APPLICATION OF OBJECT BASED APPROACH TO HETEROGENEOUS LAND COVER/USE U. Kanjir a *, T. Veljanovski a , A. Marsetič a , K. Oštir a a Institute of Anthropological and Spatial Studies, Scientific Research Centre of the Slovenian Academy of Sciences and Arts, Novi trg 2, 1000 Ljubljana, Slovenia – (ukanjir, tatjanav, alesm, kristof) @zrc-sazu.si KEY WORDS: object based image analysis, segmentation, delineation, shadows, agricultural and subalpine landscapes ABSTRACT: While analyzing satellite or aerial images different inconveniences can have negative effects on final classification accuracy. The objective of this paper is to expose the problems that were faced while classifying land use and land cover over two study areas in Slovenia. Two differently covered areas were chosen intentionally; the intensive agricultural area of Gornja Radgona in North- Eastern Slovenia and the subalpine area of Kobarid in the Western part of the country. By using object based image analysis we focus on two main problems; reducing negative effects of shadows on the image and misleading delineation of spectrally similar classes. Shadows that are present in the image and inconvenient for the interpretation were divided into two groups; small patches of shadows, which are result of higher objects, and bigger areas of shadows in the mountainous area. Both types of shadows are the consequence of the “inappropriate” position of the Sun at the time of acquisition and significant altitude difference over a relatively short distance. Another issue in the classification process is the delineation of segments that separate classes with similar spectral signature in the segmentation phase. Although segmentation parameters have been optimally set on a value where spectrally similar classes can still be separated, the delineations of classes do not always show the real situation. This paper presents approaches that were used to minimize both aforementioned problems and through which higher final classification accuracy was obtained. 1. INTRODUCTION The demand to automate image analysis in operational environments is constantly growing. Object based image analysis, which became an area of increasing research interest in the late 1990s offers an effective method for a good understanding of the Earth’s surface, especially on the high resolution (HR) images. Due to the occurrence of a large number of small objects all creating high contrast on HR imagery pixel based change metrics fail to operate successfully (taken from Wulder et. al, 2008, p. 355, Im and Jensen 2005, Niemeyer and Canty, 2003). Object based image analysis consist of two stages: contextual segmentation, where segments (objects) that have information related to shape, site and spatial relation (context) of the objects of the scene are created, and classification, where all created objects are analyzed and classified to the most representative class of land use/cover. It is important to explain here that with object oriented classification we are dealing with two concepts of objects: an image-object that is a discrete region of a digital image that is internally coherent and different from their surroundings, and that potentially represents – alone or in assemblage with other neighbours – a geo-object (Castilla and Hay, 2008). Although object based approach offers good final results with its fast, consistent and less subjective monitoring it still presents some disadvantages, especially in the stage of segmentation where creating correct shape of the image-object is desirable. One problem with object based classification is that there are no real objects recognised, but image objects, which can be spectrally confused. It is important to note that the accuracy and the significance of the final measurements, numbers and statistics directly and actually critically depend on the quality of segmentation (Baatz et al., 2008). Considering the large number of existing segmentation algorithms and their versatility (e.g. Guigues et al. 2006; Baatz and Schäpe 2000; Jung 2007; Hay et al. 2003; Pal and Pal 1993 Zhang 1997), the choice of an appropriate segmentation algorithm must rely on objective methods to assess segmentation quality (Radoux and Defourny, 2008). When segmenting and thus creating object shapes it is important to consider two aspects: 1) the appropriateness of an object's delineation (match of extend and shape with real situation) and 2) the precision of boundary delineation (match with scale applied) (Lang, 2008). Segmentation can be especially problematic in areas with low contrast or where different appearance does not imply different meaning. In this case the outcomes are represented as wrongly delineated image objects. Therefore, legend may include classes whose instances can barely be differentiated from each other in the image. The possibility exists that the boundaries of the classified image objects do not lead to an agreeable representation of real objects. This implies that there will be some classified image objects that need to be split or reshaped in part of their perimeter – the less clear the relation between two similarities, the more likely the possibility (Castilla and Hay, 2008). In general, shadows represent a great problem in remote sensing and symbolize a factor that considerably influences the results. Shadows increase the variability of the spectral response of a given feature and complicate the task of separating desired feature classes. One of the most descriptive characteristic is their low brightness. Shadows have been recognized as important determinants of canopy biophysical characteristics for many years (Richardson et al., 1975). Combinations of some factors (leaf, canopy or landscape levels) result in spatial variations of apparent shadow as observed from remote sensors (Gerard and North, 1997). In fact, changes in shadowing allow some remote sensing approaches, such as with multiangle observations, to estimate the structural attributes of ecosystems (Diner et al., 1999;