BRIDGING REMOTE SENSING AND GIS – WHAT ARE THE MAIN SUPPORTIVE PILLARS ?

This paper highlights recent developments that have led to object-based image analysis (OBIA). We summarize major trends in bridging remote sensing, image processing and GIS and we hypothesize that OBIA is an emerging paradigm in image analysis. We identify two initial foundations for this paradigm shift, namely the advent of high resolution satellite data and the market entry of a commercial OBIA software package. Timewise, these developments fell on fertile ground and were accompanied by other, more sectoral approaches to cope with new requirements in automated image processing, analysis and interpretation. An increasing demand for geospatial information in the light of environmental pressure and monitoring needs has catalyzed the development of new methods to exploit image information more intelligently. Beyond an overview on the development of OBIA we briefly pinpoint some challenges that – from the authors’ view – arise in the retinue of the upcoming paradigm. * Corresponding author. 1. BRIDGING POTENTIAL AND A PARADIGM SHIFT In 2001, a workshop on “Remote Sensing and GIS – new sensors, innovative methods” held in Salzburg, focused on the new high resolutions satellite sensors of the ‘1m-generation’ and the increasing number of applications mainly based on the commercial software eCognition (www.definiens.com) which got commercial in 2000. In one of the workshop outcomes, Blaschke & Strobl (2001) have provokingly raised the question “What’s wrong with pixels”? These authors identified an increasing dissatisfaction in pixel-by-pixel image analysis. Although this critique was not entirely new (Cracknell 1998) they observed something like a hype in applications ‘beyond pixels’. A common denominator of these applications was that they were built on image segmentation. Interestingly, much of these developments were driven by German speaking countries and some western European countries, mainly. Consequently, much of this literature was available in German language (for an overview see Blaschke, 2002b). This is remarkable for at least two reasons. First, image segmentation which builds the basis of this approach has been introduced much earlier (see Haralick & Shapiro 1985, Pal & Pal 1993). Secondly, software driven scientific developments are generally more likely to make their way from North America to the rest of the world. In this short paper we concentrate on Earth applications. The majority of algorithm development in the late 1970ies and the 1980ies evolved in industrial image processing applications. Many of these algorithms use Markov Random fields or unsupervised texture segmentation (Jain & Farrokhnia 1991, Mao & Jain 1992, Pal & Pal 1993, Panjwani & Healey 1995, Chaudhuri & Sarkar 1995). Only more recently these algorithms are used for Earth applications (Dubuisson-Jolly & Gupta 2000). Over these last about five years from this first (mainly German speaking) workshop in 2001 to now, advances in computer technology, earth observation (EO) sensors and GIScience have led to the emerging field of „object-based image analysis” (OBIA). Although segmentation is a mature technique we observe the advent of a new approach that integrates segmentation with other methodological components for an optimised analysis of very high spatial resolution (VHSR) data in earth observation. The main asset is the general potential of OBIA to tackle the complexity and multi-scale characteristics of VHSR imagery. More specifically this potential builds upon two dimensions: (1) to use segmentation for representing a complex scene content in a set of scaled and nested (i.e. hierarchical) representations, and (2) to provide means to address this complexity by means of either a rule-based production systems that makes expert knowledge explicit and formalised, or a built-in adaptive (i.e. learning) mechanism, or both of it. Typically for a new technology-driven approach, we can observe a wide range of application papers in conference proceedings and ‘grey literature’. Literally dozens, if not hundreds of such papers from mainly younger scientists appeared in the years 2000 to 2004 and this development is still going on. A variety of successful applications exists (de Kok et al. 1999a, 1999b, Blaschke et al. 2000, Blaschke et al. 2001, Bauer & Steinnocher 2001, Blaschke 2002a, Schiewe & Tufte 2002, Pilz & Strobl 2002, Lang & Langanke, 2004; Neubert & Meinel 2002, Koch et al. 2003, Grubinger 2004, Mitri & Gitas 2004, Tiede et al. 2004; Langanke et al. 2004; van der Sande et al. 2004). Scientifically, some of these papers are disputable indicating some positivistic tendencies. This is not much different from the situation of GIS papers in the 1980ies. Starting from around 1989 we could observe a significant increase in methodological and theoretical research in GIS (Goodchild 1992). This trend