Illinois State Water Survey Division SURFACE WATER SECTION SWS Contract Report 487

UPDATING LAND USE CLASSIFICATIONS OF URBANIZED AREAS IN NORTHEASTERN ILLINOIS BY USING SPOT AND TM SATELLITE DATA

by Ming T. Lee and Ying Ke

Prepared for the Illinois Department of Energy and Natural Resources

Champaign, Illinois February 1990

UPDATING LAND USE CLASSIFICATIONS OF URBANIZED AREAS IN NORTHEASTERN ILLINOIS BY USING SPOT AND TM SATELLITE DATA

by Ming T. Lee and Ying Ke

Illinois State Water Survey 2204 Griffith Drive Champaign, IL 61820-7495

Prepared for the Illinois Department of Energy and Natural Resources

February 1990

CONTENTS Page Abstract

1

Introduction Study Area Data Used Acknowledgments

1 2 2 4

Literature Review

4

Preliminary Evaluation of SPOT Imagery

6

Determination of Textural Features .................................................................... .........

6

Image Processing Methods Pre-Processing Image Classification Post-Processing

8 8 8 9

Results

10

Summary and Conclusions

11

References

23

UPDATING LAND USE CLASSIFICATIONS OF URBANIZED AREAS LN NORTHEASTERN ILLINOIS BY USING SPOT AND TM SATELLITE DATA by Ming T. Lee and Ying Ke

ABSTRACT Urban water quality assessment requires updated and detailed information on land use, pervious and impervious areas, soil, and storm sewer network features such as pipe size, detention basins, slopes, roughness, and other parameters. A research project was conducted to incorporate SPOT and TM satellite images in updating Geographic Information System (GIS) land use databases for rapidly urbanized areas in northeastern Illinois.

Both spectral and

textural features of SPOT and TM images were used for land use classification. TM images from 1985 and SPOT images from 1987 and 1988 were used.

A procedure was tested for

linking U.S. Geological Survey land use digital analysis (LUDA) data (1973) and 1:100,000 scale digital line graph (DLG) transportation coverage (1983) on the ARC/INFO system, and classified SPOT and TM files on the ERDAS system. Urbanized areas during 1985 through 1988 were delineated, and the results were verified by using recent detailed aerial photographs. This updated information is useful for regional urban runoff water quality modeling.

INTRODUCTION A land use/cover classification for a region in northeastern Illinois was developed by King et al. (1989). Their report indicated that maps classified on the basis of thematic mapper (TM) data, although less accurate than U.S. Geological Survey (USGS) land use digital analysis (LUDA) maps, are helpful in updating regional land use/cover information at a small cost (10 percent or less of the cost of the LUDA mappings).

The updated information can aid in

identifying areas that have undergone rapid changes. In some cases, the TM data do not have a high enough resolution to delineate urban areas. To correct this shortcoming, a methodology was developed for using SPOT panchromatic satellite data with 10-meter spatial resolution to update the data on recently urbanized areas in northeastern Illinois.

1

Study Area The region for which King et al. (1989) developed a land use/cover classification consists of a 36-township area as shown in figure 1. Because of the limitations of some of the 1987 SPOT data and the unavailability of USGS digital line graph (DLG) transportation data for this region, our present study area is limited to six townships in Kane County (shaded area in figure 1). The Illinois Department of Energy and Natural Resources published an environmental screening atlas for the study area (Hines, 1986). The atlas includes information on: 1) The physical setting of the study area, the model ring for the proposed Superconducting Super Collider, topography, soils, distribution and thickness of quaternary deposits, sand and gravel aquifers, bedrock geology, structural features, seismic risk, and flood hazard areas. 2) Conservation and preservation of threatened and endangered species, wetlands and water bodies, natural areas, and known paleontological sites. The present report, prepared as part of an environmental database study, describes the methods used in detecting recently urbanized areas within the six-township area on the basis of SPOT satellite data and USGS DLG data. Data Used Four kinds of data are available for the study area: 1) U.S. Geological Survey land use digital analysis (LUDA) maps. 2) Land use maps produced from 1985 landsat thematic mapper data. 3) Two SPOT panchromatic data sets (from June 13, 1987, and September 16, 1988, referred to here as SPOT87 and SPOT88). SPOT panchromatic images have high spatial resolution but only one spectral band. The spatial resolution is 10 meters, and the wavelength is from 0.51 to 0.73 micrometer. The processing level for SPOT87 is level 1B (having radiometric and geometric corrections), and the level for SPOT88 is level 1A (having radiometric corrections only). 4) 1984 USGS DLG transportation data for Kane County, Illinois. These data were digitized from 1:100,000 scale maps.

