Synthetic Training Enables Deployment on Raw Drone Data: An Attention-Based Framework for Detecting Orphan Wells

Area of Science:

• Geophysics
• Artificial Intelligence
• Environmental Science

Background:

• Undocumented orphan wells pose significant subsurface characterization and environmental management challenges.
• Magnetic surveys can identify wells via steel casing anomalies, but data processing is complex and inefficient.
• Conventional methods struggle with high-volume, noisy magnetometer data, yielding poor recall scores.

Purpose of the Study:

• To develop an efficient deep learning framework for processing hyper-resolute magnetic survey data.
• To enable accurate detection of undocumented orphan wells without extensive data downsampling.
• To improve the efficiency and scalability of remote sensing for environmental applications.

Main Methods:

• A transformer-based deep learning framework was proposed.
• Novel on-the-fly techniques and sinusoidal positional encoders were utilized for relative positional awareness.
• The model was tested on synthetic and real-life magnetic survey data.

Main Results:

• The model achieved F1-scores exceeding 90% on synthetic data for line spacings up to 140 m.
• This allows for significantly sparser flight paths, enhancing survey efficiency.
• A 70% recall rate was achieved when applied to real-life data.

Conclusions:

• The proposed deep learning framework offers an efficient and scalable solution for detecting undocumented orphan wells.
• The approach facilitates more efficient drone-based magnetic surveys.
• The framework is adaptable to other remote sensing applications requiring complex data processing.