Countering New Challenges Regarding Classification Quality Assessment Methods with the Help of Fuzzy Boundaries

Christoph Kinkeldey and Jochen Schiewe
Lab for Geoinformatics and Geovisualization (g2lab) HafenCity University (HCU) Hamburg, Germany
{christoph.kinkeldey, jochen.schiewe}

Abstract: The quality assessment of classified remote sensing data has become more challenging since the geometrical and spectral resolution of remote sensing data has increased. The project CLAIM (Classification Assessment using an Integrated Method) deals with the development of an a posteriori quality assessment method countering these challenges. We propose the consideration of uncertainties in the classification data as well as in the ground truth data (integrated method). By substituting the discrete object boundaries by symmetrical buffer areas we introduce transition zones as a model for the geometric and semantic uncertainties. A key aspect of the application is the selection of two parameters, i.e., the transition zone width and the class membership function. In this contribution we focus on the characteristics of the model parameters and how to choose appropriate values.

Keywords: classification evaluation; accuracy assessment; uncertainty; remote sensing; fuzzy logic; transition zones

KinkeldeyAccuracy2010.pdf619.32 KB