Earth Observation Semantics and Data Analytics for Coastal Environmental Areas

Kurzfassung

Current satellite images provide us with detailed information about the state of our planet, as well as about our technical infrastructure and human activities. A range of already existing commercial and scientific applications try to analyze the physical content and meaning of satellite images by exploiting the data of individual, multiple or temporal sequences of images. However, what we still need today are advanced tools to automatically analyze satellite images in order to extract and understand their full content and meaning. To remedy this exploration problem, we outline a highly automated and application-adapted data-mining and content interpretation system consisting of five main components, namely Data Sources (selection and storage of relevant images), Data Model Generation (patch cutting and generation of feature vectors), Database Management System (systematic data storage), Knowledge Discovery in Databases (clustering and content labeling), and Statistical Analytics (generation of classification maps). As test sites, we selected UNESCO-protected areas in Europe that include coastal areas for monitoring and an area known in the Mediterranean Sea that contains fish cages. The analyzed areas are: the Curonian Lagoon in Lithuania and Russia, the Danube Delta in Romania, the Hardangervidda in Norway, and the Wadden Sea in the Netherlands. For these areas, we are providing the results of our image content classification system consisting of image classification maps and additional statistical analytics based on three different use cases. The first use case is the detection of wind turbines vs. boats in the Wadden Sea. The second use case is the identification of fish cages/aquaculture along the Mediterranean coast. Finally, the third use case describes the differences between beaches, dams, dunes, and tidal flats in the Danube Delta, the Wadden Sea, etc. The average classification accuracy that we obtained is ranging from 80% to 95% depending on the type of available images.