Pixel- versus object-based image analysis

Contrary to the visually-based, solely manual interpretation of imagery, automated digital image analysis techniques provide a time-saving solution and eliminate the influence of the interpreter’s subjectivity in land cover/land use, more specifically in vegetation delineation.

However, the appropriate use of these methods requires proper experiences in theory and practice. In general digital image analysis techniques aim at (semi-)automated information extraction, where the optimal approach depends primarily on the definition of the output products (e.g., the type of the maps) and is influenced by spatial resolution and inter-pixel variance (Wulder et al., 2004). In traditional image classification methods the pixels of an image are examined in order to give particular class labels to them. The application of pixel-based approaches has been well-accepted at low or moderate spatial resolution (Ozesmi and Bauer,2002), whereas at high spatial resolution it could be problematic, since an individual feature could often be classified in distinct categories because of consisting multiple pixels with variable spectral reflectance characteristics (Johansen et al.,2010b).

Several remote sensing studies (e.g.,Addink et al.,2007;Levick and Rogers,2008;Kamagata et al.,2008;Johansen et al.,2010a) showed that due to the heterogeneous nature of the target vegetation habitats in high spatial resolution, traditional pixel-based digital image classifiers do not give satisfactory results. Instead of that the analysis of object-based methods came to the focus of interest.

Chapter 2. Classifying riparian vegetation based on remotely sensed images 12 During the analysis of high spatial resolution imagery the target features are generally larger than pixel size (Johansen et al.,2010b), which makes it very much reasonable to analyse the images based on objects like groups of pixels, instead of analysing the pixels individually.

This special type of analysis method named object-based image analysis (OBIA) - also called elsewhere as object-oriented image analysis (OOIA) - applies actually the first law of geography (Tobler, 1970), which says: “Everything is related to everything else, but near things are more related than distant things”. In that sense certain image pixels situated next to each other could be merged beforehand and be handled together. With the increasing amount of data with high spatial resolution and the development of automated digital image analysis techniques the principle of grouping pixels into meaningful objects before the classification has become crucial. This need triggered the release of the first commercially available OBIA software (eCognition, from 2000).

Object-based classifiers can use spectral and spatial patterns together for the image anal-ysis and thereby involving contextual information (Lillesand et al., 2008) and overcoming the so-called “salt and pepper effect” (Blaschke, 2010). OBIA consists of image segmenta-tion (clustering of pixels into homogeneous objects), classificasegmenta-tion (or labelling objects) and modelling based on the characteristics of objects (Johansen et al., 2010b). The following methods, e.g., edge-detection, feature extraction and classification involved in OBIA have already been used in remote sensing image analysis for decades, whereas the image segmen-tation itself was not applied extensively in geospatial applications between 1980 and 2000, although it has not been a new concept coming originally from industrial image processing (Blaschke et al.,2011). Segmentation approaches currently applied in the present research are discussed in Chapter4.1, for a more detailed review on the available segmentation meth-ods applied in the eCognition object-based software environment the reader is referred to Dezső et al.(2012).

Due to the capability of building a logical hierarchical structure between different scales of image objects (Benz et al.,2004), OBIA has a high potential in multiscale landscape analysis, where semantically significant regions are found at different spatial scales (Burnett and Blaschke,2003). The characteristics of meaningful objects can be assessed through spatial, spectral and temporal scales to generate new spatial information in GIS-ready formats representing its compatibility with vector GIS-software (Johansen et al.,2010b).

OBIA is becoming a standard image analysis approach for the analysis and extraction of

GIS-ready spatial information from VHR/HR imagery and its significance has been revealed through biennial conferences (GEOBIA 2006, 2008, 2010, 2012, 2014), the establishment of specific teaching activities in academia, special journal issues and beyond that, private and government agencies adopted OBIA-technique as an integral part of GI Science and spatial information generation (Johansen et al.,2010b).

The comparison of pixel- and object-based analysis of vegetation emphasizing the improve-ment of accuracies for the OBIA-method has been presented in various research studies applied to HR imagery (Yu et al., 2006;Yan et al., 2006; Addink et al., 2007; Levick and Rogers, 2008; Kamagata et al., 2008; Johansen et al., 2010a). For a general visual com-parison of pixel-based and object-based classification results an example is presented in Figure2.2, based on an aerial photo from the Szigetköz floodplain, which was latter used in the present research. Both types of supervised classifications (maximum likelihood for the pixel-based and class description based fuzzy algorithm for the object-based method, see later in Chapter 4.5) were based on the original bands complemented with the bands for the first principal component and for the vegetation index, after the selection of the same training samples.

Yu et al. (2006) presented improved vegetation classification results in Northern California for the object-based method based on the imagery of Digital Airborne Imaging System (a 12-bit multispectral imaging sensor system) with 1-m spatial resolution in comparison to the pixel-based analysis, overcoming the problem of salt-and-pepper effects often found in the traditional pixel-based approaches. In the study ofLevick and Rogers (2008) classification accuracy improvements by OBIA have been presented for the analysis of woody vegetation in a heterogeneous savanna system in South Africa based on black & white aerial photographs.

Johansen et al. (2010a) highlighted the advantage of geo-object-based classification in the analysis of temporal changes concerning riparian land-cover classes based on QuickBird images (under 3 m/pixel) in Central Queensland, Australia. Nevertheless, Blaschke et al.

(2011) have stated that an optimal use of object-based image classification and mapping for vegetation-related analysis is still under consideration regarding consistency and time-efficiency. Besides that, concerning the general concept of vegetation mapping Xie et al.

(2008) emphasized that the search for improved image classification algorithms presents an actual research field in the remote sensing applications because of the lack of those classification methods which could be universally applicable.

Chapter 2. Classifying riparian vegetation based on remotely sensed images 14

Figure 2.2: Comparison of pixel- and object-based classification approaches based on the latter applied aerial image scene of the Dunaremete test site, from 2008, with a spatial

resolution of 1.25 m/pixel.

Materials