Application of object-based accuracy assessment for land cover classification using rapideye images in Southeastern Brazil

Abstract

Maps of land cover generated by Geographic Object-Based Image Analysis (GEOBIA) approach provide several advantages in relation to the pixel-based methodology, thus leading to significant gains for the accuracy of the final map. The accuracy indices that frequently express such gains are obtained from the traditional pixel-based validation. Then, given the absence of a validation representing the geographical object, new object-based methodologies to assess the accuracy have been developed with the similarity matrix STEP (Shape, Thematic, Edge and Position). In this sense, from the STEP methodology, this study aims to evaluate the thematic and geometric accuracy of land cover mapping generated by GEOBIA, as well as identify and analyse the main sources of errors. The study area is located in the municipalities of Sao Jose do Barreiro (Sao Paulo state) and Resende (Rio de Janeiro state), Southeastern Brazil. The map of land cover was generated from RapidEye Images acquired in year of 2011, with 5 m of spatial resolution. The global accuracy of the classification in relation to the traditional matrix was of 92%. The methodology integrates four similarity measurements (Shape, Thematic, Edge and Position), each one generating two error matrices (individual and aggregate thematic class). A global thematic accuracy of 76.6% was obtained, and for the measurements position (81.8%), edge (99.1%) and shape (93.3%). The methodology may be considered applicable and effective to validate thematic maps generated by GEOBIA. However, there are some limitations to be considered in relation to the size and extent of the objects.