A new study in Heritage applies deep learning directly to lidar point clouds, sorting archaeology from forest and ground without the manual cleanup step that has slowed automated mapping for years. The headline results are not the accuracy scores. They are a training method that closed the last gap in the pipeline, and the finding that a simpler model matched a carefully engineered one in half the training time.
AI Brings Under-Canopy Lidar Archaeology Closer to Automation
Most AI applied to lidar archaeology has worked on pictures. Take the point cloud, render it as a hillshade or local relief model, then hand the image to a convolutional network trained on photographs. It works, and it discards the third dimension along the way. A new study in Heritage by Gabriele Mazzacca and Fabio Remondino of the Bruno Kessler Foundation in Trento keeps the deep learning inside the point cloud, labeling every return as vegetation, ground, or structure across three sites in Italy and Croatia that span a wide range in point density.
Why lidar in archaeology needs AI
Standard ground filters are built to remove everything above the terrain, which is the wrong instinct when the target is a two thousand year old, collapsed wall. What archaeologists want is a Digital Feature Model, a surface stripped of vegetation where walls, roads, and earthworks survive. No conventional filter delivers that cleanly, so someone ends up classifying points by hand, and hand classification does not scale to landscapes. That gap is what AI is being asked to close.
How the model reads a lidar point cloud
The comparison at the center of the study is between two ways of telling a model what a point’s surroundings look like.
Random Forest, the machine learning baseline here, needs that description handed to it. The authors calculated shape measurements in advance for every point, capturing how flat, upright, or scattered its neighborhood is at several radii. The model learns which combinations signal a structure. It trains in under a minute and depends entirely on someone having chosen useful measurements up front.
Point Transformer, the deep learning approach, skips that step. It reads the raw coordinates and works out the spatial relationships itself, weighing each point against its neighbors. In the class that matters, the gap was wide. Structure IoU reached 51.8 to 56.3 percent against 33.5 to 46.0 percent for Random Forest, with far less scattered noise left on the surface after vegetation filtering.
Both models also work from two signal attributes beyond point position. One is intensity, the strength of each returning pulse, normalized across sites because one sensor used a green laser while the others used near-infrared. The other is return order: in a sequence of returns from a single pulse, early returns tend to come off the canopy and later ones off the ground, which gives the model a rough read on vertical position.
Training the model on its own mistakes
One of the most interesting ideas in the paper has nothing to do with architecture. The workflow runs in two stages. Stage one strips vegetation at coarser resolution, discarding about 76 percent of the points and raising the structure share from roughly 3 percent to 12 percent. Stage two separates structure from ground on what’s left over.
Earlier versions trained both stages on clean, annotated data, which creates a mismatch that only shows up in deployment. Stage two learned from perfect input, then received stage one’s imperfect output with missed vegetation still in it. This is where the human re-enters the loop and provides the correction.
Here, stage two trains on stage one’s actual predictions. The model sees those errors during training and learns to work around them. The manual correction step disappears, and the pipeline now runs from start to finish on its own.
The rare class problem
Structures make up about 3 percent of the points. A model labeling everything vegetation or ground would score near 97 percent accurate and find no archaeology at all. The staged design is partly an answer to that, since stage one exists mainly to improve the ratio before the hard decision, and training was weighted so the rare class would not be drowned out. Modern buildings and ancient ones share a single label, because telling them apart is an interpretive judgment rather than a geometric one.
When the simpler AI wins
The authors also ran an end-to-end model, a single pass that assigns all three labels at once with no staging in front of it. Theoretically, it should have performed worse across the board.
The best staged configuration reached 75.08 percent average IoU in 4 hours 10 minutes. End-to-end at 0.5 m posted the highest structure IoU in the study, 51.92 percent, in 2 hours 4 minutes. Pushing it to 0.3 m cost 8 hours 33 minutes and moved structure IoU slightly down, to 51.42 percent, though one site (Punta Kriza) was deliberately run below the resolution its point density supports, so part of the dip may be one sparse dataset dragging the average down.
The staged pipeline still buys higher precision and a checkpoint you can inspect before spending compute on the second stage. The end-to-end method buys recall and runs considerably faster. Which one fits depends on whether you are verifying known archaeology or searching a forest nobody has surveyed yet, and on a first pass across unmapped canopy, recall is usually what you want.
Source: Mazzacca, G. and Remondino, F., “Under-Canopy Archaeological Mapping Using LiDAR Data and AI Methods,” Heritage 9(4), 134, 2026. DOI: 10.3390/heritage9040134.
Figures are reproduced from the source article, published open access under a Creative Commons CC BY license permitting reuse with citation.