2

Figure 1. Region for which a land use/cover classification was previously developed (Shaded area represents the study area for the current study)

3

Acknowledgments This study was funded by the Illinois Department of Energy and Natural Resources as part of an extensive project. Dr. K.P. Singh is the Water Survey project coordinator. Becky Howard typed the camera-ready copy of the report, and Gail Taylor edited the report.

LITERATURE REVIEW Detection of newly developing urban areas is one of the major applications of remote sensing. Jensen (1983) summarized remote sensing applications in urban/suburban land use analysis. Extensive research on land use change detection has been reported by Martin (1989), Todd et al. (1978), Estes and Simonett (1975), Swain and Davis (1978), Jensen (1985), and Place (1974). Jensen (1985) discussed different approaches to detection of urban change. The conventional method is to use visual comparisons of photographs taken at different times, but this approach is relatively time-consuming and is subject to numerous omission errors. Therefore new approaches are needed for performing automated change detection. Jensen (1985) pointed out that a remote sensing system used for urban change detection should have the following characteristics: 1) Images that are compared should have the same coverages. 2) Images for the same geographic area should be recorded at the same season (about the same day of the year) so that they have similar sun-angle effects. 3) Images for the same area should have the same scales and geometric characteristics. 4) Distortions caused by topographic relief should be removed as much as possible. 5) Images should have consistent and useful spectral regions. Recent remote sensing systems on satellites have been designed to meet these requirements. Jensen (1985) indicated that to achieve high accuracy, the spatial registration of the two images compared should be within one-fourth to one-half of a pixel. He emphasized that optimum dates should exist for a specified area and that land use changes from non-urban to urban areas can be detected on the basis of differences in the spectral response of pixel records on two different dates. Change detection approaches can be classified into four groups: 1) image differencing; 2) image ratioing classification; 3) post-classification comparison; and 4) comparison of preprocessed imagery. Each approach requires a particular kind of information. The selection of an appropriate approach is based on 1) the spatial and spectral characteristics of the data; 2) the cultural and physical characteristics of the study area; and 3) the advantages and

4

disadvantages of each approach.

The following sections describe the procedures for each

approach. Image Difference Approach (Christenson et al., 1978; Royal, 1980).

The image

difference approach involves subtracting one image from another image to create a new image. Because two images cannot have the same spectral characteristics, the critical part of this approach is to find the threshold value that is used to separate changed and unchanged pixels. Image Ratioing Approach (Kriegler et al., 1969; Todd, 1977; Robinson, 1979).

The

image ratioing approach involves computing the ratio of two spectral images to obtain a composite image. This increases the accuracy of urban change detection by reducing the effects caused by environmental factors and by variations in sensor systems. Post-Classification Comparison Approach (Weismiller et al., 1977; Toll et al., 1980). The post-classification comparison approach, which encompasses many different methods, distinguishes changes by comparing two classification maps produced from two data sets. One map, produced from early image data, is considered the standard land use map.

The

comparison between this map and the map produced from later images can show the change between each pair of land use types.

