College of Agriculture and Natural Resources
Center for Land Use Education and Research
Impervious Surfaces

Long Island Sound Regional Impervious Surface Study

Methods

Long Island Sound Study

Sub-pixel Classification | FGDC Metadata

Sub-pixel Classification

The impervious surface estimates were derived through a process known generically as spectral un-mixing. In the case of this work, Leica Geosystem’s ERDAS IMAGINE Subpixel Classifier (SPC), an add on module developed by Applied Analysis,Inc., was used to derive the IS estimates. Essentially how the SPC works is that a user identifies several pixels (spectral signatures) that represent a particular material of interest (MOI), such as impervious surfaces. The spectral signatures are processed through the SPC and all dissimilar materials are removed from the pixel resulting in a certain proportion of the pixel containing the MOI. The output is an image indicating pixels containing the MOI in eight classes, at 10 percent increments, beginning with 20 percent (the SPC is unable to detect MOIs that contribute less then 20 percent of the pixel).

The SPC process was applied to each of four dates of Landsat imagery (1985, 1990, 1995, and 2002). Several steps are necessary before the estimation can proceed. Image preprocessing and generation of an environmental correction file are first performed as required by the SPC on each date of Landsat imagery. Training pixels that represent, at minimum, 90 percent imperviousness were selected for each date concurrently to assure the same training pixel for each date of imagery was used. This process was repeated for each of four brightness sub-classes (bright, medium, dark, and very dark) with 50–60 training pixels selected for each sub-class. For a successful sub-pixel classification, the quantity of training pixels is less critical than the quality of the pixels (i.e., pixels that represent nearly 100 percent imperviousness), but it is important to try to select pixels that represent the full spectral range of impervious surfaces within each sub-class. The signatures for each sub-class were combined using the SPC Signature Combiner then MOI Classification was performed. Eight output classes were generated representing different percentages of imperviousness within a (i.e., class 1 = 20-30%, class 2 = 30-40%, … class 8 = 90-100%). The resulting IS maps were clipped to the developed area of their respective land cover map. This helped to make the data consistent among the four dates and to maintain compatibility with the land cover data. Additionally, any pixel not classified as containing IS (possibly because it fell below the 20% threshold set by the SPC) but was classified as developed in the land cover was considered to be 10% impervious. Additional post-processing was performed to eliminate the chance of a given pixel to fluctuate up and down between dates. The final result is four IS maps with nine classes (i.e., class 1 = 10-20%, class 2 = 20-30%, … class 9 = 90-100%), that is compatible with land cover data.

Landsat image
Derived Land Cover
Derived IS Estimate


Read the Project Completion Report for more detailed explanations of the data and techniques (you will need Adobe Acrobat free reader for this).

FGDC Metadata

The FGDC (Federal Geographic Data Committee) has a standard form for reporting metadata, or information about data. Visit the FGDC-compliant metadata for:

Back to Top