DeepOWT v5.25.1：深層学習による洋上風力インフラ向けの高密度Sentinel-1時系列データセット

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:深層学習とSAR時系列を組み合わせた洋上風力モニタリング手法と、公開ベンチマークデータの設計を参考にできる。

🏢実務担当者:自社の洋上風力資産やサプライチェーンの設置・稼働状況を衛星データで外部検証する用途に活用できる。

🏛政策担当者:再エネ導入進捗や洋上風力展開の独立モニタリング基盤として、政策評価や統計補完に活用できる。

DeepOWT (deep learning derived global offshore wind turbines) is an independent and openly accessible data set of offshore wind energy infrastructure locations and their temporal dynamics on a global scale. Locations are derived by applying deep learning based object detection on ESA's spaceborne Sentinel-1 synthetic aperture radar (SAR) archive, see Global Offshore Wind Infrastructure: Deployment and Operational Dynamics from Dense Sentinel-1 Time Series Dense time series are derived by inspecting each available Sentinel-1 scene at every detected infrastructure location. One-dimensional swath profiles are generated for each acquisition and infrastructure location, showing the maximum SAR backscatter value along the horizontal axis (range direction), thereby capturing directed SAR signatures. The data set is organised into three thematic groups ( spatial/ , temporal/ , training/ ). The dense time series is enriched with deep learning derived event predictions, and deployment event flags, derived from the BiLSTM predictions, indicating start and end of a deployment phase. With version v5.25.1 the event predictions come from a seed ensemble of ten BiLSTM models with self-supervised pre-training, the hand-labelled event labels are extended by 500 training facilities, and a hand-reviewed deployment start/end reference for 1,326 facilities is added. Machine-learning-ready data sets for self-supervised pre-training and supervised training are published as well. Related publications: Publication Data set version Added feature DeepOWT: a global offshore wind turbine data set derived with deep learning from Sentinel-1 data v1.21.2 Global OWT locations 2016Q3 - 2021Q2 hand labelled test locations North Sea Basin, and East China Sea for 2021Q2 Global Offshore Wind Infrastructure: Deployment and Operational Dynamics from Dense Sentinel-1 Time Series v3.25.1 Global OWT locations 2016Q1 - 2025Q1 hand labelled test locations North Sea Basin, East China Sea, and South East Vietnam for 2025Q1 1D SAR profile time series of all S-1 acquisition for each location + automatically derived semantic labels + hand labelled benchmark set Benchmarking Deep Learning Models for Dense Event Classification of Offshore Wind Infrastructure in Sentinel-1 Time Series (first submission 30.06.2026) v4/5.25.1 v5.25.1 is the revised version of v4.25.1 extended by deployment phase labels and reprocessed predictions. reorganised distribution into  spatial/ , temporal/ , and training/ groups dense 1D SAR profile time series enriched with deep learning event predictions, and per-unit deployment start/end event flags derived from the BiLSTM predictions hand-labelled event labels for 500 training facilities added to the 553 test facilities hand-labelled deployment start/end reference for 1,326 facilities machine-learning-ready training data sets for self-supervised pre-training and supervised training (monotemporal samples and uni-/bidirectional sequence windows) The dense event classification distinguishes six semantic states of an infrastructure location over time: water , vessel , platform , turbine foundation , mooring / active construction , and deployed turbine . The rule-based baseline additionally uses unclear for ambiguous acquisitions. File metadata File(s) Time Geometry Spatial extent Temporal resolution spatial/DeepOWT_pnt_locations.parquet (Derived Locations, 15,606 units) 2025Q1 points Global quarterly spatial/location_validation_2025Q1.parquet (Ground Truth Location, 9,770 polygons) 2025Q1 polygons North Sea Basin, East China Sea, Southeast Vietnam - temporal/deepowt_time_series.parquet (Analysis Ready Time Series, Derived Baseline & Deep Learning Labels, Hand Labels, Deployment Event Flags and Reviewed Deployment Reference Flags) 2016Q1-2025Q1 - (index into DeepOWT points via unit_id) - for each available S1-acquisition (~1-12 days) temporal/deployment_reference.parquet (Hand-Reviewed Deployment Start/End, 1,326 units) 2016Q1-2025Q1 - (index into temporal/deepowt_time_series.parquet via unit_id / sequence_id) - one row per unit training/training_monotemporal.parquet, training/training_windows_unidirectional.parquet, training/training_windows_bidirectional.parquet (Supervised Training Samples) 2016Q1-2025Q1 - (index into temporal/deepowt_time_series.parquet via unit_id / sequence_id) - - training/ssl_monotemporal.parquet, training/ssl_windows_unidirectional.parquet, training/ssl_windows_bidirectional.parquet (Self-Supervised Pre-training Samples) 2016Q1-2025Q1 - (index into temporal/deepowt_time_series.parquet via unit_id / sequence_id) - - Column description of temporal/deepowt_time_series.parquet Column Type Description unit_id string Infrastructure location identifier; joins to spatial/DeepOWT_pnt_locations.parquet. acquisition_date string Acquisition timestamp of the Sentinel-1 scene (ISO 8601). orbit_direction string Sentinel-1 orbit direction (ascending / descending). swath_profile_horizontal_max array of double One-dimensional swath profile of maximum SAR backscatter along the horizontal axis. baseline_label string Event label from the rule-based baseline classifier. sequence_id integer Zero-based position of the acquisition within the unit's chronological time series. gold_label string Hand-labelled event label of the 553 test facilities and the 500 training facilities; empty for all other facilities. is_test boolean True for acquisitions belonging to the hand-labelled benchmark test set. Use to create the hold out test (benchmark) subset from gold_label; gold_label with is_test = False is the training subset. bilstm_label string Event label predicted by the seed ensemble of ten BiLSTM models with self-supervised pre-training. ensemble_source string Which model provided the ensemble label for this unit (baseline or bilstm): the source with fewer label transitions over the unit's time series, ties go to bilstm. ensemble_label string Ensemble event label, fuses baseline and BiLSTM predictions. is_depl_start boolean True on the single event marking the start of the unit's deployment phase, derived from bilstm_label; a water/vessel run of at least 2 observations before the deployment end gates the search for the start. is_depl_end boolean True on the single event marking the end of the unit's deployment phase (first confirmed deployed turbine), derived from bilstm_label. is_depl_start_ref boolean True on the hand-reviewed deployment start of the unit (see temporal/deployment_reference.parquet); False for units without a reviewed start. is_depl_end_ref boolean True on the hand-reviewed deployment end of the unit (see temporal/deployment_reference.parquet); False for units without a reviewed end. Column description of the training data sets File group Columns Description *_monotemporal.parquet unit_id, sequence_id Single 1D swath profiles, referenced into temporal/deepowt_time_series.parquet via (unit_id, sequence_id). *_windows_unidirectional.parquet, *_windows_bidirectional.parquet unit_id, start_id, end_id, change_sequence_id, from_label, to_label, is_synth_regression Sequence windows spanning start_id..end_id with a label change at change_sequence_id (from_label to to_label). Unidirectional windows end ~at the change; bidirectional windows extend around it. is_synth_regression flags synthetically generated regression samples. The training/training_* files provide the labelled samples used for supervised training (the 500 training facilities with gold_label and is_test = False); the training/ssl_* files provide the (label-free with respect to human annotation) samples used for self-supervised pre-training. Both index into the dense time series temporal/deepowt_time_series.parquet.  

• Zenodo https://zenodo.org/records/23160399first seen 2026-10-06 04:18:27