Anomaly Detection for Locating Abandoned Oil and Gas Wells Using ResNet models

Abandoned oil and gas wells present serious environmental hazards, including uncontrolled greenhouse gas emissions and contamination of soil and groundwater. Detecting and accurately locating these wells, particularly in remote and difficult-to-access regions, is a critical prerequisite for effective remediation. This study proposes a novel anomaly-based framework that integrates high-resolution LiDAR-derived land surface parameters with deep learning to identify abandoned wells in Alaska. LiDAR point clouds were collected using a drone-mounted 3D mapping system equipped with a Light Detection and Ranging sensor, a high-gain GNSS antenna, and a tactical-grade inertial measurement unit. After preprocessing and classification, digital terrain models were used to derive profile curvatures, which highlight geomorphological anomalies indicative of abandoned well sites. These features served as inputs to fine-tune the ResNet50 convolutional neural network. The model achieved an accuracy of 96.2 percent with zero false positives at a confidence threshold of 0.8, outperforming the baseline model trained without data augmentation. Results demonstrate the effectiveness of combining LiDAR-based terrain metrics with deep learning to detect abandoned wells with high precision, even under dense vegetation cover or site degradation. The proposed approach provides a scalable, minimally invasive tool for well detection in remote environments, enabling targeted remediation strategies and reducing methane emissions.