In the spectral/temporal change classification method,

two images are combined into a multiple-date data set, and a single classification is performed on this data set. The changed area should have very different spectral responses from the unchanged area. Pre-Processing Comparison Approach (Swain and Hauska, 1977; Jensen, 1978; Weismiller et al., 1977). This approach also encompasses many different methods. The layered spectral/temporal classification method uses multistage decision logic during the classification. An unclassified pixel can be classified into a class type according to one or more different decision features. Analysts should have full knowledge of the study area, and special programs should be designed for each specific area. In the clustering comparison method, a clustering classification is used to identify nonurban and urban areas (including developing areas). Results of previous research (Irons, 1981; Haralick, 1979) showed that the approach worked only in some specific areas because the spectral data may not provide sufficient information to detect change. Some data preprocessing techniques such as the use of high-frequency filters (Royal, 1980); low-frequency filters (Maxwell, 1976; Nagao and Matsuyama, 1979); brightness images (Jensen, 1980; Purcell, 1978); and principal component analysis (Walthall and Knapp, 1978; Friedman, 1980; Toll, 1980) are used to detect land use changes.

5

PRELIMINARY EVALUATION OF SPOT IMAGERY Because SPOT panchromatic images have only one spectral band, the common multispectral classification method could not distinguish land cover types very well, and the image ratioing approach also could not be used. Because the different data sets used were not obtained on the same day of the year, they show significant differences in reflectance caused by differing environmental factors. Because the image differencing and ratioing approaches are not appropriate and because early land use maps are available, the post-classification comparison approach was used in this study. Visual interpretation of the SPOT data revealed the following characteristics of nonurban and urban cover (including developing areas) in the study area: 1) Residential areas have significant variations in spectral reflectance. 2) The spectral characteristics of grassy areas and agricultural fields are not significantly different. 3) Large industrial areas, parking lots, and commercial areas have relatively low variations in spectral reflectance, but they have very high spectral reflectance. 4) Most newly developed areas have relatively high spectral reflectance. 5) Agricultural areas have relatively low spectral reflectance and relatively low variations in spectral reflectance. 6) The spectral characteristics of the SPOT88 data are better than those of the SPOT87 data for distinguishing urban and nonurban areas. Because of the above characteristics, the urbanized areas could not be accurately identified by using either spectral or textural characteristics. Therefore a classification was designed to use both spectral and textural information.

DETERMINATION OF TEXTURAL FEATURES Texture, which is defined as the characteristic placement and arrangement of spectral tones, is an important factor in the interpretation of aerial photography. Spectral characteristics are related to the amount of light reflected or emitted from a surface. Textural characteristics refer to the impression of roughness or smoothness caused by variations in tone or repetitions of visual patterns across a surface. Generally, information on both spectral and textural features is needed to distinguish different features, especially in very complex pictures (Haralick, 1979). To use textural concepts in digital data analyses, textural information should be extracted from reflectance measurements and converted into digital format. No precise 6

