The northwest of the Iberian Peninsula has one of the largest Roman gold mining complexes in the world, mainly exploited between the 1st and 3rd centuries AD. During that period, more than 6.5 tons of the precious metal were extracted, according to Latin authors such as Pliny the Elder, who described the situation in what is now Galicia, Asturias, and Leon.
In this last province, Las Médulas is located, one of the best-preserved Roman mining remnants in Europe. This ancient gold mine, declared a World Heritage Site in 1997 by UNESCO, has a hydraulic system consisting of a network of channels excavated in rock that exceeds 700 km in length. The water that flowed through them was used to collapse the mountain and wash the gold-bearing materials.
In the study of this 2000-year-old mining complex, new technologies had already been used, such as drones and airborne LiDAR sensors, but now researchers from the University of Leon have incorporated artificial intelligence (AI) to recognize and map ancient remains preserved in the landscape.
Machine learning algorithms and drone images are combined to identify Roman mining sites and their elements, such as hydraulic channels. The work, published in the journal “Applied Intelligence,” “combines machine learning algorithms (deep learning) and georeferenced drone images for the identification of mining sites and other elements of the Roman system, such as hydraulic channels,” says SINC’s lead author, Daniel Fernández Alonso.
“For this,” he continues, “we have trained this intelligent system based on images with similar geometric patterns, which could easily be confused with mining remains (such as roads, paths, and other elements that make up the landscape), adjusting the system to achieve a 95% accuracy in identifying the different structures of the ancient gold mines.”
Within the field of deep learning, which allows for automatic recognition without human subjectivity, the authors have used convolutional neural networks, capable of learning to enhance the features that best fit the elements defined during training. For example, differentiating a path from a channel like the one in Peña Aguda, a structure almost 43 km long inside Las Médulas.
“This type of neural network continues to refine a set of filters that, when applied to the images, highlight those parts with the elements we are trying to find and differentiate, in our case, mining remains and road intersections,” says co-author María Teresa García Ordás, “and it also has a layer of connected neurons that correctly classify those images generated after applying the filters.”
The researchers point out that mountainous areas like the ones studied are, in many cases, difficult to access, with a landscape transformed over the centuries by increased vegetation and human activity. This makes it difficult to identify ancient mining infrastructures like the Roman ones, but the new methodology offers a useful tool to solve this and help archaeologists, according to the authors.
Another author, Javier Fernández Lozano, also points out that this is the first time that AI has been used to identify channels and ponds. Additionally, “this is a modeling that in the future will help in locating more elements of Roman gold mining and even discovering new deposits or gold mines.”
In the future, this method could be implemented and include other images to recognize patterns of new gold deposits. “For this,” he clarifies, “our method would have to be implemented and include other types of images, such as those taken with multi and hyperspectral cameras, which allow for the recognition of characteristic patterns of gold deposits that can then be compared based on a prospective study.”
He also highlights that this method, “for which we have already applied for a patent,” allows for the identification of remnants of more modern mining present in the landscape. “The province of Leon – he gives as an example – has had, in addition to gold, a long history of coal, iron, or tungsten mining, among others. Being able to recognize the traces they left on the territory facilitates proper management by the authorities, thus avoiding accidents and costs associated with the abandonment of wells and other mining infrastructures.”
In fact, the researchers conclude: “This innovative application of deep learning can be implemented to reduce the potential risks caused by abandoned mines (especially those that are not marked), which can cause significant annual human and economic losses worldwide.”
Source: MiMub in Spanish
