AI & Archaeology · c. 5400 BC–AD 220 · 18 sites across China, China
AI learns to identify charred crop seeds up to 7,400 years old from China
A model called APSNet, built on 8,340 microscope images of seeds from 18 Chinese sites in 17 plant types, scored 90.2% accuracy.
Peer-reviewed paper
18 sites across China, China A computer model can sort charred seeds from ancient Chinese sites into 17 plant types, and it gets about nine in ten right. The seeds come from 18 archaeological sites across northern and southern China. They date from around 5400 BC to AD 220.
The researchers call the model APSNet. To train it, they built what they describe as the first large image dataset for classifying ancient plant seeds. It holds 8,340 microscope photographs. The work is published in the journal Heritage Science.
Seeds are a main source of evidence for early farming and diet. Identifying them still depends on expert knowledge. The authors note that charred seeds in particular take years of specialised training before a researcher can work alone. That slows down large studies.
Barley, wheat, millet and peach stones
The dataset covers common ancient crops. Among them are barley, bread wheat, foxtail millet and broomcorn millet. It also includes fruit seeds such as peach. All samples were collected in real archaeological research settings, in close collaboration with professional archaeobotanists.
The seeds were photographed after cleaning and flotation. Flotation means soaking excavated soil in water so that light plant remains float to the surface. The team used an Olympus SC180 digital microscope camera. Every image shows the embryo side of the seed at the same 1.6× magnification.
The authors list several difficulties. Charred seeds are often broken, and the collection spans a wide range of places and periods. So seeds of the same type can look quite different. Charring also changes texture and colour, and closely related species look alike. Some categories are far more common than others, because excavations find what they happen to find.
To cope, the team added seed size as a clue, based on the experience of archaeobotanical experts. Because the magnification is fixed, each image keeps the seed's true size. A part of the model called Size Perception and Embedding feeds size information in early. A second part, Asynchronous Decoupled Decoding, handles spatial cues such as size separately from other image features.
Tested against 28 other models
The team compared APSNet with 28 existing image classification models. These included standard models, models built for telling apart very similar objects, and models designed for uneven category sizes. APSNet scored 90.2% accuracy and outperformed the others, according to the paper. The authors say the comparison sets the first benchmark for ancient seed classification.
An accuracy of 90.2% means roughly one image in ten was put in the wrong category. The authors present the model as an aid for the repetitive first stage of sorting. Experts would still review the results.
They see it as most useful for large excavation projects and regional surveys, where many seeds need sorting. They also argue that a shared dataset and model could reduce differences between individual researchers. That would make results from different sites and periods easier to compare.
Sources
Paper: Heritage Science, 2026, DOI: 10.1038/s40494-026-02736-9
Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.
Readers' picks show on the front page.
Follow @AncientHeraldHQ on X for the day's finds as they come in.
Comments
House rules. Stay on the subject, be kind to other readers and say where your facts come from. No advertising and no personal attacks. The editor reads along and hides comments that break these rules.
Loading comments…