Technical Summary: Benchmarking Deep Learning Models for Dense Event Classification of Offshore Wind Infrastructure in Sentinel-1 Time Series

Problem Statement

The global expansion of offshore wind energy, driven by ambitious targets in China, the European Union, and the United Kingdom, necessitates rigorous monitoring of infrastructure life cycles, particularly the deployment phase. While Earth Observation (EO) missions like ESA's Sentinel-1 Synthetic Aperture Radar (SAR) provide vast data archives suitable for global monitoring, extracting meaningful information from these high-volume archives remains challenging.

Previous approaches have largely relied on expert-driven, rule-based classifiers to interpret Sentinel-1 time series. While these methods can distinguish between "turbine" and "no-turbine" states, they often lack the semantic granularity to disentangle specific deployment events (e.g., site investigation, foundation construction, installation of nacelle/rotor) at a dense, single-acquisition level. Furthermore, existing global-scale products often aggregate data to monthly or quarterly resolutions, underestimating the complexity of the deployment process. The core problem addressed is the need for an automated, scalable algorithm that can perform dense event classification on Sentinel-1 time series to provide a nuanced, global view of offshore wind infrastructure deployment, moving beyond binary signals to detailed temporal sequences.

Methodology

Data Preparation

The study utilizes the DeepOWT v3.25.1 dataset, comprising 15,606 time series (14.8 million events) derived from Sentinel-1 GRD IW VH-polarized imagery (2016–2025).

• Feature Representation: 2D SAR patches are reduced to 1D backscatter intensity profiles along the range direction to capture spatial features like layover effects. Profiles are standardized (z-scored) and padded/cropped to a fixed length of 64 bins.
• Labeling: Events are classified into six classes: water, vessel, foundation, mooring/construction, turbine, and platform.
• Training Sets:
  • Supervised: 500 hand-annotated time series (661,732 events) were created. To address class imbalance (where vessel and mooring/construction are rare), specific sampling strategies were employed:
    • Monotemporal: Random undersampling of dominant classes and oversampling of minority classes to create balanced event-level datasets.
    • Multitemporal: Windows of 64 events centered around label transitions (unidirectional or bidirectional) to capture context.
  • Self-Supervised Learning (SSL): Leveraging the remaining 14.5 million unlabeled events, SSL datasets were constructed using rule-based classifier labels to mask and reconstruct event profiles or predict next events.

Model Architectures and Experimental Design

The study benchmarks ten model-training variants across three architectural streams:

1. Monotemporal (No Sequential Awareness): Fully Connected Neural Networks (FCNN) operating on individual event profiles.
2. Causal (Unidirectional Context): LSTM and Transformer models using only past context (masked self-attention), suitable for near-real-time inference.
3. Bidirectional (Full Context): BiLSTM and Transformer models using both past and future context (full self-attention), suitable for offline historical analysis.

Each architecture was tested in two configurations:

• Supervised Only: Trained directly on hand-labeled data.
• SSL + Supervised: Pretrained on unlabeled data via reconstruction/next-prediction tasks, followed by fine-tuning on labeled data using discriminative learning rates and gradual unfreezing.

Evaluation Metrics

Performance was evaluated on a held-out test set (553 time series) using:

• Area Under the Collapsed Edit Similarity Curve (AUCEditSim): The primary metric, measuring sequence alignment quality while accounting for class imbalance and transition sensitivity.
• Perfect Match Rate: The percentage of sequences with 100% collapsed edit similarity.
• Label Transition Analysis: Comparing the number of class transitions in predictions against the test set to assess sequence homogeneity and noise.
• Per-class F1, Precision, and Recall.

Key Contributions

1. Dataset Expansion: Release of 500 additional hand-labeled, globally distributed offshore wind infrastructure time series (661,732 events) to facilitate deep learning research.
2. Systematic Benchmarking: A structured comparison of ten deep learning variants (LSTM, Transformer, FCNN) across monotemporal, causal, and bidirectional contexts, with and without self-supervised pretraining.
3. Ensemble Strategy: Development of a label-transition-minimizing ensemble that combines the best-performing deep learning model with the existing rule-based baseline to optimize sequence consistency and recover performance on under-represented classes.
4. Global Deployment Analysis: Application of the improved labels to isolate deployment phases for 15,606 turbines globally, providing median deployment durations and regional dynamics for China, the EU, and the UK.

Results

Model Performance

• Bidirectional Superiority: All bidirectional deep learning models outperformed the rule-based baseline (AUCEditSim 0.7853). The supervised BiLSTM achieved the highest performance with an AUCEditSim of 0.8509 and a perfect match rate of 0.5063 (vs. 0.3508 for the baseline).
• Causal Models: Causal models (LSTM and Transformer) generally underperformed the rule-based baseline in AUCEditSim. However, Self-Supervised Learning (SSL) provided significant gains for causal Transformers, improving AUCEditSim from ~0.70 to ~0.75 and the perfect match rate from 0.15 to 0.26.
• Monotemporal Models: FCNNs performed poorly (AUCEditSim ~0.30), confirming that sequential context is essential for this task.
• Impact of SSL: For bidirectional models, SSL did not improve performance, suggesting that the supervised labels already provided sufficient representation learning. For causal models, SSL was beneficial, particularly for Transformers.

Ensemble Performance

Combining the supervised BiLSTM predictions with the rule-based baseline via a transition-minimizing ensemble yielded the most robust results:

• Sequence Consistency: Spearman's rank correlation for label transitions increased to 0.924 (vs. 0.89 for BiLSTM alone).
• Class Recovery: The ensemble recovered performance on under-represented classes. Notably, the platform class F1 score jumped from 0.4677 (BiLSTM) to 0.7352, matching the baseline's 0.7357. The foundation class also saw improvements, though it remained slightly below the baseline.

Regional Analysis

Using the improved labels, the study isolated deployment phases and reported median deployment durations:

• China: 84 days
• European Union: 242 days
• United Kingdom: 258 days

The analysis revealed distinct regional patterns:

• EU/UK: Construction concentrates in favorable windows (April–September). The UK exhibits a biennial pattern with longer deployment phases.
• China: Deployment activity shows sharp drops at year-end followed by surges in Q1, driven by regulatory subsidy deadlines (e.g., the 2021 feed-in tariff expiration). The Jiangsu (Rudong) cluster was identified as the world's densest offshore wind hub, with activity shifting to other provinces (Guangdong, Shandong, Zhejiang) as local subsidies expired.

Significance and Claims

The paper claims to provide the first systematic investigation of how deep learning can improve the dense classification of Sentinel-1 based offshore wind infrastructure time series on a global scale.

• Advancement over Rule-Based Systems: The study demonstrates that deep learning, specifically bidirectional architectures, can surpass expert-driven rule-based classifiers in sequence alignment and homogeneity, offering a more nuanced view of the deployment lifecycle.
• Practical Utility: The derived labels enable the isolation of specific deployment phases, moving beyond binary "turbine/no-turbine" signals. This allows for the quantification of deployment durations and the identification of regulatory and environmental drivers (e.g., subsidy deadlines, seasonal weather windows) at a global scale.
• Limitations and Future Work: The authors modestly note that while bidirectional models excel in offline analysis, causal models currently lag behind rule-based baselines, limiting their utility for near-real-time monitoring. They also highlight that the fixed 64-event window may miss long-range temporal patterns (e.g., implausible early foundation appearances), suggesting future work should address variable-length sequence modeling.
• Data Availability: The study emphasizes the public availability of the training data, predicted labels, and derived deployment phases to foster further research in offshore wind infrastructure monitoring.