DBPapers
DOI: 10.5593/SGEM2014/B23/S10.043

TWO-STAGE SUBPIXEL IMPERVIOUS SURFACE COVERAGE ESTIMATION: COMPARING C 5.0/CUBIST AND RANDOM FOREST

K. Bernat, W. Drzewiecki, M. Twardowski
Wednesday 1 October 2014 by Libadmin2014

References: 14th International Multidisciplinary Scientific GeoConference SGEM 2014, www.sgem.org, SGEM2014 Conference Proceedings, ISBN 978-619-7105-12-4 / ISSN 1314-2704, June 19-25, 2014, Book 2, Vol. 3, 343-350 pp

ABSTRACT
The paper presents accuracy comparison of subpixel classification based on medium resolution satellite images (Landsat 5TM), performed using two machine learning algorithms (C5.0/Cubist and Random Forest) built on decision and regression trees method. The research was conducted for the immediate catchment of the Dobczyce Reservoir (which is the main source of water supply for the city of Cracow, Poland) along with an adjacent area of towns (Myślenice and Dobczyce). The land use in the catchment is dominated by agriculture with numerous villages of dispersed development. The southern part of the study area is covered mainly by forests. The aim of the study was to obtain image of percentage impervious surface coverage, valid for the period of 2009-2010. Imperviousness index map generation was a two-stage procedure based on two algorithms. The first step was hard classification, performed using decision trees method, which divided the study area into two categories: i) completely permeable (imperviousness index less than 1%) and ii) fully or partially impervious areas. In a second stage, for pixels classified as impervious, the percentage of impervious surface coverage in pixel area was estimated using regression trees approach. Accuracy of the final imperviousness index map was checked based on validation data set, which was not used for learning and testing of classifiers. The root mean square error (RMS) of determination of the percentage of the impervious surfaces within a single pixel was ±12.8% for C5.0/Cubist method and ±13.9% in case of Random Forest method. Further results analysis shown, that in intensively urbanized areas small imperviousness index value differences occurred. Larger differences of up to a few percent were found in agriculture and forests areas, with more accurate results obtained using C5.0/Cubist method.

Keywords: subpixel classification, impervious surfaces, decision and regression trees, Landsat TM