Subpixel Estimates of Impervious Cover from Landsat TM Image
Impervious Surfaces and Their Relevance
Impervious surfaces are generally understood to be any material, natural or man-made, that prevents the infiltration of surface water. These surfaces include concrete, asphalt, brick, roofing, and other man-made materials as well as compacted soils and exposed rock outcroppings. Impervious areas dramatically alter surface hydrological properties, affecting the dynamic flow of water and the rate of subsurface reservoir recharge. Impervious surfaces also impact the physical and chemical properties of surface reservoirs resulting in reduced water quality and poor watershed health. The characterization and quantification of impervious surfaces has been suggested as a way to manage environmental restoration efforts in watersheds (Schueler, 1994) and as a way to map urban environments (Ridd, 1995). Obtaining improved estimates of impervious surface area at local and regional scales has been identified as an essential component of planning decisions.
Previous Efforts to Map Impervious Surfaces
Initial efforts to map impervious surfaces from remote sensing platforms included the use of low-altitude aerial photography and manual digitization techniques. With the launch of the Landsat Multispectral Sensor (MSS) in 1972 and the Landsat Thematic Mapper (TM) in 1984, digital satellite imagery began providing a synoptic view of the Earth's surface which allowed for the production of regular, repeatable land cover maps. Automated spectral analysis of MSS and TM data reduced the amount of labor necessary for impervious surface delineation and lowered operating costs (Ragan and Jackson, 1975).
Many of the methods using spectral information from satellite sensors are based on supervised and unsupervised classification techniques and other forms of spectral clustering, thresholding, and modeling. Other estimates of impervious cover rely on lookup tables derived from surrogate measures such as parcel size (Monday et al., 1994; Sleavin et al., 2000) and land use/land cover information (Deguchi and Sugio, 1994; Williams and Norton, 2000; Ward et al., 2000). These techniques are successful at quantifying relative impervious surface cover across an entire watershed but they lack impervious surface information at the subpixel level for specific locations.
The Need for Subpixel Information
A mismatch in sensor resolution compared to the fine spatial resolution of features in the urban environment may result in errors of impervious estimation. New techniques in the field of digital image processing permit the derivation of information at the sub-pixel level. Classification techniques such as spectral mixture modeling (Ji and Jenson, 1999; Ward et al.; 2000 and Phinn et al., 2000) and neural network based classification methods (Civco and Hurd, 1997) have been capable of extracting sub-pixel information. The study presented here describes another technique capable of extracting subpixel information on impervious surface cover using Landsat TM imagery and a decision tree classifier.
Planimetric Data Processing Steps
Impervious features were first identified and then extracted from a GIS planimetric data set based on polygon attribute codes provided by Montgomery County Maryland and the Maryland National Capital Park and Planning Commission (MNCPPC). This impervious vector coverage was then processed into a 3m-raster image, maintaining the spatial fidelity of the data. A 10x10 summary filter was then used to degrade the 3m impervious image; essentially counting the number of 3m pixels in a 100-pixel window, thereby estimating the percent of impervious cover corresponding to a 30m TM pixel. This process was applied to the entire county producing a GIS-based training data set at TM resolution.
Click image above for full resolution view
Figure 1. Planimetric Data Processing Steps: Impervious features are extracted from the GIS database (a) and rasterized to 3m (b). The 3m raster is then degraded to 30m using a 10x10 pixel summary filter, essentially counting the number of 3m pixels in a 100-pixel block. This produces a 30m estimate of impervious cover comparable to Landsat TM data.
Decision Tree Classifier
Classification trees provide a robust statistical method useful for mapping land cover types at regional to global scales (Hansen et al., 1996), assessing topographic and hydrological interactions (Townsend, in review) and classifying soil drainage classes (Cialella et al., 1997).
Click image above for full resolution view
Figure 2. Through a recursive process of binary partitioning, Classification Trees are capable of teasing out hierarchical data structures and non-linear interactions between predictor variables.
Decision tree classifiers are capable of handling both parametric and non-parametric predictor variables. Inputs can be in the form of continuous numeric values or thematic classes and missing data values are acceptable (Breiman et al., 1984; Quinlan, 1993). Through the recursive binary partitioning of predictor variables into smaller more homogeneous groups, decision tree classifiers are capable of "teasing out" hierarchical data structures and non-linear interactions between predictor variables (Quinlan, 1993). This process of binary partitioning results in traceable splits of the predictor variables that contain meaningful information on the biophysical and reflectance properties of the predicted classes.
Using the planimetric GIS data produced by the MNCPPC and multi-temporal
Landsat 5 TM data to train c5.0, the decision tree software package
developed by RuleQuest Research (Quinalan, 1993), we quantified subpixel
impervious surface cover for Montgomery County. We also extended the predictive capability of the algorithm across a larger area encompassing the Baltimore-Washington Metropolitan Area.
Click image above for full resolution view
Figure 3. GIS Derived Estimates of Impervious Cover.
