Ancient agriculture leaves behind a hidden record in the form of microscopic silica “skeletons” called phytoliths. These tiny fragments are key to understanding how early civilisations grew food, but identifying them under a microscope is an incredibly slow task, requiring years of specialised training.
While modern Artificial Intelligence (AI) can now automate this process in seconds, a major scientific hurdle remained: these systems often act as a “black box” (see image 1). Without knowing how the computer reaches its conclusions, scientists cannot fully trust if the results are based on real botanical evidence. To understand why this transparency is vital, consider the “cat vs. dog” problem (see image 2): a poorly trained AI might correctly identify pets only because it recognizes the grass in the background of dog photos rather than the animals themselves.
Our research successfully “opened the box” using a visual tool, called Guided Grad-CAM, to map exactly which features the AI focuses on when classifying ancient crops. We discovered that the algorithm independently learned to recognise the same complex wave patterns and surface bumps (papillae) (see Image 3) that human experts use to distinguish between ancient wheat, barley, and oat. In 91% of cases, the AI’s primary decision-maker was the genuine biological wave pattern of the plant.
By making AI “explainable”, we are setting a new gold standard for computational archaeology. This ensures that the digital tools used to reconstruct human history are as reliable.
