Researchers at Yamagata University have identified 303 new figurative geoglyphs on Peru’s Nazca desert, nearly doubling the known catalog of these ancient ground drawings. Among the newly confirmed figures is a killer whale shown gripping a knife, an image that connects to a well-documented supernatural motif found on Nasca ceramics and rock carvings. The discovery, announced in September 2024, was driven by a deep-learning system that, according to the university, made the survey process sixteen times more efficient than traditional fieldwork alone.
How deep learning reshaped the Nazca survey timeline
The speed of this discovery matters because it compresses decades of expected fieldwork into a few seasons. According to Yamagata University’s September 2024 press release, the AI system flagged 1,309 promising geoglyph candidates from high-resolution aerial imagery. Roughly one quarter of those candidates survived ground-truth verification, yielding the 303 confirmed figures. That ratio points to a practical screening tool rather than a finished detector: three out of four flagged sites turned out to be false positives or ambiguous marks once archaeologists walked the ground.
The 303 figure also sits in a longer arc of accelerating discovery. In 2019, the same research group reported 143 new geoglyphs on the Nasca Pampa and surrounding area, a count that itself represented years of combined aerial and pedestrian survey. Moving from 143 confirmed figures to 303 in roughly five years suggests the deep-learning pipeline is scaling faster than the team’s earlier methods allowed. Whether that pace can be sustained, or whether the most detectable figures have already been found, is an open question the published data does not yet answer.
One testable implication is whether the same models, retrained on an expanded motif catalog, could locate additional figurative geoglyphs in the adjacent Palpa valleys using drone surveys already flown between 2022 and 2023. If the system’s detection logic generalizes beyond the Nazca Pampa, dozens more figures could surface without new flights. If it does not generalize, that would reveal how site-specific the training data remains and how much manual retraining each new survey zone requires.
The killer whale motif and what the peer-reviewed record shows
The killer whale gripping a knife is not a random image. Nasca iconography features a recurring supernatural creature, part orca and part human, often depicted holding a trophy head or a blade. The Art Institute of Chicago has published a detailed scholarly treatment of this figure across ceramics, petroglyphs, and geoglyphs in a digital study of mythical killer-whale imagery. That publication traces the motif’s evolution through multiple phases of Nasca culture, showing how the creature’s attributes shifted over centuries and how its fanged mouth, knife, and dismembered heads signaled power over life and death.
The newly detected geoglyph appears to fit within this iconographic tradition, echoing the curved dorsal fin, prominent teeth, and weapon-bearing posture that scholars have linked to ritual violence and water symbolism. Yet the match remains stylistic rather than stratigraphic. No published excavation or ceramic association has yet tied this specific ground figure to a dated phase of the sequence outlined in the Art Institute’s analysis, leaving its precise position within Nasca chronology unresolved.
The technical workflow behind the discovery was first detailed in a peer-reviewed paper in the Journal of Archaeological Science, which described the deep-learning object-detection approach and its precision–recall tradeoffs. That earlier study served as a feasibility demonstration, training convolutional networks on labeled geoglyph imagery and then testing them on unseen tiles. The larger-scale deployment, covering the full Nazca survey area and producing the 303-figure count, was subsequently published in PNAS, according to the university’s press materials. Together, the two papers establish a method lineage: a controlled pilot followed by a scaled application, each subjected to peer review before the headline numbers were released.
Gaps in the data and what to watch next
Several pieces of evidence that would allow independent scrutiny remain unavailable. The full list of 1,309 AI-flagged coordinates and their confidence scores has not been deposited in a public repository. Field-verification logs, including dates, team composition, and ground-truth photographs for each of the 303 confirmed figures, are likewise unpublished. Without these records, outside researchers cannot replicate the claimed sixteen-fold efficiency gain or evaluate how borderline cases were classified.
The two headline totals from Yamagata University also require careful reading. The 2019 release reported 143 new geoglyphs; the 2024 release reports 303. These are not simply additive. The earlier count covered a different survey period and methodology mix, and the university’s own framing treats the 303 as a distinct result tied to the scaled AI pipeline. Readers should treat the two numbers as separate campaign outputs rather than a running tally, because the overlap between the two datasets has not been publicly clarified. Some of the earlier figures may have been re-detected by the new system, while others could have been excluded from the AI-focused reporting to keep the metrics clean.
The practical next step for the field is whether the raw high-resolution imagery tiles used in the PNAS-scale analysis will be released for independent replication. Open access to those tiles would let other teams apply alternative detection models, experiment with different training–validation splits, and test how robust the reported efficiency gains really are. It would also allow researchers to probe potential biases: for example, whether high-contrast animal figures are overrepresented among detections compared with more abstract or eroded designs that might be harder for algorithms to recognize.
There are also cultural and ethical dimensions to consider. Many of the newly mapped geoglyphs lie outside the most heavily visited sections of the Nazca Pampa, raising questions about how publicizing their locations might affect looting risks, tourism pressure, and local stewardship. Institutions that help interpret Nasca visual culture for the public, such as the Art Institute of Chicago, routinely navigate similar tensions when they present sensitive ritual imagery in galleries or digital formats. Visitors who want to situate the killer-whale motif within a broader Andean art context can explore on-site collections through standard museum admission or engage more deeply via member programs promoted through the institute’s membership offerings, which often support research and conservation.
For archaeologists, the Nazca case is becoming a benchmark for how AI can be integrated into long-term landscape surveys without displacing traditional field skills. The Yamagata team still relies on pedestrian verification, stratigraphic observation where possible, and iconographic comparison with ceramics and rock art. The deep-learning models function as triage tools, not oracles: they prioritize where to walk and what to photograph, but they do not decide what counts as a geoglyph in the final catalog. As more datasets are released and additional teams test comparable pipelines in other regions, the Nazca experience will likely shape emerging norms around transparency, reproducibility, and the balance between computational speed and interpretive care.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
