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

Undocumented orphan wells present challenges for subsurface characterization and environmental management due to their unknown locations and varied physical conditions. Magnetic surveys offer a promising pathway for identifying these wells by detecting the magnetic anomalies associated with steel casings. However, magnetometer data are typically high-volume, noisy, and complex, making them difficult to process efficiently with conventional methods. Existing processing methods require heavy preprocessing and achieve unsatisfactory recall scores. In this study, we propose a transformer-based deep learning framework designed to efficiently process hyper-resolute data without extensive downsampling. This is achieved through novel on-the-fly techniques as well as the use of sinusoidal positional encoders to allow the model relative positional awareness. Tests on purely synthetic data show that our model achieves F1-scores of over 90% for line spacings between successive flight paths up to 140 m, enabling surveys to take much sparser flight paths, resulting in more efficient coverage. When applied to real-life data, our model achieves a recall of 70%. This flexible and scalable framework enables the detection of orphan wells from drone data and can be readily adapted to other remote sensing applications.