Twelve circular shapes turned up in a lidar scan of Kisatchie National Forest, in the pine hills of north-central Louisiana. Each was a low ring of raised earth wrapped around a shallow depression, with a small pit set off to one side. They sat close to Camp Livingston, an Army installation from the Second World War. And at a glance, they looked like tar kilns.
Tar kilns are how people pulled tar, pitch, and resin out of pine before the petrochemical era. From the eighteenth century into the early twentieth, workers stacked pine staves inside a circular earthen mound, packed soil over them from a surrounding trench, and burned the wood in a slow, low-oxygen fire. The tar drained through a buried pipe into a collection pit nearby. What survives in the ground today is usually a ring-shaped ditch around a slightly raised interior, sometimes with that collection pit still legible just outside the ring.
The Louisiana structures had the circle and the adjacent pit. The relief was backwards, though. A standard southeastern tar kiln reads as a raised center sitting inside a ditch. These read as a sunken center inside a raised berm. Same footprint, opposite topography. That inversion is a small detail with an annoying consequence, because it breaks the usual way archaeologists would go looking for more of them.
The usual way, increasingly, is a neural network. You show a model many labeled examples of a shape and it learns to flag that shape in new terrain. The approach has found Maya buildings, Dutch burial mounds, Samoan agricultural terraces, New England charcoal hearths. The catch never changes: you need the examples first, hundreds or thousands of them, each one traced by hand. For a common feature that is merely tedious. For a feature you have seen exactly twelve times, it is impossible. Twelve examples will teach a model to recognize those twelve and nothing else, a failure with a name, overfitting. And you cannot borrow a model trained on ordinary tar kilns, because these are not shaped like ordinary tar kilns.
So the team behind the study,1 Katherine Peck, Claudine Gravel-Miguel, Grant Snitker, and Matthew Helmer, tried the obvious inversion of the problem. If you do not have enough real examples, invent some.
