An initial minimum distance to means (MDM; Jensen, 1996; Mesev, this volume) supervised classification was performed on the ASTER mosaic using 16 classes: Desert Soil, Low Vegetation; Desert Soil, Vegetated; Bedrock; Fluvial Sediments; Bare Soil; Fallow Agricultural Soil; Water; Canopied Vegetation; Grass; Riparian Vegetation; Active Agricultural Vegetation; Mesic Built Materials; Xeric Built Materials; White Rooftops; Blue Rooftops; and Asphalt. The term mesic refers to land cover types with significant vegetation in the form of grass, shrubs, and canopied woody plants. Xeric land cover types are typified by little to no grass or shrub cover and open-canopy plant types with significant bare rock and soil (i.e. similar to equatorial deserts). The MDM classification was run using each sub-area as a separate class, followed by aggregation of the results into the original sixteen classes.
The ASTER mosaic was also used to calculate a Normalized Difference Vegetation Index (NDVI; Botkin et al., 1984). This index highlights actively photosynthesizing vegetation by comparing reflectance values in the visible red (low for vegetation) and near infrared (high for vegetation) bands. It is computed as follows: (Band3-Band2) / (Band3+Band2), where Band 3 is near infrared and Band 2 is visible red reflectance. The index returns pixel values ranging from -1 (no vegetation; low reflectance in both bands 2 and 3) to 1 (pixel dominated by actively photosynthesizing vegetation).
Spatial variance texture was also calculated from the VNIR mosaic. This operation highlights large changes in brightness value (or reflectance) between adjacent pixels and has been shown to correlate well with urban versus non-urban land cover types. Spatial variance texture was calculated for all three VNIR bands using both a 3 x 3 and 5 x 5 pixel moving window. This was done to capture fine-scale spatial texture in urbanized regions as well as coarser-scale texture in undeveloped regions. The NDVI and variance texture raster data were then each separated into low, medium, and high data values using an unsupervised ISODATA algorithm. This approach takes advantage of the inherent statistical clustering within each NDVI and texture dataset, and provides a simple means of objective thresholding of the data.
Qualitative assessment of the MDM classification results indicated that significant misclassification was present both within and between the various soil, vegetation, and built classes. We then constructed an expert classification system similar to that used by Stefanov et al. (2001b, 2003) to perform post-classification recoding of the MDM classification result. An expert classification system applies a sequence of decision rules to a set of georeferenced datasets using Boolean logic (Vogelmann et al., 1998; Stuckens et al., 2000). This approach allows for the introduction of a priori knowledge into the classification data space and can significantly reduce errors of omission and commission. Figure 13.1 presents a schematic example where the dashed rectangle indicates the hypothesized pixel classification (“Soil and Bedrock”), hexagons are alternative decision pathways, and solid rectangles indicate the variables being tested. If any one of the decision pathways (“Path”) is satisfied by the variables, the pixel will receive the hypothesized classification value. There is no limitation on the number of variables or decision pathways that can be combined within an expert system framework. Most image processing software packages now include tools for constructing expert system classification frameworks.
The datasets combined in the expert system framework include the initial MDM land cover classification, unsupervised classifications of the NDVI and spatial variance texture data, and a land use vector polygon dataset. The land use data were acquired from the Maricopa Association of Governments (MAG; Maricopa Association of Governments, 2000) and are contemporaneous with both the ASTER and MODIS data. The land use data are constructed from a combination of survey questionnaires, site visit, and aerial photograph data. This dataset contains 46 separate land use categories which were aggregated to seven for use in the expert system model: Open Residential, Built, Cemeteries, Open Space, Golf Courses, Water, and Agriculture. Incorporation of land use polygon data provides additional discriminatory power for spectrally similar pixels such as asphalt and bedrock. For example, a pixel classified as Asphalt located within an Open Space polygon would be reclassified as Soil and Bedrock. A series of decision rules were then constructed to recode misclassified pixels in the MDM classification product. The MDM classes White Rooftops and Blue Rooftops were also recoded into one class, Reflective Built Surfaces, within the expert system model. The expert classification model was run using the area of overlap of the MDM classification and the MAG land use dataset only