definition or mathematical formula exists for quantifying image texture. Most textural analysis describes the spatial distribution of gray levels across an image segment (Irons, 1981). The most important texture measurements can be derived from the distribution of local characteristics. Hsu (1978) listed eleven local characteristics that were computed by moving 3by-3 or 5-by-5 pixel windows through an image. The change detection procedure used in this study consists of three steps, the first two of which are performed in parallel. The first step is to use spectral information. A threshold value is defined visually through investigation of several test areas. If the reflectance of a cell is larger than the threshold value, the cell is considered to be a candidate for the presence of urban or developing areas; otherwise, it is categorized as a nonurban area. In the second step, textural information is used. If the value of texture transform for a cell is larger than the threshold value, the cell is considered an urban area. The third step is to use the results from the first two steps. If a cell is considered as an urban area in either or both previous results, it is considered to be an urban or developing area. The texture transform used in this study is the variance contained in a 3-by-3 pixel window that is moved through the image. The texture transform is computed by a TEXTURE command in the ERDAS system (ERDAS, Inc., 1989). Although the high resolution of the SPOT panchromatic data improves the accuracy of land use classification, it also creates new problems because the data have only one spectral band. Digital data represent average reflectance values in certain geometrical areas and in certain spectral wavelengths. For example, grassy areas inside an urban area may have similar spectral characteristics to an agricultural area; but house roofs and surfaces of roads, walkways, or parking lots have very different spectral characteristics from agricultural cover. When the spatial resolution of the remote sensor is low, large ground areas are converted into a single measurement unit. The units in urban areas differ according to grassland and other land cover types, and the unit in agricultural areas is uniform. These two types of land cover areas should have distinguishable spectral characteristics, but they may be significantly misclassified along their boundaries, especially when the boundaries are very complex. The high-resolution SPOT data provide an opportunity for more accurate classification of the areas along the boundaries, and when the spatial resolution increases, many large plots of grassland inside urban area can distinguished from roof surfaces. This is good for studying the land use structure of urban areas. However, these data are not good for classification if the difference between grassland and agricultural areas cannot be distinguished spectrally. In this study, the spatial resolution of the SPOT data was very high, so many detailed urban characteristics could be distinguished. For example, a single house in an agricultural area could be identified. In urban areas, the rooftops of houses and the surfaces of roads, 7

walkways, and parking lots sometimes could be distinguished from grassy areas. But it was difficult to distinguish the grassland in an urban area from that in an agricultural area because of the similarity of spectral characteristics. As mentioned before, the high resolution of the SPOT data increases the accuracy of classification along the boundaries, but decreases the accuracy of areal identification. The structure of urban land use is not the main object of interest in this study. To increase the accuracy of classification, the data should be run through a low-pass filter before each subclassification. It is obvious that the texture transform and low-pass filter will artificially extend hightexture (high-variance) and high-reflectance areas into low-texture and low-reflectance areas. To reduce this kind of misclassification, a special program was designed to correct the artificial extensions. Results of the above processing show that some grassy plots and shadows of trees and houses in urban areas were classified as nonurban areas, and some road surfaces were classified as nonurban areas. But the results also show that newly developing areas - which are the areas of interest in this study - were well identified. A land use map classified from 1985 landsat TM data was combined with USGS DLG data to form a land use base map. A new land use map can be derived by overlaying the results from the SPOT data on the land use base map. It is assumed that the urban change is only from agricultural areas to urban areas. The results show that most urban change areas have been detected by this approach.

IMAGE PROCESSING METHODS For this study, image processing was summarized into three steps. Pre-Processing 1) The SPOT panchromatic data were run through a low-frequency filter (ERDAS command: CONVLV). This process enhances the homogeneity. 2) The SPOT panchromatic data were processed through a 3-by-3 window by defining the texture as the variance within the window. A low-frequency filter was used to enhance the homogeneity. Image Classification 1) The observation of SPOT spectral data showed that low values will indicate agricultural land, urban parks, and residential lawns. High spectral areas will indicate newly developed commercial and service areas. On the basis of these properties, a threshold value was selected to delineate these two groups. 8

2) Observations of SPOT textural data show that urban residential areas and newly developed urban areas have high textural values (defined here as variances in 3-by3 pixel windows). Agricultural land has very low textural values. After inspection of the histographs of the textural data, a threshold value was selected to delineate these two groups. Post-Processing 1) To combine the results of textural and spectral groups, a decision table was set up. Only land low in both spectral and textural areas will be classified as agricultural. The rest of the area will be classified as urban lands. The decision table is as follows:

TEXTURAL VALUES High

Low

SPECTRAL VALUES High Low 2) Because the calculation of the variance artificially created high textural values on urban-agricultural fringes, a procedure was used to define the approximate zones within 2 pixels from the fringes. The ERDAS command for this operation is SEARCH. The areas recorded within these zones were reclassified as agricultural (ERDAS command: RECODE). 3) As stated before, the 1985 land use map was developed from thematic mapper (TM) data (King et al., 1989). This map was categorized into six classes (urban and five others). Most of the urbanized areas had been converted from agricultural lands to urban areas. The combination procedures are as follows:

