ASSESSING THEMATIC AND GEOMETRIC ACCURACY FOR REMOTE SENSING PRODUCT CLASSIFICATION PROCESS: PER-OBJECT APPROACH

Thematic accuracy assessments are usually based on non-spatial statistics, which summarize the characteristics of an error matrix. These statistics do not normally consider the spatial distribution of pixels that are wrongly classified. Consequently, thematic accuracy assessment using a pixel approach has obvious limitations when applied to object-based image analysis. Thus, it is necessary to develop methods to assess both the thematic accuracy and the geometric accuracy (location and shape) of classified objects. This article aims to propose an object-based error matrix (OBEM) for thematic validations of classifications following the per-object approach, based on the calculation of similarity and positional measures. This OBEM makes it possible to assess both the thematic accuracy and the geometric accuracy (location and shape) of classified regions or objects. Results show that the spatial incidences of error arising from the thematic classification process present a characteristic and spatially correlated pattern. These errors are usually grouped, with a high proportion of erroneously labeled pixels occurring in the vicinity of the boundaries between spatial features, regardless of the approach used. Research needs to be carried out before the correct spatial characterization of errors from the classification process to be properly communicated in standardized formats and legends.