Kleebauer, MaximilianMaximilianKleebauerBraun, AxelAxelBraunHorst, DanielDanielHorstPape, CarstenCarstenPape2024-10-312025-02-172025-05-052024-10-312024https://publica.fraunhofer.de/handle/publica/47824010.1109/IGARSS53475.2024.10641018Remote sensing and deep learning-based methods can be combined to obtain location information automatically on a large scale. This paper introduces an approach for enhancing the geo-coordinate accuracy of existing wind turbines. By employing a RetinaNet-based method for regressive object localization, turbines can be precisely located in images in addition to being identified. Utilizing semi-automatically processed and manually filtered high-resolution image data, a model is trained with an average precision of 96 %. Subsequently, the model is applied to Germany’s MaStR wind turbine dataset. The application illustrates the advantageous implementation of the method and emphasizes its considerable potential for improving the accuracy of geo-coordinates. While 73.72 % of existing coordinates can be confirmed as correct with a deviation of less than 10 meter, for more than 15 % of the turbine locations coordinates between 10 and 100 meters can be corrected, and for 5.6 % locations a deviation of more than 100 meter can be determined. This showcases the real-world application of the proposed methodology and underscores its significant potential for enhancing the quality of geo-coordinates.enremote sensingwind turbinesrenewable energy systemsobject detectionobject regressiongeo-coordinate validationretinanetMarktstammdatenregisterconference paper