Uncertainty in Image Interpretation as Reference for Accuracy Assessment in Object-based Image Analysis
Centre for Geoinformatics, Salzburg University, Salburg, Austria
Abstract: For accuracy assessment in object-based image analysis (OBIA) image interpretations can be used as reference data. This approach has some drawbacks as image interpretation contains uncertainty that may affect the reliability of the accuracy assessment of an OBIA product. Therefore, thematic and spatial uncertainties in image interpretation need to be quantified so that it can be accounted for by an object-based accuracy assessment. In this study the spatial uncertainty in object delineation was emphasised, because boundaries are more relevant for OBIA products than they had been for traditional pixel-based image analysis approaches. Twenty-two image interpretations acquired through an assignment within a postgraduate distance-learning programme were the basis to analyse uncertainty. The interpretations were grouped according to their delineation detail and overlaid to identify uncertain areas. The group of detailed interpretations agreed on the same class for 83.4% of the area, with 16.6% of the area remaining uncertain. Based on the overlay, median interpretations serving as a reference were retrieved for each group. Object Fate Analysis (OFA) was applied to estimate boundary deviation. The area that was affected by uncertainty was measured for several levels of agreement between the interpreters. E.g., for the detailed group the boundaries of 75% of the objects deviated less than 4.91 meters from their corresponding reference object. The deviation values of boundaries can be interpreted as distances for an error band along the delineations.
Keywords: classification accuracy; remote sensing; OBIA; object; fate analysis