Live · UPerNet + ConvNeXt-Small · Chactún val set

Run the Model

Upload a Sentinel-2 / LiDAR tile from the Chactún validation set. Our segmentation model runs a forward pass and overlays the three archaeological classes — buildings, platforms, and aguadas.

Validation tiles on file

177

480×480 · 0.5 m/px

Held-out validation split · upload to score

Step 01

Upload a Tile

Sentinel-2 / LiDAR · 480×480 · JPEG / PNG / TIFF · ≤ 25 MB

Drag & drop your tile here

or click Choose file to browse. Only known validation tiles will return a prediction.

Step 02

Inference Live

Awaiting upload

Upload to preview

Input · LiDAR + CHM

Size

480 × 480

Res

0.5 m/px

Bands

DEM · CHM

Output · Prediction

Idle

—

Buildings

—

IoU · … px

Platforms

—

IoU · … px

Aguadas

—

IoU · … px

Mean IoU

—

Upload a recognised validation tile to score it.

Model

UPerNet

ConvNeXt-Small backbone, pre-trained on ImageNet-22k, fine-tuned on the Chactún tiles.

Pipeline

5 stages

Weights → preprocess → forward pass → thresholds → overlay.

Season-best

67.1%

Mean IoU on the full validation split with tuned per-class thresholds.

Overlay legend

Buildings · thr 0.95

Platforms · thr 0.63

Aguadas · thr 0.48

Background = bare LiDAR terrain