An Approach for Accuracy Assessment Comparison between Per-pixel Supervised and Object-oriented Classifications on a QuickBird Image

Abstract submitted to "30th EARSeL Symposium: Remote Sensing for Science, Education and Culture"
An Approach for Accuracy Assessment Comparison between Per-pixel Supervised and Object-oriented Classifications on a QuickBird Image
Vassil Vassilev
Space Research Institute - Bulgarian Academy of Sciences
Keywords: QuickBird, Land use/land cover, Image analysis, per-pixel supervised classification, object-oriented classification, Feature Analyst
Presentation preference: oral

An approach for accuracy assessment comparison between two types of Land use/Land cover image classifications based on extraction of thematic information from very high resolution multi-spectral QuickBird image acquired on 31.05.2008 is presented in the present paper. Two types of classifications are compared in the present paper: 1) per-pixel supervised classification, and 2) object-oriented classification. The object-oriented classification was applied in ArcGIS software using the Feature Analyst 4.2. extension tool, while for the per-pixel supervised classification ERDAS Imagine software was used. The proposed approach includes several work stages and it has been applied on a highly fragmented urban and agricultural land, district of Novi Iskur, Sofia municipality, Bulgaria. A land use/land cover classification scheme for the area studied was created. A large scale land use map for the Novi Iskur district is composed on the base of the final results, and the differences of the land use classes are assessed using image analysis. Essential part of this approach for image processing is using a combination of spectral reflectance and texture differences for extracting different land use/land cover classes. Unsupervised classification, fuzzy convolution filter, relief data, and Normalized difference vegetation index (NDVI), were supplemented and used as an ancillary data in the classification process for additional analysis. The accuracy assessment was calculated for both classifications using accuracy assessment tool in ERDAS Imagine software. It was found that the object-oriented classification has better overall classification accuracy (94.15%) than the per-pixel supervised classification (89.51%), and the overall Kappa statistics are respectively 0.9335 and 0.8776. It can be concluded that the object-oriented classification deals more sufficiently with the urban environment including streets and buildings, while the per-pixel supervised classification is more efficient regarding the forest canopy and agricultural land after applying fuzzy convolution filter to reduce the mixed pixel problem. The presented approach proposes an opportunity for quick and objective accuracy assessment evaluation of thematic information from multi spectral image classifications. The accurate detection of land use/land cover classes can be achieved only by having expert knowledge on what kind of classification is appropriate for each particular ground surface. The ability to use different kinds of image classifications, either per-pixel supervised or object-oriented, and utilizing their benefits is vital for making each classification process a success.