Click image above for full resolution view
Figure 4. Decision Tree Estimates of Impervious Cover
Accuracy Assessment
The classification tree had an overall within-class accuracy of 83.82% (Kappa = 0.707) when compared to the more complete areas of the GIS planimetric data. We note that maintaining a GIS planimetric database the size of Montgomery County is an enormous task and several impervious features were omitted or incorrectly coded in the database. This uncertainty in the planimetric data is obvious in certain areas. Locations used to extract training data were selected and visually verified for completeness using color-IR Digital Ortho Quarter Quads (DOQQ) from a similar time period as the planimetric map. At first glance the decision tree classification seemed to overestimate impervious features but, in fact, the planimetric map underestimated impervious cover as a result of missing features (e.g. some parking lots). The decision tree, however, missed small impervious features at scales below the sensor resolution (e.g. single lane rural roads).
A substantial amount of misclassification occurred in adjacent impervious classes. We used a non-traditional measure that accounts for a slight misclassification across adjacent classes to assess the classification accuracy. This across adjacent-class overall accuracy of 87.68% (Kappa = 0.777) showed more promising results, as expected, and paints a better picture of how well the algorithm is working.
| Producer's Accuracy |
| Impervious Class | Across Class Accuracy |
| 0 | 98.36% |
| 1-10 | 86.90% |
| 11-20 | 61.45% |
| 21-30 | 76.35% |
| 31-40 | 78.80% |
| 41-50 | 69.17% |
| 51-60 | 61.46% |
| 61-70 | 53.99% |
| 71-80 | 56.10% |
| 81-90 | 92.46% |
| 91-100 | 94.86% |
|
| User's Accuracy |
| Impervious Class | Across Class Accuracy |
| 0 | 92.35% |
| 1-10 | 88.44% |
| 11-20 | 84.81% |
| 21-30 | 82.42% |
| 31-40 | 80.82% |
| 41-50 | 77.30% |
| 51-60 | 75.61% |
| 61-70 | 78.21% |
| 71-80 | 88.77% |
| 81-90 | 96.38% |
| 91-100 | 94.03% |
|
The quantification of impervious surfaces using decision tree classifiers and Landsat TM imagery performed well when compared to the planimetric map estimates. Future research will be focused on the application of this technique across a diverse range of conditions and development patterns through time.
Reference
Breiman, L., Friedman, J., Olshen, R. and Stone, C., 1984. Classification and Regression Trees. Chapman and Hall, New York, 358 pp.
Cialella, A.T., Dubayah, R., Lawrence, W. and Levine, E., 1997. Predicting Soil Drainage Class Using Remotely Sensed Data and Digital Elevation Data. Photogrammetric Engineering & Remote Sensing, 62(2): 171-178.
Civco, D.L. and Hurd, J.D., 1997. Impervious Surface Mapping for the State of Connecticut. Proceedings of the 1997 ASPRS Annual Conference, Seattle, WA, 3: 124-135.
Deguchi, C. and Sugio, S., 1994. Estimations for Percent Impervious Area by the Use of Satellite Remote Sensing Imagery. Water Science and Technology, 29(1-2): 135-144.
Hansen, M., Dubayah, R. and DeFries, R., 1996. Classification trees: an alternative to traditional land cover classifiers. International Journal of Remote Sensing, 17(5): 1075-1081.
Ji, M. and Jensen, J.R., 1999. Effectiveness of Subpixel Analysis in Detecting and Quantifying Urban Imperviousness from Landsat Thematic Mapper Imagery. Geocarto International, 14(4): 33-41.
Monday, H.M., Urban, J.S., Mulawa, D. and Benkelman, C.A., 1994. City of Irvine Utilizes High Resolution Multispectral Imagery for N.P.D.E.S. Compliance. Photogrammetric Engineering & Remote Sensing, 60(4): 411-416.
Phinn, S.R., Stanford, M., Shyy, P.T. and Murray, A., 2000. A Sub-Pixel Scale Approach for Monitoring the Composition and Condition of Urban Environments Based on the (VIS) Vegetation-Impervious-Surface Model, 10th Australasian Remote Sensing and Photogrammetry Conference (ARSPC), Adelaide, Australia.
Quinlan, J.R., 1993. C4.5 : programs for machine learning. Morgan Kaufmann series in machine learning. Morgan Kaufmann Publishers, San Mateo, CA, 302 pp.
Ragan, R.M. and Jackson, T.J., 1975. Use of Satellite Data in Urban Hydrologic Models. Journal of the Hydraulics Division, 101(HY12): 1469-1475.
Ridd, M.K., 1995. Exploring a V-I-S (vegetation-impervious surface-soil) model for urban ecosystem analysis through remote sensing: comparative anatomy for cities. International Journal of Remote Sensing, 16(12): 2165-2185.
Schueler, T.R., 1994. The Importance of Imperviousness. Watershed Protection Techniques, 1(3): 100-111.
Sleavin, W.J., Civco, D.L., Prisole, S. and Giannotti, L., 2000. Measuring Impervious Surfaces for Non-Point Source Pollution Modeling. Proceedings of the ASPRS Annual Conference May 22-26, 2000, Washington D.C.
Townsend, P.A., in review. Mapping seasonal flooding in forested wetlands using multi-temporal Radarsat SAR. submitted to: Photogrammetric Engineering & Remote Sensing.
Ward, D., Phinn, S.R. and Murry, A.T., 2000. Monitoring Growth in Rapidly Urbanizing Areas Using Remotely Sensed Data. Professional Geographer, 52(3): 371-386.
Williams, D.J. and Norton, S.B., 2000. Determining Impervious Surfaces in Satellite Imagery using Digital Orthophotography. Proceedings of the ASPRS Annual Conference May 22-26, 2000, Washington D.C.
|
