Detection of impervious soil area in multispectral remote sensing - a comparison
Julius-Maximilians-Universität Würzburg, Bachelorarbeit, (August 2018)Abstract
Urban areas are bound to huge changes to accommodate infrastructure, living- and workspaces. The monitoring of these changes plays a vital role in planning and managing sustainable land use and land cover. Impervious surface, although not attributable to the whole area of urban sprawl, influences local climate known as urban heat island (UHI) effect, distribution of watershed, water quality and many other factors which ultimately affect the inhabitants of an area. To approximate consequences, exact measurements of Impervious Surface Area (ISA) are needed, which have to be cost-effective and up to date. Current operational methods to detect ISA offer only low temporal resolutions, are mostly based upon costly airborne surveying techniques or are not freely available.
In this study, an approach based on remote sensing satellite data for the mapping of impervious surface is evaluated on Munich, Bavaria for the year 2011 and 2017. With the use of very high spatially resolved WorldView-2 and WorldView-4 sensors as reference data, the Percentage of Impervious Surface (PIS) is estimated with a Support Vector Regression (SVR) for medium spatially resolved and freely available sensors Landsat 5, Landsat 8 and Sentinel 2. PIS calculation is done per-pixel on a spatial resolution of 30, 15 and 10 meters. Comparison is made for different inputs, which are the reflectance data of the sensor, Principal Component Analysis (PCA), spectral indices and combination of PCA and the Normalized Difference Vegetation Index (NDVI). Subsequently the needed amount of samples for an accurate result is analyzed, as well as the noise which results out of a randomized input sample selection. The results are compared to Copernicus' imperviousness status maps and a municipal imperviousness study from 2011 by the city of Munich.
Results show that PIS and extent of ISA can be predicted at 11.12\% Mean Absolute Error (MAE) (16.15\% Root Mean Square Error (RMSE)) for the city of Munich, with a slight loss in accuracy for higher spatial resolutions. Differences in accuracy related to spatial resolution are within 3\% range in MAE and 6\% RMSE for the same coherent Area Of Interest (AOI) as reference sample. A combination of PCA and the NDVI showed to deliver the most consistent results in the comparison. A sample size of 200 to 400 data points was preferred, given the amount of calculation time and reduced noise with more samples. The final comparison with the municipal imperviousness study for 2011 and Copernicus' imperviousness maps for 2012 and 2015 showed MAE and RMSE values ranging from 14.23 to 19.92\% and 20.64 to 29.33\% for all sensors.
