SyntEO: Synthetic training data generation for offshore wind energy infrastructure detection in Sentinel-1 imagery

Kurzfassung

Over the last decade, deep learning models have become increasingly important in the toolbox of the remote sensing community. Nevertheless, one major drawback is the necessity of large, precisely annotated training data sets in the widely used supervised optimisation approach. If a researcher wants to investigate an object class or land surface process for which no training data exists, a large amount of potential data which contains the target of interest has to be examined and annotated manually. Three major problems are related to this process, the availability of data showing the targets of interest, the time which has to be spent to explore and annotate the data, and the subjective annotation of the targets. To solve these issues, we present a synthetic data generation approach for Earth observation images called SyntEO. SyntEO is based on an ontology representing expert knowledge that describes the characteristics and relations of scene elements appearing in a remote sensing scene. Thereby, the domain expert does not include single fixed values to describe the characteristics of scene elements but multiple values contained in ranges, distributions or sample databases. That way, a data generator can combine highly variable scene element characteristics. An image generator uses the ontology to compose a two-dimensional, discrete scene composition. In a proceeding step, the texture is added to the discrete scene composition to generate the final image. The desired annotation is derived from the scene elements, which are defined as targets at the same time. To provide an intuition of an entire SyntEO workflow, we present the generation of synthetic Sentinel-1 scenes which show coastal areas and annotated offshore wind energy infrastructures. The synthetic training examples are then used to train a convolutional neural network in order to detect offshore wind energy infrastructure in real-world Sentinel-1 imagery on a global scale.