE International Geoscience andRemote Sensing Symposium edited by M.C. Dobson, Ann Arbor,Michigan, USA: Univ. Of Michigan, pp:337-342.

Weidner, U., and W. Forstner, 1995. Towards automatic buildingextraction from high resolution digital elevation models, ISPRSJournal of Photogrammetry and Remote Sensing, 50(4):38-44.

84

GEOMania ProfessinalIntroduction

GIS/CAD data inputtig and editing tool

Main Functions

To: Geocarto International CentreG.P.O. Box 4122, Hong Kong Date:Tel: (852) 2546-4262 Fax: (852) 2559-3419E-mail: [email protected] Website: http//www.geocarto.com

q Encolsed is my/our check/bank draft for US$1,495 (HK$11,660) for one GEOMania Professional.

Name (Mr. Mrs. Miss/Dr.) / Institution:

Address:

City: State: Country: Code:

Tel: Fax: E-mail: Signed:

GEOMania is developed by GEOMania Co., Ltd. Geocarto International Centre is authorized foregin reseller.

GEOMania Professional is the product which containspowerful GIS and CAD function and provides convenientand various functions.

Especially, it supports the strong map editing functionssuch as Rubber Sheeting, Edge Matching, AutomaticPolygon Creation, and Line Error Correction. It also cancreate Network Topology, analyze the shortest path andgenerate isochrone.

Readjust Map Coordinate- Moving function by inputting the X

and Y coordinate values and a basecoordinate position after selectingmap scale

Edge Matching- Automatically or manually correct the

errors among contiguous maps

Creating polygons automatically- Transfer Line Object into polygon for

convenient inputting & editing

Rubber sheeting- Moves and change scale, shape of

map, coordinate base on convertingpoint. It can be applied for one part orwhole map

Input data collection- Input the same value in selected

area collectively

Text Filtering- Input a text layer into attribute data in

polygon layer

Clipping- Clip a specific area from map of screen

Create the thematic map- Classify certain values of fields from

map data and visualize different coloraccording to the value

Print Map- Print a map with coordinates,

legends, scales, directions and indexmaps

Command Window- Support CAD command for CAD user

in command window

Create Network Topology- Create Network Topology and

analyze the shortest path andgenerate isochrone