of the first years of the 3 Millennium is, at least from the perspective of application-oriented research, often associated with one single commercial software. The perceived ‘omnipresence’ of this product, which was formerly called eCognition and is now available under the brand name Definiens, is an obstacle for the scientific community. Just in brief we shall mention at this point that this software was never the only option for scientists. Many different image segmentation algorithms are available for years (Pal & Pal, 1993, Hofmann & Böhner 1999, Blaschke 2000, Tilton 2003). Beyond algorithms also fully functional image segmentation software is available for several years (for a comparison see Meinel & Neubert 2004). In fact, the approach developed by Kettig and Landgrebe (1976) is still used and widely available in an open-source environment. Nevertheless, it is a fact that with the advent of the software eCognition in 2000 the number of segmentation-based image processing applications has taken off. Only in 2003 and 2004 first journal publications were published around the OBIA idea and/or applications based on eCognition. Burnett & Blaschke (2003) developed a multiscale segmentation / object relationship methodology (MSS/ORM) building on Koestler’s ideas of multi-levelled hierarchies and on an extended and on a more applied vision of a scaling ladder (Wu 1999). The multiscale segmentation based approach is designed to utilize information in the scales inherent in our spatial (image) data sets in addition to a range of auxiliary data sets, including for airborne and satellite data, but also to the scales of information inherent in single images. Flanders et al. (2003) describe transferability experiments of classifications by using eCognition but with a more technical focus. In 2004, Benz et al. publish an ‘eCognition paper’ in the ISPRS journal. Although there is also not too much of theory basis this paper is very often referenced since prior to its appearance for detailed explanations of the object-based methods beyond the underlying segmentation algorithm which was published earlier (Baatz & Schäpe, 2000) only the eCognition user guide could be referenced. Various authors have developed their own methods independently from the approaches mentioned so far. Gorte (1998), Melgani & Serpico (2002), Walter (2004) or Castilla et al. (2004), just to mention a few. Most of these approaches have in common that they are demonstrated for specific applications and their availability to other interested potential users is limited. Especially developments in academia very often demonstrate that something works in principle. It is more the exception rather than the rule that a commercial or open source software is developed out of these undertakings as it is the case with the algorithms of Kettig and Landgrebe (1976), Hofmann and Böhner (1999) or Tilton (2003). Clearly, OBIA has more roots than the software-centric ones and the selected developments mentioned above. A much more theory driven approach starting from the question of scale is represented by a Canadian group (Hay et al. 2001, 2002, 2003) and international co-workers (Hall et al. 2004, Castilla et al. 2004, Hay et al. 2005). They started very much from systematically exploring scale as a ‘window of perception’ (Marceau 1999). Meanwhile OBIA has also been introduced in text books and book chapters under this or similar terms (Jensen 2004, de Mer and de Jong 2005, Schöpfer et al, in press). The rapid spread and further maturation of the OBIA approach has triggered the demand for the “1 International Conference on Object-based Image Analysis – Bridging Remote Sensing and GIS (OBIA 2006)” which has been organized and hosted by the Centre for Geoinformatics (Salzburg University, Austria), and co-organized by ISPRS working groups IV/4 & VIII/11 as well as ESA. The two-days conference has likewise stimulated advanced methodological discussions on this paradigm shift in image analysis as it has opened the stage for presenting straight-forward, tangible solutions to problems of complex classification, change detection and accuracy assessment. Three conference themes were addressed from various aspects: (1) automated classification, mapping and updating techniques, (2) potential and problems of multiscale representation and (3) further development of standard methodologies. The world wide response to the conference announcement demonstrated that the challenge of linking methods and concepts from both remote sensing and geoinformatics is coupled with a range of expectations from various fields of applications including urban planning, mapping of settlements and infrastructures, forest management, land-use/land-cover mapping and change analysis, assessment of wetlands, habitats and species composition, natural resource management and geological exploitation, agricultural land use and crop monitoring.