Drone LiDAR and AI map Chachapoya structures beneath Peru’s forest canopy

In the cloud forests of north-western Peru, an old stone wall can be concealed by vegetation, soil and the passage of centuries. From ground level its outline may be invisible; from a conventional aerial photograph it can disappear into a solid-looking tree canopy. A drone equipped with a laser scanner can produce a different picture, sampling the geometry beneath the leaves. Researchers working around the Chachapoya sites of Kuélap and Cambolín asked how much machine learning could accelerate the painstaking interpretation that follows.[1]

The study compared 17 ways of displaying laser-derived terrain with a Mask R-CNN system that outlined architectural features. At Kuélap it detected 137 of 185 reference structures; at Cambolín, 53 of 73. These were mapped archaeological sites, and the exercise measured agreement with specialists’ interpretation of terrain imagery—not the discovery of a previously unknown city.[1]

How to look through a canopy

The team used drone-borne airborne laser scanning to construct digital terrain models at a half-metre resolution. Rather than interpreting a single height map, they applied seventeen visualisation techniques that emphasised relief in different ways—slopes, shadows, openness and more elaborate combinations. They then trained Mask R-CNN models, a type of neural network that can recognise and outline regions corresponding to target objects. A stone ring or terrace became a candidate pattern to be compared with expert-delineated reference features.[1]

Different lighting and terrain treatments revealed different details. In a steep landscape, a visualisation blending complementary features may make an enclosure stand out. In comparatively smoother terrain, slopes may be more informative. There is no universal filter that turns every landscape into an archaeological inventory. The study trained separately for the two sites and across the tested visualisation methods, reflecting those local differences.[1]

The overall results were concrete but moderate. At Kuélap, one leading detection setup recognised 137 of 185 reference features; at Cambolín, 53 of 73. The reported performance also depended on the chosen metric and visualisation. For example, the leading F1 scores described in the results were about 0.72 for one Kuélap terrain treatment and 0.77 for a slope treatment at Cambolín. These values should not be conflated with the number of verified ancient buildings found in the field.[1]

A detector's disagreement is informative

In object detection, precision asks how often a proposed location matches the reference; recall asks how many reference structures were recovered. A model can achieve apparently strong precision by identifying only the most conspicuous enclosures and ignoring half the smaller walls. Conversely, one that outlines every bump may retrieve more structures at the cost of a map full of false leads. Archaeologists need to choose the trade-off according to their purpose: a survey intended to guide a field team might tolerate more candidates than a map presented as a definitive site plan.[1]

The paper's validation labels were generated through expert visual interpretation and manual digitisation of laser-derived terrain maps. The authors expressly note that independent field verification of these reference data was not performed. That caveat limits a seductive conclusion: the model was tested against what specialists interpreted from the same type of remote evidence, not against an independently excavated inventory of everything buried beneath the trees.[1]

A better map for planning the fieldwork

The Chachapoya archaeological landscape extends far beyond any one iconic fortress. Prospecting large forested areas by foot is demanding and expensive; a drone survey offers a way to narrow a field team's priorities. The researchers describe reduced manual interpretation time in their workflow, but the savings depend on the quality of inputs, site type and level of verification needed afterward. Speed in making candidate polygons is not the same as completing an archaeological investigation.[1]

The site difference is instructive. Recovering a similar proportion of reference structures did not produce equal average-precision scores, because detection quality also depends on false candidates, confidence rankings and how closely the predicted outlines match the reference polygons. A count of recovered structures describes coverage; it cannot substitute for the full quality assessment.[1]

For an archaeological team, the value is a more focused survey. Terrain processing can reveal faint relief over a wide area, and a model can mark candidates for closer attention. The map then becomes a working hypothesis: a way to organise the next walk through the forest, compare uncertain outlines and decide where direct investigation will add the most knowledge.

Report an error or suggest a correction