Ontology Based Framework for Geospatial Objects Retrieval in Remote Sensing Images Rainer Larin-Fonseca; Eduardo Garea-Llano Advanced Technologies Application Centre CENATAV 7ma A. No. 21406, Playa, Havana, Cuba (rlarin, egarea)@cenatav.co.cu Abstract— Detection and retrieval of geospatial objects in remote sensing images is a challenging and expensive task. The main problem that we are facing in this paper is the detection and retrieval of geospatial objects based on their semantic abstraction. We present a geographical object retrieval framework. This framework employs ontology-based techniques for detecting and retrieving geospatial objects in remote sensing images. The semantic gap in the geospatial object retrieval is reduced by the use of the Data-Representation Ontology (DRO) combined with High Level Ontologies (HLO). This is achieved by representing scale and rotation invariants low levels features in the DRO and representing the semantic abstraction of these objects in the HLO. For validating the proposal, an experiment based on geospatial objects retrieval is presented showing good results. Keywords—semantics, object, detection, retrieval, images. I. INTRODUCTION Remote sensing images are one of the most common data sources used in Geospatial Information Systems (GIS). With the development of the new technologies the amount of remote sensing images has considerably increased as well as their spatial resolution and extension. Searching and retrieving geospatial objects contained in these images are both challenging and expensive tasks from the computational point of view. Therefore, the creation of a mechanism for carrying out these tasks automatically is very important. The use of ontologies for geospatial object retrieval can reduce the semantic gap since they can be used as a middle layer between end-users and computers allowing a better understanding between them. In this paper, an ontology-based framework for geospatial object retrieval in remote sensing images is proposed. We assume that the geospatial objects to be retrieved can have both scale and rotation problems. Another aspect to be considered is the insufficient number of instances of these objects therefore; the thousands of training examples required by several classification algorithms are not available. The proposal presented in this paper is ontology-based and it uses three ontologies to provide the semantics of domain context and of task context. The remainder of this paper is organized as follows. In Section 1, we review related works. In the section 2 the proposal is presented. In Section 3 the main experimental results are shown and finally the main conclusions are stated in Section 4. II. RELATED WORKS Several ontology-based approaches have been proposed [1],[2],[3]. These approaches use ontologies for structuring and managing knowledge. Lately, there is an increasing interest in the use of ontologies because they are essential to create human computer interfaces and to face semantic gap problems. An example is the use of ontologies as middle layer between end user queries and low level image features. Thus, a conceptual level is included allowing to use semantic concepts directly in the query. A. Object detection and retrieval in remote sensing images Some approaches like CBIR [4] uses low-level features such as color, texture and shape to retrieve images, however the results are still not good enough. This is mainly due to the semantic gap between visual features and semantic concepts. The SIGMA approach [5] is strongly domain-dependent since prior knowledge about the scene is integrated in their algorithms for image understanding. Moreover it is based on knowledge bases which are difficult to produce. Some recent approaches such as the presented in [6], [7] are based on shape features on target segmentation. Nevertheless they can only be used in some specific cases. Methods based on machine learning algorithms such as presented in [8] have a limited applicability in real applications when insufficient leaning data is available. However, the use of invariant local descriptors can provide relevant results for specific object retrieval task [9],[10],[11]. Recent works use ontologies to explicitly describe knowledge domains. In [2] and [3], the authors present ontology-based approaches for semantic-based remote sensing image retrieval. In [12] the authors propose an ontology-based recognition method. This approach uses very high resolution (spatial or spectral) remote sensing images; therefore it does