Object-based classification algorithms for mapping

Abstract

An adequate maintenance and protection of urban green spaces requires update and accurate information of these features. Remote sensing techniques provide an important source of information to automate urban land-cover mapping. Nevertheless, some techniques tend to be more appropriate than others in distinguishing specific categories, such as vegetation, particularly when classifying high resolution imagery in urban environments. Previous studies have shown that the conventional pixel by pixel classification cannot obtain very satisfactory results in urban spaces whereas an Object-Oriented approach tends to achieved better results (e.g. Laliberte et al., 2007, Cleve et al., 2008). This paper aims to assess the suitability and effectiveness of using object-based classification of high resolution Ikonos imagery to map and resolve different types of green spaces in the city of Kuala Lumpur. A set of fuzzy membership rules was developed within a training subset of IKONOS imagery for identifying four land-cover classes in different segmentation levels. Focus was placed on the vegetation class. The classification result for the vegetation class yielded to producer’s accuracies higher than 64% and user’s accuracies higher than 90%, suggesting that this is an efficient approach to deal with the spatial and spectral variability of urban environment to map general vegetation spaces. However improvement is needed in order to discriminate between different vegetation types and qualities.