9

1985 LAND USE MAPS Urban

Non-urban (5 classes)

SPOT DATA Urban Non-urban 4) The results from the last table were then combined with USGS DL6 data in the following way: DLG DATA Road

Non-road

Non-urban New-urban Old-urban

RESULTS Figure 2 shows the 1985 land use for the six-township area in Kane County, derived from the TM data of King et al. (1989). Figure 3 shows the 1987 SPOT panchromatic image of the same area. To illustrate the details of this image, a portion of it was enlarged as shown in figure 4. In this figure, the streets in newly urbanized areas have a high reflectance; agricultural areas are relatively dark and have less variation of reflectance. Major highways and commercial developments have high reflectance. Figure 5 shows the same image with an overlay of DLG transportation data. The transportation data were created by buffering the original line data according to the widths of the different street and highway classes. Most of the streets were matched with the urbanized areas. In figure 6, the 1985 urban areas based on the TM data are shown by the color red, and the newly urbanized areas are shown by green. As the figure illustrates, the newly urbanized areas are adjacent to the old urban areas, which is

10

the normal pattern of urban expansion. Some small tracts in old urban areas were shown as newly urbanized areas, which may be due to omissions in the 1985 TM data because of lower spatial resolution.

SUMMARY AND CONCLUSIONS The experience gained in this project leads to the following conclusions: 1) The approach used in this study is efficient and inexpensive because the algorithm is not complicated. 2) The SPOT data proved to be useful for updating the classification of urbanized areas. 3) Additional accuracy can be achieved by using SPOT data obtained in August or September, when agricultural areas are covered by high levels of vegetation. 4) SPOT processing level 1B should be selected because level 1A had view-angle distortion. This distortion is not linear, and it is difficult to remove it by using the geometric correction method.

11

Figure 2. Land use classification map based on 1985 thematic mapper (TM) data

13

REFERENCES Christenson, J.W., D.L. Dietrich, H.M. Lachowski, and M.L. Stauffer. 1978. Urbanized Area Analysis Using Landsat Data. Paper presented at the International User's Conference on Computer Mapping, Cambridge, MA, July, 21 p. ERDAS, Inc. 1989. ERDAS User's Guide, Version 7.2. Atlanta, GA. Estes, J.E., and D.S. Simonett. 1975. Fundamentals of Image Interpretation. In Manual of Remote Sensing, L. W. Bowden and E. L. Pruit, eds., Fall Church: American Society of Photogrammetry, pp. 869-1076. Friedman, S.Z. 1980. Mapping Urbanized Area Expansion Through Digital Image Processing of Landsat and Conventional Data. Jet Propulsion Laboratory, Pasadena, CA, Publication 79-113 (March), 90 p. Haralick, Robert M. 1979. Statistical and Structural Approaches to Texture. Proceedings of the IEEE, Vol. 67, No. 5, pp. 786-804. Hines, J.K., ed. 1986. Siting the Superconducting Super Collider in Northeastern Illinois: Environmental Screening Atlas. Illinois State Geological Survey, Champaign, IL. Hsu, S.

1978. Texture-Tone Analysis for Automated Land-Use Mapping. Engineering and Remote Sensing, Vol. 44, pp. 1393-1404.

Photogrammetric

Irons, J.R. 1981. Texture Transforms of Remote Sensing Data. Remote Sensing Environment, Vol. 11, pp. 359-370. Jensen, J.R. 1978. Digital Land Cover Mapping Using Layered Classificaiton Logic and Physical Composition Attributes. The American Cartographer, Vol. 5, pp. 121-132. Jensen, J.R. 1980. Urban Area Change Detection Procedures with Remote Sensing Data. Final Report, contract NAS 5-26129, Goddard Space Flight Center, Greenbelt, MD, 50 p. Jensen, J.R., ed. 1983. Urban/Suburban Land Use Analysis. Remote Sensing, American Society of Photogrammetry.

