Using the OpenStreetMap database to improve an object-based classification process. Application to a landcover product on small islands in the Indian Ocean

The southwest of the Indian Ocean comprises numerous islands of less than 3000 sq km (Comoros, Seychelles, Mascarene Islands). These small island territories have very fragmented and diversified environments. Satellite imagery and products at low and medium spatial resolution (usually from 4 kilometers to 250 meters) are not or poorly suited to the study of these areas. This is particularly the case for ecological studies such as the study of vector-borne diseases where the evaluation of interactions at a fine scale is critical to understand their spatial dynamics. To overcome this need, we realized a homogeneous land cover mapping of these small islands, by analysing SPOT 5 satellite images acquired between July 2013 and July 2014 by SEAS-OI Station. We used an object-based image analysis method to identify the 12 major classes of land cover / land use of these tropical islands. To improve the results of this classification, we used some data from the free and participative OpenStreetMap (OSM) database as training data. These data were first checked and even completed to ensure their quality. This information helped us in particular to improve our results on agricultural areas (sugar cane, market gardening) and to define more precisely the main roads. This methodology together with a good knowledge of the field has enabled us to achieve an overall accuracy of 85%, making it an operational product. In return, we fed the OSM database by integrating the classes for which we had the best accuracy (mangrove, surface water and forest). Indeed, OSM is a particularly interesting platform for releasing and disseminating land use data.