Object Oriented Classification: Case Studies Using Different Image Types with Different Spatial Resolutions

This paper presents the case for information extraction from images using what has been termed “object oriented classification” (OOC). Traditional methods of computer classification of images focus of the colour or spectra of individual pixels. However, human interpretation of images uses many more image attributes, including relative size, shape, distribution (texture and pattern) and context of groups of pixels which are spatially associated together to form “objects”. Image segmentation, a process of examining the spatial clustering of pixels which display some degree of homogeneity at a particular scale, has been used in the last decade to assist image classification, particularly when the spectral dimensionality of an image is low. More recently, OOC has become available in off the shelf software. OOC provides a process of computer based learning in respect to not only colour, but also in respect to size, shape, orientation, texture and to some extent, context of groups of pixels. This paper illustrates the use of OOC image classifiers via the use of four examples. Firstly in the stratification of ecological zones an Australian desert using six banded Landsat TM data, the software eCognitionTM from Definiens was used with better results compared with pixel based methods. Secondly the results of classification of high resolution multi-spectral and panchromatic QuickBird satellite imagery over South Australian vineyards, which was the subject of undergraduate UniSA student research by Martin Nolan, are presented. In his work Nolan was trying to achieve or exceed 96% accuracy as compared with human interpretation combined with field work. Results of using the object oriented classifier, in this case Feature AnalystTM from Visual Learning Systems, showed improvement over the use of traditional supervised classification, but did not meet the high tolerances required for this technique to be universally adopted. Thirdly, OOC was applied to classifying urban features in a rural township in South Australia. Feature AnalystTM was again used on multi-spectral Quickbird imagery and compared with supervised classification. Whilst in the main good results were observed, shadows from tall trees, and larger buildings created classification uncertainty. Finally, the paper examines the results of applying OCC to very high resolution (pixel size of 0.15m) aerial, multi-spectral orthophotos over an urban area, where the objective was to automatically extract building roof information. Results show that OOC software, in this case Feature Analyst, provided a very quick method of directly extracting GIS polygons from the imagery. Hierarchical learning used in the project requires that context from landuse and cadastre is utilized to assist the differentiation of grey roof tops from roads. Other learning is based on the shape of roofs versus the shape of paved driveways, which share similar areas and colours to some roof tops, but different shapes. Again shadows presented significant problems for the OOC algorithm and suggest that whilst humans interpret object height from shadows, this attribute needs to be explicitly injected into the computer classification methodology. Airborne laser scanning (ALS) is currently being evaluated for provision of this additional attribute. However, it is noted that relief displacement in the orthophoto imagery will cause geometric displacement of roofs relative to their correct position and that observed in the ALS data. In conclusion the paper summarizes accuracy from the four examples to show how, for at least images exhibiting low spectral dimensionality, object oriented techniques are superior to the traditional pixel based methods, but still inferior to human interpretation.