Chapter 30 of Manual on

Jensen, J.R. 1985. Urban Change Detection Mapping Using Landsat Digital Data. The Surveillant Science: Remote Sensing of the Environment, Robert K. Holz, ed., John Wiley & Sons. King, R., M.T. Lee, and K.P. Singh. 1989. Land Use/Cover Classification for the Proposed Superconducting Super Collider Study Area, Northeastern Illinois. Illinois State Water Survey Contract Report 458. Kriegler, F.J., W.A. Mailila, R.F. Nalepka, and W. Richardson. 1969. Preprocessing Transformations and Their Effects on Multispectral Recognition. Proceedings, Sixth International Symposium on Remote Sensing of Environment, Oct., pp. 97-131. Martin, L.R.G. 1989. Accuracy Assessment ofLandsat-Based Visual Change Detection Methods Applied to the Rural-Urban Fringe. Photogrammetric Engineering and Remote Sensing, Vol. 55, No. 2, February, pp. 209-215.

23

Maxwell, E.L. 1976. Multivariate System Analysis of Multispectral Imagery. Photogrammetric Engineering and Remote Sensing, Vol. 42, pp. 1173-1186. Nagao, M., and T. Matsuyama. 1979. Edge Preserving Smoothing. Computer Graphics and Image Processing, Vol. 9, pp. 394-407. Place, J.L. 1974. Change in Land Use in the Phoenix (1260,000) Quadrangle, Arizona, between 1970 and 1973 - ERTS as an Aid to a Nationwide Program for Mapping General Land Use. Proceedings, Earth Resources Technology Satellite - Symposium, Washington, DC, Vol. 1, pp. 393-423. Purcell, E.J. 1978. Calculus with Analytic Geometry. Prentice-Hall, Englewood Cliffs, NJ, pp. 442-447. Robinson, J.W. 1979. A Critical Review of the Change Detection and Urban Classification Literature. Technical Memorandum 79/6235, Computer Sciences Corp., Silver Springs, MD, 50 p. Royal, J.A. 1980. Change Detection Method Development: Census Urban Area Application Pilot Test. Final Report, contract NAS5-25707, General Electric Company, Beltsville, MD, May, 74 p. Swain, P.H., and S.M. Davis. 1978. Remote Sensing: The Quantitative Approach. McGraw Hill, New York, 396 p. Swain, P.H., and H. Hauska. 1977. The Decision Tree Classifier: Design and Potential. IEEE Transactions on Geoscience Electronics, GE-15, pp. 142-147. Todd, W.J. 1977. Urban and Regional Land Use Change Detected by Using Landsat Data. Journal of Research, U.S. Geological Survey, Vol. 5, pp. 529-534. Todd, W.J., R.N. Hall, C.C. Henry, and B.L. Lake. 1978. Metropolitan Land Cover Inventory Using Multiseasonal Landsat Data. USGS Open File Report 78-378, Sioux Falls, SD, EROS Data Center, 26 p. Toll, D.L., J.A. Royal, and J.B. Davis. 1980. Urban Area Update Procedures Using Landsat Data. Proceedings, American Society of Photogrammetry, 12 p. Walthall, C.L., and E.M. Knapp. 1978. The Classification of Landsat Data for the Orlando, Florida, Urban Fringe Area. Report on contract NAS 5-24350, Computer Sciences Corp. Silver Spring, MD. Weismiller, R A , S.S. Kristof, D.K. Scholz, P.E. Anuta, and S.A. Momin. 1977. Change Detection in Coastal Zone Environments. Photogrammetric Engineering and Remote Sensing, Vol. 43, pp. 1533-1539.

24