Validation against a reference

The Validation tab compares the prediction with a reference raster you trust, such as a LiDAR nDSM, a stereo DSM or a surveyed model.

1. Open Inspector → Validation, or Tools → Validate against reference.
2. Load a reference .tif (GeoTIFF) or .npy array.
3. Set Reference type:
   • Height above ground (nDSM): compared directly with the prediction.
   • Elevation (DSM): absolute heights. Turn on Remove ground offset to compare its shape with the height-above-ground estimate, or anchor the prediction first (Anchoring).
4. Read the metrics and charts, then use the layers to find where the errors are.

GeoTIFF references are aligned by coordinates. Plain arrays must match the prediction grid. Reference aligned confirms the overlap.

Metrics

Metric	Meaning
RMSE	Root-mean-square error (m)
MAE	Mean absolute error (m)
Bias	Mean signed error. Positive means the model predicts too high
Pearson r	Correlation
Within ±1 m / ±2 m	Share of pixels with error inside the band
Balanced RMSE	Mean RMSE across height bands 0–2, 2–5, 5–10, 10–20 and 20+ m
N	Number of compared pixels

These use the same definitions as the model evaluation (Metrics).

Charts

• Density scatter: predicted vs reference height. A tight diagonal is good, and a slope below 1 means tall objects are under-predicted.
• Error histogram: the distribution of prediction − reference.
• RMSE per height band (Per-stratum metrics): shows whether error comes from flat ground or tall structures.

Visual comparison

• Error layer: drapes signed error on the terrain (blue under, red over).
• Reference layer: shows the reference’s own heights.
• Compare swipe (prediction | reference): a slider splits the view, with the prediction on one side and the reference on the other.

Export a report

Export report saves a self-contained HTML file with the metrics, charts and scene details. Anyone can open it in a browser, with no server needed.

Caution

A 30 m DEM (SRTM, Copernicus) is a poor reference for building heights: it smooths every structure over 30 m cells. Use LiDAR or stereo products for per-pixel validation. For a DEM reference, compare 30 m cell averages instead, which the self-hosted service reports as per_cell.