Remote sensing derived onshore wind turbine locations for Germany

The rapid expansion of renewable energy sources poses significant challenges in reconciling energy development with competing interests. This underscores the necessity for precise spatial data to facilitate effective balancing, management, or evaluation of compliance with regulatory frameworks.

In Germany, a widely used data source for location and plant specific data is the [Marktstammdatenregister](https://www.marktstammdatenregister.de/MaStR) operated by the Federal Network Agency. This dataset builds upon the Marktstammdatenregister and enhances its spatial accuracy of onshore wind power plant locations using high resolution remote sensing data and object detection deep learning.

The dataset consists of point geometries for onshore wind turbines in Germany. The wind turbines were detected using YOLO an object detection algorithm, on high-resolution PlanetScope satellite image time series. The inference on PlanetScope time series, in particular Global Monthly basemaps, include basemaps for the months of April to October for the years 2018 to 2024 with 45 German wide inferences in total. The training dataset was build using wind turbine sites from Manske & Schmiedt (2023) from the UFZ, which semi-manually corrected the site information of the Marktstammdatenregister for the years 2021, 2022 and 2023. Hereby the training dataset was curated by evenly sampling wind power plants based on geographic location. Additionally negative samples were added to the dataset both picked randomly from all over Germany, excluding wind power plant sites by a certain distance, as well as from specific objects prone to be detected as false positives such as power towers, cell phone towers or similar based on Open Street Map.

The dataset was refined in multiple iterations by filtering bad initial labels which could not be re-detected by a trained model over multiple time steps, as well as by adding more samples of scenes where the model had difficulties.

The resulting monthly detections of onshore wind turbines over Germany were aggregated throughout time by merging all detections to a single detection time series if it would occur in a certain buffer throughout time. Hereby, we added the detected time stamp time series as an addition attribute to the resulting aggregate points. In order to filter detection noise, the dataset can individually be filtered for points which exceed a certain detection count.

In future works we will update this dataset to recent month, an enhanced post-processing to filter false detections, as well as links to IDs of the German Marktstammdatenregister per detection.