SUBPIXEL IMPERVIOUS SURFACE MAPPING

Identified by the EPA as the leading threat to surface water quality in the United States, nonpoint source (NPS) pollution is channeled into rivers and streams via impervious surfaces. Impervious surfaces are anthropogenic features, such as roads, buildings, and parking lots, through which water cannot infiltrate into the soil. Research from the past 15 years shows a consistent, inverse relationship between the percentage of impervious surfaces in a watershed and the health of its receiving stream. In conjuction with remote sensing satellite imagery, this relationship may be utilized as a time- and cost-effective indicator of overall ecosystem health and water quality. Impervious estimates are typically calculated by multiplying a land use specific percent impervious coefficient by the total area of that land use within a drainage basin. Though a widely used method, this approach does little to promote accurate, standardized, measures upon which to base land use planning decisions. Artificial neural networks and the ERDAS Imagine SubPixel Classifier were investigated as methods for the improved characterization and quantification of impervious surface cover. The principal goal of this research was to develop an accurate, standardized, and geographically extensible impervious surface prediction model. This model was based upon Landsat Thematic Mapper data and was used to quantify, by land cover type, the percent imperviousness at the subpixel (30 m) level. High accuracy planimetric data, in the form of an impervious footprint, were used to calibrate both models for four municipal study areas in Connecticut. Considering only impervious-pervious detection at the pixel level, overall accuracies for the artificial neural network and the ERDAS Imagine Subpixel Classifier predictions, respectively, were 92% and 94% for Marlborough, 90% and 92% for Waterford, 84% and 86% for Woodbridge, and 74% and 71% for West Hartford. At the local watershed level, the RMSE for the four towns for the neural network approach and Subpixel Classifier, were, respectively: Marlborough, 1.29 and 0.66; Woodbridge, 2.51 and 0.99; West Hartford, 4.97 and 5.97; and Waterford, 1.24 and 2.98. Results from this research will provide the foundation for subsequent efforts to quantify impervious surface cover using satellite remote sensing imagery.