SITE-SPECIFIC AREA-BASED VALIDATION OF CLASSIFIED OBJECTS T. G. Whiteside a, *, S. W. Maier b , G. S. Boggs c a Environmental Research Institute of the Supervising Scientist, Pederson Rd, Eaton, Darwin, NT, Australia – tim.whiteside@environment.gov.au b Research Institute for the Environment and Livelihoods, Charles Darwin University, Ellengowan Drive, Casuarina, NT, Australia – stefan.maier@cdu.edu.au c Wheatbelt Natural Resources Management Inc., York Rd, Northam, WA, Australia – gboggs@wheatbeltnrm.org.au KEY WORDS: Thematic accuracy, geometric accuracy, quality, multispectral classification, object-based. ABSTRACT: The establishment of geographic object based image analysis (GEOBIA) as a group of methodologies for analysing and classifying remotely sensed data as objects suggests accuracy assessment should incorporate some form of geometric validation of the classified objects against the real world objects they are meant to represent. Site-specific accuracy assessment methods, such as those associated with per-pixel classifications provide information on the accuracy of a classification at particular locations (x,y) across an image. Applied to GEOBIA classifications, there is uncertainty whether that class is consistent across the entire object. In addition, as the output of an object-based classification is ready for inclusion in GIS analysis, it should be assessed for the geometric accuracy (shape, symmetry and location) of the classified objects. This study describes a novel method of validating both the geometric and thematic accuracy of a multi-class classification against reference data. The accuracy assessment used a hierarchical object-based approach applied to randomly selected sample areas containing both classified (C) and reference (R) objects. Proportional overlap between the C and R objects, is used in the calculation of a number of measures of similarity that provide both thematic and spatial accuracies for the sample areas and classified objects within. In this study, the GEOBIA classification showed an overall accuracy of 72%, 90% sample areas showed a good match (>50% overlap) between C and R objects within and of these 11 out of 20 have over 70% correspondence. The measures of similarity also indicate strong correspondence between C and R objects within these samples. In the sample areas where there was poorer accuracy, and the omission and commission errors greater, the values for the dissimilarity measures were noticeably higher. Visual inspection of the sample areas shows that error is greatest in the sample areas with greater heterogeneity in cover types. Most of the non-matching objects occurred on the boundaries between land covers. This paper presents a novel application of area-based methods for the quality or accuracy assessment of object-based image analysis that have an advantage over the conventional site-specific assessment. Given appropriate reference data, the measures provide not only an overall accuracy for the image but also per object, per class and per sample area accuracies. In addition, classification uncertainty can be visualised enabling further analysis of error and where it occurs. * Corresponding author. Proceedings of the 4th GEOBIA, May 7-9, 2012 - Rio de Janeiro - Brazil. p.153 153