Key concepts

These terms are used throughout the docs.

Surface model products

The model predicts the nDSM: the height of each pixel above bare ground. Adding a terrain model (DTM) gives the absolute top surface (DSM).

Product	Meaning	When DepthWizard produces it
nDSM	Normalised DSM: metres above ground for buildings, trees and other objects. Bare ground is about 0	Always. This is what the network predicts
rDSM	Relative DSM: an nDSM whose metric scale depends on an assumed GSD	Plain PNG/JPG input with no known resolution
DTM	Digital terrain model: bare-earth elevation above a datum	Georeferenced input, fitted from a public DEM
DSM	Digital surface model: elevation of the top surface above a datum, equal to DTM + nDSM	Georeferenced input with absolute enabled

Ground sample distance (GSD)

The GSD is the ground size of one pixel, in metres. It is the key that lets heights come out in metres. The model always works at a canonical 0.5 m GSD, so every image is resampled to 0.5 m before inference and the result is resampled back to the input grid afterwards.

The GSD is resolved in this order:

1. a value you supply (--gsd, or Resolution in the app),
2. the pixel size in the GeoTIFF’s geotransform,
3. otherwise an assumed 0.5 m, in which case the output is labelled rDSM.

A wrong GSD scales every height by roughly the same factor. For example, a 1.0 m image processed as 0.5 m would show buildings about twice as tall.

Vertical datums

Absolute elevations are only meaningful relative to a stated datum:

Datum	Used by
EGM2008 geoid	Copernicus GLO-30 (the default DEM)
EGM96 geoid	SRTM GL1, NASADEM, Terrarium tiles
WGS84 ellipsoid	CartoDEM, raw GNSS / LiDAR

DepthWizard converts between them with PROJ geoid grids and writes a compound vertical CRS into every DSM GeoTIFF.

Test-time augmentation (TTA)

TTA runs the model on the 8 flips and rotations of each tile (the dihedral group D4) and averages the predictions. Across our runs it lowered RMSE by 0.01–0.12 m, at 8× the compute (see Benchmarks). In the app this is the Higher quality (TTA) switch.

Glossary

Term	Meaning
RMSE / MAE	Root-mean-square / mean absolute error of predicted height, in metres
δ1	Share of pixels where the ratio of (prediction + 1) to (truth + 1) is below 1.25
Balanced RMSE	Mean of the RMSEs in five height bands, so tall structures count as much as flat ground
Height stratum	One of the bands 0–2, 2–5, 5–10, 10–20 and 20+ m
Sliding window	Whole-scene inference over overlapping 512 px tiles, the path the app uses
OOD	Out-of-domain: data from a sensor or region never seen in training