Abandoned oil wells pose a significant environment and health risk, as they can leak methane — a potent greenhouse gas — into the atmosphere and contaminate groundwater. The total estimated number of undocumented orphaned wells (wells that are not formally owned or documented) is between 310,000 and 800,000. These wells are often “out of sight and out of mind” which makes them a big hazard that’s almost impossible to track.

But according to a new study, AI may be of help. A new, state-of-the-art deep learning framework was trained to recognize oil and gas well symbols on historical topographic maps spanning decades. Using this approach, researchers highlighted over 1,000 undocumented wells and subsequently confirmed their presence to within 10 meters.

Zombie wells

For decades, oil and gas wells have dotted the American landscape — after all, the US was one of the first countries to have a large-scale oil industry. But the less visible legacy of this industry is an environmental disaster. These wells are silent polluters that no one watches, environmental ticking bombs. Methane emissions from abandoned wells (also called ‘zombie wells’) contribute significantly to climate change and additionally, these wells can leak brine and hydrocarbons, threatening aquifers and ecosystems.

Plugging these abandoned wells would cost billions of dollars. But before plugging them, you need to find them.

At some point, we knew where these wells were; they were marked on some maps. But orphaned wells were lost in the fray, and not mentioned in any database. They simply exist as dots on old maps — but there are a lot of old maps. Since 2011, the United States Geological Survey (USGS) has uploaded 190,000 scans of historical USGS topographic maps made between 1884 and 2006.

“This problem is equivalent to finding a needle in a haystack, since we are trying to find a few unknown wells that are scattered in the midst of many more documented wells,” said Charuleka Varadharajan, a scientist at Berkeley Lab and senior author of the study.

The AI developed in this study identifies existing wells marked on historical topographic maps, and then determines which of these wells are undocumented in official databases, classifying them as potential undocumented orphaned wells (UOWs). The process involves two steps:

1. Well Detection on Maps: The AI scans georeferenced historical maps, using a trained neural network to recognize well symbols. These symbols represent wells marked by cartographers on maps created decades ago.
2. Comparison with Databases: The detected wells are cross-referenced with state and federal databases of documented wells. If a well detected by the AI is more than 100 meters away from any documented well, it is flagged as a potential UOW.

Training on the U.S. Department of Energy’s NERSC supercomputer took approximately two hours, highlighting the efficiency of the approach​​.

Confirming the AI

To validate findings, the researchers employed modern satellite imagery and on-the-ground surveys. Magnetic field detectors identified the presence of buried metal pipes, confirming the existence of wells in previously undocumented locations.

The study focused on four counties: Kern and Los Angeles in California, and Osage and Oklahoma counties in Oklahoma. These areas were selected for their historical significance in early oil production. Overall, the team uncovered 1,301 potential undocumented orphaned wells:

• California: The AI identified 539 potential UOWs in Kern and Los Angeles counties. However, the dense urban landscape of Los Angeles posed challenges, with false positives arising from features like roundabouts and cul-de-sacs.
• Oklahoma: The rural setting of Osage and Oklahoma counties yielded higher detection accuracy. Here, 762 potential UOWs were identified, with 29 confirmed using satellite imagery and six through field surveys.

Overall, however, the method was successful, and the researchers are confident the method could be used to find more orphaned wells.

“With our method, we were conservative about what would be considered as a potential undocumented orphaned well,” Varadharajan said. “We intentionally chose to have more false negatives than false positives, since we wanted to be careful about the individual well locations identified through our approach. We think that the number of potential wells we’ve found is an underestimate, and we might find more wells with more refinement of our methods.”

A game-changer

If the method really works, it can be a paradigm shift in addressing the orphaned well crisis. By automating detection over large areas, the AI framework significantly reduces the time and cost of locating undocumented wells. Moreover, the approach is scalable, capable of analyzing maps nationwide.

Regulatory agencies can use this technology to prioritize high-risk wells for remediation. Identifying methane-leaking wells, for example, has long been considered one of the low-hanging fruits for addressing our climate crisis. Furthermore, the framework provides a template for integrating historical data with modern AI tools, opening avenues for similar applications in other fields, such as archaeology or urban planning.

The next step now is to assess how much methane is leaking. High-tech methane sensors are expensive, so researchers are working on low-cost sensors to quantify the leak. Berkeley Lab scientist Sebastien Biraud is spearheading this effort.

“We don’t need to know if it’s leaking exactly 2.3 grams per hour,” Biraud said. “We need to know if it’s not leaking, if it’s leaking between 10 and 100 grams per hour, or if it’s leaking kilograms per hour. And we need to be able to do it in five minutes.” From there, after finding and quantifying the leaks, we can finally move on to prioritizing the biggest culprits and start plugging them.

“There’s a requirement now to quantify emissions before and after plugging an oil and gas well,” Biraud said. “Both because you want to make sure the plugging is done right, and you also want to quantify the impact of the program itself on our climate mitigation strategies — particularly for methane emissions, which can cause global warming impacts more quickly than carbon dioxide.”

Journal Reference: Fabio Ciulla et al, A Deep Learning Based Framework to Identify Undocumented Orphaned Oil and Gas Wells from Historical Maps: A Case Study for California and Oklahoma, Environmental Science & Technology (2024). DOI: 10.1021/acs.est.4c04413