SyntEO: Synthetic dataset generation for earth observation and deep learning – Demonstrated for offshore wind farm detection

DeepOWT: a global offshore wind turbine data set derived with deep learning from Sentinel-1 data

“…Multiple studies have demonstrated that CNNs can learn the spatial representation of target classes in Sentinel-1 images and that they can also consider the spatial context in these images to reduce false positives (Dirscherl et al, 2021;Kang et al, 2017;Hoeser and Kuenzer, 2022a). This property of CNNs is a particularly important argument for their use in extracting object classes from extensive, unfiltered satellite data archives, as demonstrated in this study.…”

“…In order to train both object detector CNNs, two synthetic training data sets were generated in a preceding step. This is the first application of the recently introduced SyntEO approach (Hoeser and Kuenzer, 2022a) embedded in a complete workflow to generate a global data set. After the OWT locations are detected in the 2021Q2 data, a time series of 19 quarterly periods is investigated to derive deployment dynamics for each OWT location from 2016Q3 until 2021Q1.…”

“…All of the aforementioned studies have the commonality that they strongly depend on the high contrast between offshore infrastructure and the surrounding open sea. In a preceding study to this data set, Hoeser and Kuenzer (2022a) proposed an adaptive, deep-learning-based object detection approach that takes multiscale spatial patterns of the target objects into account to distinguish further different types of offshore infrastructure in a single model. Radar imagery, especially that provided by the Sentinel-1 mission, is an auspicious data source for studies related to the offshore wind energy sector.…”

“…the application of the recently proposed SyntEO framework by Hoeser and Kuenzer (2022a) for using synthetic data to train the two supervised deep learning models;…”

“…The proposed DeepOWT data set has been derived from the Sentinel-1 radar archive by applying deep-learning-based object detection. The employed object detection models were completely trained on synthetic training examples generated by the novel Synthetic data generation in Earth Observation (SyntEO) approach (Hoeser and Kuenzer, 2022a). Thus, the methodological workflow enables the extraction of highly detailed information from extensive Earth observation archives.…”