The agreement between Landsat estimates and measurements of impervious made from high resolution aerial images was R-Square = 0.9200

Data provides complete coverage over the stated extent of the data.

RMS Error is 0.25 - pixel or <7.5 meters

RMS Error is 0.25 - pixel or <7.5 meters

The process is generally described as follows. The imagery was rectified to the UTM coordinate and projection system. A combination of spring and summer dates of Landsat data for each year (1986, 1991, 1998 or 2002) was used to classify the land cover types into urban/developed vs. rural (agricultural, forest, and wetland) and water classes. Measurements of impervious surface area were made on high resolution aerial imagery for a random sample of areas for calibration of the Landsat classification model. The calibration sites were used to develop a least squares model relating percent impervious to the spectral-radiometric response of the Landsat TM data for the summer image of each year. The particular Landsat parameter used to estimate percent impervious was "tasseled cap" greenness which is sensitive to the amount of green vegetation and therefore is inversely related to the amount of impervious surface area. The regression model was used to estimate the amount of impervious surface within the area of each pixel. After applying the regression model to the Landsat TM image the land cover classification was used to assign all non-urban areas a zero percent impervious value. Accuracy assessment of the Landsat derived impervious surface estimates was performed with a second independent sample of measurements of impervious from the aerial imagery. Accuracy assessment was based on the linear fit of actual versus predicted impervioius surface.

Image processing:

 

Impervious surface classification was performed with a Landsat TM satellite image (path: 27 row:28/29) acquired on 04 September, 1991. The imagery was rectified to the UTM coordinate and projection system to an RMS error of 0.25 pixel (<7.5-meters) using 25 ground control points. The base layer used was the MNDOT Major Roads map. Nearest neighbor resampling to a 30-meter pixel size and the coordinates of the final image were adjusted to values evenly divisible by 30. Following rectification the imagery was transformed to tasseled cap values using ERDAS Imagine default processes. The tasseled cap data were stretched to unsigned 8-bit.

 

Impervious Model Calibration Data:

 

Model calibration sites were randomly selected and high resolution aerial imagery were used to determine the amount of impervious surface within each calibration site. 53 sites, distributed across the image area, were collected for each image. Imperious surface area was determined from the 1991 DOQQ's of 1-meter resolution. Percent impervious surface area was calculated for each site.

Modeling Impervious Surface Area:

The measurements of impervious surface area from the calibration sites were used to develop a least squares regression model relating percent impervious to the spectral-radiometric response of the Landsat TM data. The Landsat parameter used to estimate percent impervious was tasseled cap greenness which is sensitive to the amount of green vegetation and therefore is inversely related to the amount of impervious. The regression model used for the 04 September, 1991 Landsat TM image was:

Percent Impervious Surface = 48.35453 - 1.8741616*(tasseled cap greenness) + 0.0043972*(tasseled cap greenness-2.8306)^2

R-Square: 0.85999999999999999, Adjusted R-Square: 0.84999999999999998, Root Mean Square Error: 11.33, Observations: 53

Accuracy Assessment/ Inverse Calibration:

 To measure the accuracy of the Landsat-derived impervious surface estimates, an independent sample of  104 accuracy assessment sites were selected. The sites were randomly selected and compared with high resolution aerial imagery to determine the amount of impervious surface within each accuracy assessment sites. Accuracy was determined with a plot of actual vs. predicted plots and results were as follows:

Linear Fit: Predicted Percent Impervious Surface Area = 3.4995858 + 0.9009355*(Actual Percent Impervious Surface Area)

R-Square: 0.92000000000000004, Adjusted R-Square: 0.92000000000000004, Root Mean Square Error: 8.5999999999999996, Observations: 104

To remove model bias an inverse calibration was computed from the linear fit of the actual vs. predicted plot and applied to the impervious surface classification. The inverse equation (Actual Percent Impervious Surface Area = (3.4995858 - Predicted Percent Impervious Surface Area)/0.9009355).  Accuracy of Landsat derived impervious surface estimates was reassessed following the inverse calibration. The results follow: Linear Fit:

Predicted Percent Impervious Surface Area = -1.070047 + 0.9755037*(Actual Percent Impervious Surface Area)

R-Square: 0.92000000000000004, Adjusted R-Square: 0.92000000000000004, Root Mean Square Error: 8.5999999999999996, Observations: 104

Although, the inverse calibration did not affect the R-Square and the RMSE increased, the bias is removed. This is apparent from the reduction in the intercept and an increase in the slope.

Removing Non-Urban Areas:

This step identified non-urban areas within the Landsat image area and reclassified these areas in the impervious surface classification with a percent impervious surface value of zero. To identify non-urban areas a combination of spring and summer dates of Landsat imagery, taken in 1991, were used to classify the land cover types into urban/developed vs. rural (agricultural, forest, and wetland) and water classes. Classification accuracy, for the TCMA, MN region, had producers and users accuracies of 96% and 98.2%, respectively, for the urban class and an overall classification accuracy 94.599999999999994% with a Kappa statistic of 0.93. As an additional step urban areas from previous years were added to the current year (i.e. the 1998 classified urban area would be added to the 1991 and 1986 classified urban areas). The final impervious surface classification was then overlaid with the land cover classification and areas identified as non-urban were reclassified with a percent impervious surface value of zero.

Limiting to MCD Boundaries:

The final impervious surface classification was then clipped to match the boundaries of the MCD of interest. Geospatial data used to identify MCD boundaries was acquired from the Minnesota Legislative GIS office, <ftp://ftp.commissions.leg.state.mn.us/pub/gis/shape/mcd2003.zip>.

The pixel values represent the area (in percent) of completely impervious surface within the 30-m by 30-m pixel area. Values range from 0 to 100 percent.

This data may be used for educational and non-commercial purposes, provided proper attribution is given. Secondary distribution of the data is permitted, but not supported by the University of Minnesota. By accepting the data, the user agrees not to transmit this data or provide access to it or any part of it to another party unless the user includes with the data a copy of this disclaimer.

see website or contact info