Ontologies Contribution to link thematic and remote sensing knowledge: preliminary discussions Samuel Andrés 1 Damien Arvor 1 Marie-Angélique Laporte 1 Laurent Durieux 1 Thérèse Libourel 1 Isabelle Mougenot 1 Christelle Pierkot 1 1 UMR 228 ESPACE-DEV 500 rue Jean-François Breton, 34093 Montpellier, France {firstname.lastname}@ird.fr Abstract. The semantic gap between image data and expert knowledge has been well identified for several years. Knowledge representation techniques such as ontologies are expected to help reducing this gap. Such techniques have already been used in different fields of application, from medical images to landscape pictures. We think they also have a great potential in remote sensing applications due to the interdisciplinary expert knowledge necessary for image interpretation. In this paper, we introduce what are ontologies and we discuss a potential ontological architecture for remote sensing applications. This architecture involves top-domain ontologies (OBOE, SWEET) and domain ontologies (land cover, image, spatio-temporal relations, protocol, sensor). We illustrate our thoughts with an example of Landsat TM image interpretation for detecting beaches. Keywords: Remote Sensing, Image Processing, Ontology, Expert knowledge 1. Introduction Global change is a major issue to be faced by today’s societies requiring interdisciplinary approaches involving thematic knowledge from diverse scientific domains such as geography, ecology, biology, etc. However, large part of this knowledge is focused on a specific scientific area and consequently is hardly shared with other domains. We argue that environmental sciences would be largely improved if heterogeneous expert domain-knowledge could be unified on the basis of shared core concepts. Earth Observation (EO) data has a major role to play for supporting interdisciplinary environmental researches. The development of enhanced remote sensors with increasingly high spatial, temporal and spectral resolutions allows the access to more details in satellite images, which has broadened the spectrum of remote sensing applications (see the nine Societal Benefit Areas defined by the Group on Earth Observation - GEO). Many scientific domains are now concerned by the use of EO data. As a consequence there is a need for improving the ability of researchers from different areas to process and interpret remote sensing images according to their specific purposes. One step toward such goal has consisted in implementing new image processing techniques allowing users to include their expert knowledge in the interpretation process. Especially, Geographic Object-Based Image Analysis (GEOBIA) represents a paradigm shift for image interpretation. Its main interest lies in its capacity for considering semantics based on a descriptive assessment and knowledge, i.e. it incorporates the wisdom of the user (Blaschke and Strobl, 2001; Hay and Castilla, 2008). However, although such approach has led to tremendous improvements in the interpretation of remote sensing images, it highlighted some important methodological issues: 1) each GEOBIA expert has its own conceptualization of the reality he intends to represent on the image; 2) the image processing is performed in a very laborious way at the end of