Quantitative evaluation of the smectite content of samples with different origins using deep learning with hyperspectral data in the visible to near-infrared range
This study demonstrates that deep learning models, specifically convolutional neural networks, can accurately classify smectite content ranges in hyperspectral VNIR data by leveraging spectral features in the 760–950 nm range, with the findings indicating that geological origin exerts a greater influence on spectral signatures than lithology, thereby offering significant potential for applications in geotechnical engineering, resource exploration, and waste storage site selection.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Deep underground, where tunnels are carved through mountains, the stability of the rock itself is a matter of life and death for the structures built within it. Some rocks contain a specific type of clay mineral called smectite. When this mineral is present in significant amounts, it acts like a sponge that swells when it gets wet, exerting immense pressure that can crack tunnel walls, deform roadways, or even trigger landslides. For decades, engineers have relied on laboratory tests to measure how much of this swelling clay exists in a rock sample. These traditional methods are accurate but slow, often taking days to return results. In the fast-paced world of construction, waiting for a lab report can mean delaying critical safety decisions or missing the window to reinforce a tunnel before it collapses. The question has long been whether a faster way exists—one that could tell an engineer the danger level of the rock immediately, right at the edge of the excavation site, without needing a degree in geology or a trip to a laboratory.
A team of researchers from Shimizu Corporation and Hokkaido University has taken a significant step toward answering this question by combining advanced imaging with artificial intelligence. They focused on a range of light that the human eye cannot see, known as the visible to near-infrared spectrum. While our eyes see only a narrow band of colors, hyperspectral cameras can capture hundreds of subtle shades across a much wider range, revealing the unique chemical fingerprints of minerals. The researchers gathered rock samples from two very different geological settings in Japan: one set of volcanic rocks from a mountain tunnel in Kyushu that had been altered by heat and steam, and another set of sedimentary rocks from a tunnel in Hokkaido that had changed slowly over millions of years through natural burial and pressure. They ground these rocks into powder and measured their exact smectite content using standard X-ray diffraction, a precise but time-consuming technique. Then, they took pictures of the same powders using a hyperspectral camera, capturing the way light bounced off the minerals across the spectrum.
The team then trained a computer program, specifically a type of artificial intelligence called a convolutional neural network, to learn the relationship between the light patterns and the amount of smectite present. The goal was to see if the computer could look at a new rock sample, analyze its light signature, and accurately predict how much swelling clay it contained. The results were promising but revealed a crucial nuance. When the computer was tested on rocks that came from the same geological family as the ones it learned from, it performed with high accuracy, correctly identifying the smectite content in nearly all cases. However, when the researchers asked the same computer to analyze rocks from a completely different geological origin, its performance dropped significantly. The light patterns that signaled a high amount of smectite in the volcanic rocks were different from those in the sedimentary rocks. The study suggests that the history of how a rock formed—its origin—has a stronger influence on its light signature than the type of rock it is. The artificial intelligence learned to spot the smectite by focusing on a specific increase in light reflection between 760 and 950 nanometers, a range where the presence of this clay causes the rock to appear brighter.
To prove this method could work in the real world, the researchers built a portable darkroom that could be set up directly at a construction site. They demonstrated that they could take rock samples, grind them, and capture the necessary light data right there, without needing to send anything to a distant lab. The study concludes that while this technology cannot yet be a universal tool for every rock type on Earth, it offers a powerful new capability for specific geological contexts. If engineers know the origin of the rocks they are working with, they can use this deep learning system to instantly map out where dangerous levels of smectite exist. This could allow for immediate decisions on whether to reinforce a tunnel or change its design, preventing deformation and ensuring safety long before a problem becomes visible. The research indicates that by understanding the unique light fingerprint of a rock's origin, we can turn a slow, complex analysis into a rapid, on-site assessment, transforming how we manage the hidden risks beneath our feet.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.