Estimating percent impervious cover from Landsat-based land cover with a simple and transferable regression model

Percent impervious cover (PIC) is a widely used metric in ecological and hydrological analyses because it is highly correlated with pollutant and storm water run-off. The moderate-resolution satellite data (e.g. Landsat), that are typically used to calculate PIC, tend to overestimate PIC for all but very rural and very urban landscapes. Existing models for calibrating PIC estimates (e.g. ISAT, ETIS) are limited in that they are applicable only for specific land cover datasets and may also require population data; furthermore, these models have not been tested for performance outside of the geographic locations in which they were developed. The goal of this study was to explore simple but widely applicable regression models as tools for calibrating PIC estimates based on moderate resolution satellite data. The regression models used impervious land cover, from Landsat-based datasets, as the sole predictor of actual PIC. PIC was calculated for analysis units, ranging in size from 2.25 ha to ≥100 ha, for locations in Connecticut, Massachusetts, and Ohio in the United States. Regression models were fit for each size class of analysis unit at each study location; generalized versions of the models were created by fitting a regression to all size classes of analysis units at a given study location. Calibrated PIC estimates had root mean square error (RMSE) values that ranged from 1.5–10.7%; these values were considerably better than RMSE values for uncalibrated PIC estimates which ranged from 10.1–23.3%. For both calibrated and uncalibrated PIC, the accuracy of the estimates improved with the increasing size of the analysis units. Model regression coefficients were similar regardless of the analysis unit size, geographic location, or land cover dataset; model performance declined only slightly when applied outside the area in which it was developed. The simple regression models developed in this study had similar performance to previous calibration models (i.e. ISAT, ETIS) but are easier to apply and more widely applicable.