Unveiling the Past: AI Identifies Hidden Oil and Gas Wells Across the U.S.

By Dr Silpaja Chandrasekar, PhDReviewed by Joel ScanlonDec 18 2024

Study reveals how advanced AI and satellite imagery are transforming the detection of undocumented oil wells, helping to mitigate environmental hazards and refine energy practices.

Research: A Deep Learning Based Framework to Identify Undocumented Orphaned Oil and Gas Wells from Historical Maps: A Case Study for California and Oklahoma

In a research paper published in the journal Environmental Science & Technology, researchers developed a framework using a U-Net model to identify unmapped oil and gas wells (UOWs) from historical topographic maps.

They identified potential UOWs as symbols located over 100 meters from documented wells and validated 44 sites using satellite images and field surveys. This validation included an innovative custom script that rapidly confirmed well symbols by isolating and displaying areas for manual inspection, significantly improving the efficiency of the process.

Applied across California and Oklahoma, this method discovered 1,301 potential UOWs with high spatial accuracy. The results demonstrated the feasibility of applying the workflow at a regional scale, with potential to scale up for nationwide application.

The framework was designed to scale nationwide, aiding in addressing environmental risks. Challenges to scalability, such as confounding features in urban areas and variations in map quality, were also highlighted as areas for further improvement.

Related Work

Past work highlighted challenges in identifying UOWs across diverse United States (U.S.) landscapes due to varying documentation and practices.

Traditional methods relied on ground-based surveys, remote sensing, or manual data mining but needed more scalability for regional studies.

Challenges in identifying UOWs included inconsistent documentation across U.S. states and the need for scalable methods for regional detection.

Traditional techniques were limited to small-scale studies, hindering nationwide application.

Mapping Uncharted Oil Wells

The study utilized georeferenced raster images from the Historical Topographic Map Collection (HTMC), covering the contiguous U.S., Hawaii, and parts of Alaska, to identify UOWs. These maps, spanning 1947 to 1992 and issued at a scale of 1:24,000, provided consistent symbology to distinguish natural and artificial features, including oil and gas wells marked as hollow black circles.

To validate the workflow, the researchers focused on California and Oklahoma, two states with significant early oil production history. Kern and Los Angeles counties in California and Osage and Oklahoma counties in Oklahoma were selected based on their historical production records and diverse land use patterns.

The data preparation involved selecting 79 maps from California and manually labeling 11,046 well symbols using Labelme software, resulting in 440 tiles. These tiles were further processed into 7,040 image-mask pairs, augmented tenfold to create 70,400 samples for training a U-Net neural network.

The segmentation model was trained and validated to detect well symbols, achieving a precision of 0.99 and recall of 0.88. This trade-off was deliberately chosen to prioritize precision and minimize false positives, reflecting the need to reduce errors in identifying potential UOWs. Researchers did not use maps from Oklahoma during training to test the model's generalizability.

The trained model was applied to detect wells on maps by slicing them into smaller tiles and generating binary masks from probabilistic outputs. Detected wells were matched with documented well databases from state agencies, and those located over 100 meters away from documented wells were flagged as potential UOWs.

To address potential underestimation, the researchers acknowledged limitations such as temporal gaps between map issues, map distortions, and merging duplicates across overlapping maps. These factors contributed to challenges in identifying all potential UOWs.

Manual verification using a custom script confirmed the presence of hollow black circle symbols for potential UOWs. Detections from overlapping maps were merged to eliminate redundancy, producing a final list of vetted potential UOW locations.

To further verify potential UOWs, modern satellite imagery and historical photographs were reviewed for well-related structures, such as oil rigs or storage tanks. For instance, historical aerial photos were prioritized in California’s Kern County, helping flag sites with evidence of past oil extraction activities despite the absence of visible surface structures in modern imagery.

The workflow demonstrated the feasibility of combining historical maps, deep learning (DL), and satellite imagery to identify UOWs effectively, addressing a critical gap in oil and gas well documentation and enabling further environmental and regulatory assessments.

Unmapped Wells Identification

The study leveraged a computer vision model trained on georeferenced historical topographic maps to identify UOWs across California and Oklahoma.

Tested on 14,080 unseen data points, the model achieved a precision of 0.98 and a recall of 0.88. Despite this high algorithmic precision, the ratio of vetted-to-unvetted UOWs varied significantly by region, highlighting the impact of map quality and confounding features. For example, urban areas like Los Angeles County showed lower performance due to symbols resembling wells, such as cul-de-sacs and roundabouts.

The workflow identified 1,301 potential UOWs—539 in California and 762 in Oklahoma—with detailed geospatial data in the supplementary material. However, potential underestimation was noted due to documented well-location errors and the absence of aboveground structures at many sites.

Satellite imagery and historical aerial photos were used to verify detected UOWs. In Oklahoma, visual inspection confirmed well-related structures for 29 of 261 potential UOW sites in Osage County, with detected well coordinates showing an average deviation of 9.4 meters from satellite-based estimates.

California posed more significant challenges as many wells lacked surface features due to plugging and abandonment practices. Aerial photos were used to prioritize 25 of 50 sites in Kern County for further investigation based on evidence of historical production activities, though low-resolution images limited confirmation accuracy.

Field campaigns conducted in Kern County and Osage County utilized magnetometer surveys to detect buried wells. In Kern County, 9 out of 13 accessible sites showed magnetic anomalies compatible with UOWs, while in Osage County, 6 out of 14 sites were confirmed as wells.

Across all field investigations, 15 of 27 potential UOW sites were verified, with an average deviation of 11.7 meters from algorithm-detected coordinates.

The successful transfer of a model trained on California data to Oklahoma highlights the workflow's generalizability, showing promise for scaling to new regions.

Conclusion

To sum up, the study developed a DL-based workflow using historical topographic maps to identify potential UOWs across California and Oklahoma.

The model achieved high precision, detecting 1,301 potential UOWs, though underestimates were noted due to mapping limitations and well-location errors. The workflow combined satellite imagery, aerial photos, and field surveys, confirming the accuracy of well locations within 10 meters.

Despite challenges like map distortions and confounding features, the method demonstrated scalability and transferability, successfully identifying UOWs in different regions. Further improvements, such as expanding training data and refining image segmentation models, could enhance performance, particularly in urban settings.