Global Offshore Wind Energy Infrastrucutre Dynamics Derived from Sentinel-1 Imagery with CNNs based on Synthetic Training Data

Abstract

Over the last years, deep learning has become an important component in the Earth observation toolset. Especially the convolutional neural network is the most widely used deep learning model in Earth observation. The supervised optimisation of neural networks relies on large datasets, which are necessary to predict on complex data and train models to be transferable in time and space. In contrast to the efficient processing of large data archives by trained neural networks is their need for large training datasets, which are labour-intensive to create. Another drawback is that only those research questions can be investigated where enough data are available to build datasets large enough to train a deep learning model. In order to solve the problem of labour-intensive data annotation and a potential lack of raw data, we have developed SyntEO, an approach to synthetically generate Earth observation data and corresponding labels simultaneously. This approach specifically addresses the needs of Earth observation data and composes a remote sensing scene with harmonised spatial and temporal order of nested entities. SyntEO uses an ontology formulated by domain experts to make their knowledge explicit and machine-readable. Upon that ontology, an artificial data generator composes an abstract scene composition that is used to finally generate the synthetic remote sensing scene by adding texture and to derive the corresponding label. To give an intuitive introduction to SyntEO, we demonstrate the detection of offshore wind farms and their components by using deep learning models that are only trained with synthetic data generated with the new approach. The resulting deep learning models detect offshore wind farms as well as single offshore wind turbines in real-world remote sensing imagery. The underlying data are IW GRD acquisitions of the Sentinel-1 mission in VH polarisation, which lie within a distance of 200 km of the coastline towards the sea. The trained models are used to detect offshore wind farms and turbines for the entire global coastline at a quarterly frequency between 2016 and 2021. The results are validated by assessing their performance on a hand labelled ground truth data set which includes all offshore wind turbines in the North Sea Basin and the East China Sea.