U-Net Model Output

See what a trained segmentation model predicts on real Maya LiDAR data

What you're seeing: A U-Net with a ResNet34 encoder (pre-trained on ImageNet, fine-tuned on 100 Chactun tiles) processes 3-band LiDAR visualizations (Sky-View Factor, Positive Openness, Slope) and outputs a per-pixel probability of archaeological structure. Brighter = higher confidence. Compare the model's prediction to the expert-annotated ground truth.

Tile:

Side by side

Ground truth: Buildings Platforms Aguadas (reservoirs) Prediction: Low → High confidence
Input (SVF)
Model prediction
Ground truth
Loading...

Overlay view

Prediction heatmap overlaid on the terrain. Adjust opacity and threshold.

50%
0.50
-- Precision
-- Recall
-- IoU
-- F1

Adjust the threshold slider to see how precision, recall, and IoU change — the same tradeoff from the Threshold Lab, but now with a real model's output.