How a candidate emerges

Begin with a narrow task

“Find archaeology” is not a sufficiently defined machine-learning task. A useful project might instead ask for possible mound-like forms in a particular terrain model, or outlines resembling a known class of surface feature in aerial imagery. The narrower definition makes errors interpretable. It also makes clear what the model was never designed to detect: a detector trained on raised earthworks may say little about buried occupation without surviving relief.

Different tasks produce different outputs. Classification assigns a category to an image or patch. Object detection proposes locations and bounding boxes. Segmentation assigns classes to pixels. Anomaly detection highlights departures from an expected pattern. None of these operations supplies chronology, cultural affiliation or a complete archaeological explanation. Those conclusions require evidence outside the model output.

The training record carries old decisions

Training labels often come from existing inventories and expert annotation. Those records reflect earlier survey coverage, access and research priorities. A well-surveyed agricultural region may contribute many examples, while woodland, politically restricted areas or poorly funded regions contribute few. The model can reproduce that imbalance even when the code treats every image identically. Missing entries should not automatically become reliable examples of “no archaeology”.

Annotators also disagree. One specialist may draw a tight boundary around a visible bank; another may include the surrounding enclosure. Both choices alter what the system learns. Good documentation records how labels were defined, how uncertain examples were handled and whether specialists checked one another’s decisions. Keeping a versioned label set allows later researchers to distinguish improved interpretation from a change in the underlying evidence.

Negative examples deserve equal attention. Modern field boundaries, extraction pits, natural depressions and recent tracks can help reveal which confusions matter locally. A training collection filled only with obvious monuments and empty backgrounds is unlikely to prepare a model for the ambiguous landscape where it will actually be used.

Test somewhere meaningfully different

Randomly dividing neighbouring image tiles between training and testing can produce an overly encouraging result. Adjacent tiles may share the same site, flight conditions, soil or processing artefacts. A model can appear to generalise while recognising the signature of a survey it has effectively seen before. Geographic separation, and where possible separation by survey campaign, offers a more demanding test.

Precision asks how many reported candidates are correct under the evaluation definition. Recall asks how many reference examples were recovered. These are different questions. A low reporting threshold may find more known features but send many false positives to reviewers. A high threshold may reduce review time while excluding faint or unfamiliar examples. Neither trade-off can be judged without knowing the intended use and available verification capacity.

A score is not automatically a probability that archaeology exists. Calibration must be assessed before scores receive that interpretation. Even a calibrated score within one test population may behave differently after a shift in vegetation, sensor resolution or regional monument form. Reports should describe those limits in plain language rather than hiding them behind a single aggregate metric.

Read a real example carefully

The 2024 PNAS study by Masato Sakai and colleagues reported 303 newly identified figurative geoglyphs in the Nazca region through an AI-assisted research programme. The important phrase is “AI-assisted”. Candidate generation was part of a wider process involving archaeological review and field investigation. The study is evidence that computation can improve a targeted search workflow; it is not evidence that a general model can discover any kind of archaeological site without specialists.

When reading a headline about this work, separate the published discoveries from the full set of model suggestions and from the research team’s interpretations. Ask what imagery was used, how candidates were prioritised and how ground investigation contributed to the final record. Those questions make the example transferable without pretending that its performance can simply be copied into a different landscape.

Make the shortlist useful to people

A responsible candidate record includes the source image, processing history, model version, score, review notes and reasons for uncertainty. It should also record rejected explanations. If a reviewer identifies a modern ditch, that correction is valuable evidence about the system’s behaviour rather than an embarrassing item to discard. Repeated errors can guide new training examples or reveal that the chosen task is poorly defined.

Reproducibility does not require unrestricted publication of sensitive coordinates. Researchers can share methods, aggregated results and appropriately controlled data. The FAIR principles encourage findable, accessible, interoperable and reusable research outputs; accessible does not necessarily mean open to everyone without conditions. A model should arrive at field verification with a documented history and an honest question: what evidence would support or overturn this suggestion?