Impervious surface acreage is mapped in percent ranges by individual grid cell. Users may generate area estimates by factoring the cell size (8742.25 sq. ft. or .2 acres) by the low, midpoint, or high end of the assigned range.
Two related data sets, Impervious Surfaces in Coastal New Hampshire - 1990 and Impervious Surfaces in Coastal New Hampshire - 2005, are also available. Derived from similar data and using similar techniques, they provide prior and subsequent estimates of impervious surface coverage.
Development of the 2000 Impervious Surface data was funded in part by a grant from the Office of State Planning, New Hampshire Estuaries Project, as authorized by the U.S. Environmental Protection Agency pursuant to Section 320 of the Clean Water Act.
Please cite as "New Hampshire GRANIT. 2001. Impervious Surfaces in Coastal New Hampshire - 2000. University of New Hampshire, Durham, NH."
DESCRIPTION PRODUCER'S ACCURACY USER'S ACCURACY Not impervious 87.8% 91.5% Impervious 95.8% 93.9%
The above error matrix reports the approximate accuracy of the results. It presents classified data results (e.g. derived from image processing) relative to reference data (e.g. data acquired via field visits or from some other source of known reliability). However, it is important to note that this standard methodology does not fully characterize the reliability of the results because the impervious surface pixels were mapped on a percentage basis. The accuracy assessment only evaluates the presence/absence of imperviousness at a given site, not the specific percentage impervious.
Further, two constraints were applied during selection of the assessment sites. First, a road proximity constraint was applied (within 5 pixels or approximately 467 feet of a road centerline) to facilitate the completion of the assessment. Second, each impervious surface feature was "shrunk" by 1 pixel width prior to the selection process to exclude confusion among edge pixels.
By constraining the accuracy assessment selection technique, the site selections were probably biased in favor of those areas that are most easily mapped (e.g. large parking lots, buildings, and residential subdivisions rather than single houses and isolated features). Nevertheless, the assessment provides a general estimate of the data reliability.
A geometric model was generated from the source imagery, which was then used to reference the data to New Hampshire Stateplane coordinates, NAD83. The model was derived using approximately 25-30 ground control points selected from a Landsat 5 TM reference image.
Ancillary data comprised numerous holdings from the GRANIT archive (the NH statewide GIS), including watershed boundaries, panchromatic Digital Orthophotoquads (DOQs), Digital Raster Graphics (DRGs), NH Department of Transportation road centerlines, Digital Elevation Models (DEMs), SPOT panchromatic (10 meter resolution) images, and US Fish and Wildlife Service National Wetlands Inventory (NWI) maps.
A body of 75 training sites, representing various types of impervious surfaces, was utilized in the traditional classification. These data were available as a result of numerous land cover classifications conducted within the project area over the past several years. Coupled with local knowledge, the training data were used to perform maximum likelihood classifications on the satellite imagery, yielding a data set of developed/undeveloped features for each year. The developed/urban class included areas characterized by a high percentage (typically 50% or greater) of constructed materials (asphalt, concrete, buildings, etc.). The identification of specific areas as urban was based strictly on features visible in the imagery, and thus only the areas within large subdivisions that were actually constructed were classified as urban.
Some obvious misclassifications were identified in the preliminary results. Tidal flats and wetlands, shallow water and scrub-shrub wetlands most often contributed to the problematic situations. These "problem pixels" were addressed using either an iterative process, whereby training data were added/deleted and the classification re-run, or by using on-screen editing to delete misclassified pixels in the final data set. After satisfactory results were obtained, the data were available for subsequent use.
The ERDAS Imagine Subpixel analysis tool was then applied to derive additional estimates of "proportion of imperviousness" for each urban cell in the study area. This methodology (more fully described at www.discover-aai.com and www.erdas.com) is capable of detecting materials of interest (MOI) - in this case, impervious surfaces - that occur within each pixel. The classification describes each pixel as having a percentage of the MOI ranging from 20 to 100, reported in increments of 10%. Additional processing using road centerline data, described further below, resulted in the inclusion of the lower, 0-19% range. Note that the spatial extent of the impervious surface (the MOI) within each pixel is not identified. Rather, the entire pixel is reported as having a certain percentage of the MOI. By factoring the area of each pixel by the percent of that pixel containing the MOI, acreage summaries may be generated.
The subpixel processing approach followed generally accepted techniques (Flanagan, 2000; Flanagan and Civco, 2001; ERDAS, 2000). The 2000 TM data set was initially used to generate 15-20 potential signatures, which were evaluated by running an MOI classification and displaying the results on the underlying imagery. The results were evaluated both by visual inspection of 1998 USGS Digital Orthophotoquads (DOQs), and by reference to personal knowledge of the area. However, it is important to recognize that the evaluation of each classification compared the presence/absence of impervious surface MOI and not the actual percentage mapped per image pixel, as we had no data to effect the latter type of comparison.
Signatures were marked as "good", having "potential", or "unusable". Good signatures were those that provided tight classifications and would require little if any on-screen editing. Signatures having "potential" were those that mapped much of an area correctly, but would need some data clean up. Potential signatures were also those that could be altered using classification tolerances, (a standard feature of the subpixel classification routine), such that more or fewer image pixels would be included in the classification set. Signatures were considered "unusable" when too many pixels were included in the classification and an unreasonable amount of on-screen editing would be required to produce an acceptable data set. As a result of these signature derivations and classification tests, 12 signatures were accepted to generate the final impervious surface data set. These signatures provided a reasonable classification that could be edited to derive a provisional impervious surface data set.
Unlike traditional supervised classifications, the subpixel approach typically produces classifications based on a single signature. Accordingly, 12 data sets were produced and subsequently merged into one. This was achieved by "layer stacking" the images and then using Imagine statistical functions to select the maximum layer value (e.g. maximum percentage of imperviousness) at each pixel.
The post processing phase of the project was designed to enhance the classification phase by addressing two specific issues - the correction of any remaining, obvious errors in the classification results, and the incorporation (or "burning in") of road centerline data to optimize the mapping of pavement as an impervious surface feature. Two ancillary data sets were obtained for this phase:
- US Fish & Wildlife Service National Wetlands Inventory (NWI) data, based on aerial photography acquired in the mid-1980's, as archived in the GRANIT database; and - New Hampshire Department of Transportation (NHDOT) road centerline data - both public and private roads, as of August, 2002
The provisional impervious surface classification included some recurring errors - typically misclassified pixels occurring in open water, wetland and forests. The image analyst could often quickly identify these errors using pattern recognition, past experience and in some cases, DOQ reference images. Errors were removed from the classification by defining polygons around the misclassifications and recoding, as appropriate. Because many of the misclassified pixels occurred in wetlands, NWI data were converted to a grid format and used as a mask to rapidly isolate and review potential problem areas. However, pixels concurrent with the NWI grid were not simply converted to non-impervious status, because of numerous cases where wetlands had been filled since the NWI photo date and were properly coded as impervious.
Finally, the methodology included the incorporation of NHDOT public and private road data in the final product, where the imperviousness of each pixel was assigned based on the road pavement width. (Because of their relatively narrow, linear shape, road features are occasionally omitted in the classification phase.) However, the pavement characteristic was only available for the public road data set. Thus, an editing task was required to identify the surface type (paved/unpaved) of private roads, and the paved private roads were assigned a default pavement width of 20 ft. The pavement width characteristic was then used to "burn" the paved roads (public and private) into the classified data set.