Application of XGBoost Extreme Gradient Boosting Model in the Evaluation of Shallow Gas Hazards in Urban Strata—A Case Study of a Typical Area in LG City

Beneath the streets of many modern cities lies a hidden layer of natural gas, trapped in the loose, water-saturated soil just below the surface. This is not the deep, high-pressure gas used to power homes, but a shallow, often biologically generated mixture of methane that accumulates in pockets within the sand and mud. Because it sits so close to the surface, this gas can build up significant pressure, creating a silent threat to construction projects, subway tunnels, and even the foundations of buildings. When workers drill into the ground or excavate for a new basement, they risk piercing these pressurized pockets, which can lead to sudden ground collapses, gas explosions, or the ejection of soil and water. For city planners and engineers, the challenge has long been knowing exactly where these dangerous pockets exist before they start digging. Traditionally, predicting these hazards has relied heavily on the experience of geologists and the interpretation of local rock layers, a method that works well in some places but struggles to capture the complex, shifting patterns of gas accumulation in dense urban environments.

To tackle this uncertainty, a team of researchers turned to a different approach, using a powerful computer learning tool to map the hidden risks in Longgang, a coastal city in Zhejiang Province, China. Instead of relying solely on human intuition, the team fed a massive amount of geological data into an algorithm known as XGBoost, a type of machine learning model designed to find subtle patterns in complex datasets. The researchers gathered information on eight specific conditions that influence whether gas will accumulate and cause trouble: how deep the gas lies, how much pressure it exerts, the number of gas-rich layers, the amount of organic carbon in the soil, the water content, the porosity of the ground, the gas saturation levels, and the overall pressure of the formation. They treated the city like a giant grid, dividing it into small squares and labeling each one based on whether it was a known site of ground collapse or a safe area. By training the computer to recognize the combination of factors present in the dangerous spots versus the safe ones, the model learned to predict where similar hazards might exist elsewhere.

The results of this digital training were striking. The computer model, once fine-tuned with a sophisticated method to find the perfect settings for its calculations, proved highly accurate at distinguishing between safe ground and hazardous zones. When tested against real-world data, the model achieved a high score that indicates it can reliably separate dangerous areas from safe ones, far better than random guessing. The analysis revealed that the most critical factor in predicting a disaster was simply the number of gas-bearing layers present in the soil; the more layers of gas-rich soil stacked together, the higher the risk. Following closely behind were the saturation of the gas and the porosity of the ground. The model also clarified how certain factors work against danger, showing that higher organic carbon content and specific pressure conditions actually help to suppress the risk of an eruption.

Using these insights, the researchers generated a detailed map of Longgang that color-coded the city according to its risk level. The map showed that while the majority of the city falls into a moderate risk category, specific neighborhoods, particularly the Dongpai and Xipai communities, sit in zones of high susceptibility. This finding aligns closely with historical records, as nearly all the ground collapses that have already occurred in the area happened within the zones the model identified as high-risk. The study demonstrates that data-driven tools can successfully decode the complex geological signals of shallow gas, offering a more precise way to identify danger zones than traditional methods alone. While the current model relies on static geological data and does not yet account for dynamic human activities or chain reactions that might trigger a collapse, it provides a solid foundation for future safety planning. By pinpointing where the ground is most likely to fail, this approach helps cities like Longgang prepare for the invisible forces beneath their feet, ensuring that future construction can proceed with a clearer understanding of the risks below.