AI for Archaeology: Simulated Lidar Training Data Uncovers Hard-to-Find Site Patterns

The Core Problem: Rare Sites Break Standard ML Workflows

Most deep learning models need hundreds or thousands of labeled examples to perform well. Archaeological site detection is already data-scarce. When you’re looking for a site type that barely exists in the record, standard supervised learning approaches fall apart quickly.

Manual labeling of real sites is also slow. It requires expert time, careful annotation, and often field access. That bottleneck limits how fast researchers can build usable training sets.

The Louisiana study tested whether simulated data could sidestep that bottleneck entirely.

How Simulated Lidar Training Data Works

The team created artificial versions of the target structures and placed them into real lidar-derived digital elevation models. Lidar captures small changes in ground height with high precision, making it useful for mapping features hidden under forest canopy.

The simulation process involved:

• Generating idealized kiln shapes as circular buffers around placed points
• Adding random noise to make the shapes look more natural
• Including collection pit geometry alongside the main structure
• Placing objects only in suitable terrain locations to avoid unrealistic placements

Two simulated-data models were trained using this approach. A third model, trained on real tar kilns from South Carolina with adjusted elevation data, served as a comparison baseline.

All three used Mask R-CNN, a well-established computer vision architecture for instance segmentation. Training used tiled lidar images split 80/10/10 across training, validation, and testing sets.

What the Models Actually Found

The results varied significantly between the two simulated-data models, and the tradeoffs are worth understanding clearly.

Model 1 produced 183 predictions. After filtering, 142 remained for review. It found 9 of the 12 known targets and flagged 6 additional objects worth investigating.

Model 2 was more aggressive. It predicted 2,032 objects. After filtering, 709 remained — and it found all 12 known targets plus 11 additional candidates. The catch: 686 of those remaining predictions were false positives, many triggered by reservoirs, drainage features, and natural mima mounds.

The South Carolina comparison model found 11 of the 12 targets and 7 additional candidates, performing competitively despite being trained on real but geographically mismatched data.

Higher recall came with a significant false positive cost. That’s a real tradeoff, not a flaw to dismiss — it shapes how much post-processing work the human reviewer has to do.

Filtering False Positives: Where Human Judgment Still Matters

The team used ground shape analysis to separate mound-like forms from the pit-like profile of the actual targets. Additional spatial filters removed some of the noise from drainage and reservoir features.

Even after filtering, reviewing hundreds of flagged objects takes time. But the researchers noted that cleaning model outputs was still faster than conducting a pedestrian survey of the same landscape. That’s the practical argument for using these models as a first-pass screening tool before committing field resources.

Field Validation Changed Everything

Here’s where the study gets genuinely interesting — and honest.

Researchers visited 11 of the original 12 targets and tested two with augers. They found none of the physical evidence you’d expect at a tar kiln: no charcoal-rich deposits, no charred pine wood, no buried wooden tar pipes, no hard clay floors.

The structures also sat near a former World War II military training site. Cross-referencing a wartime training manual, the team concluded the circular features were more likely howitzer emplacements than tar kilns.

The AI found the right shapes. The initial interpretation of what those shapes meant was wrong.

That outcome isn’t a failure of the method — it’s a demonstration of what the method is actually for. The model narrowed a large landscape to a manageable set of locations worth investigating. Field work then answered the question the model couldn’t: what are these things?

Why Synthetic Training Data Matters for Geospatial AI

The broader implication here extends well beyond archaeology. Any domain where labeled examples are scarce — rare geological features, specific infrastructure types, low-frequency environmental events — faces the same training data bottleneck.

Simulated data offers a practical path forward when:

• Real examples number in the dozens, not thousands
• Manual annotation is too slow or expensive to scale
• The target feature has a predictable geometric form that can be modeled

The Louisiana study suggests simulated objects can be generated in minutes, compared to the hours required to manually label real sites. That speed advantage compounds when you need to iterate on model design or test different feature geometries. Related challenges in Geospatial AI show how much depends on data quality and constraints.

Practical Takeaway for AI Practitioners

If you’re working on a detection or classification problem with limited labeled data, the lesson here is straightforward: don’t wait for more real examples if you can simulate plausible ones.

The key conditions for this approach to work are:

1. The target has a describable shape — something you can approximate geometrically
2. You have real background data — in this case, actual lidar terrain models
3. You can validate outputs in the real world — simulation handles discovery, not confirmation

The Kisatchie study is a clean example of AI doing what it’s actually good at: processing large spatial datasets faster than humans can, surfacing candidates worth checking, and compressing the search space before expensive field work begins.

It didn’t replace the archaeologists. It gave them fewer places to dig.