GSD & scale recovery

The network predicts height in metres at 0.5 m GSD, so every result depends on knowing the input’s true pixel size.

Resolution order

flowchart TD
  A[Input image] --> B{GSD supplied?<br/>--gsd / Resolution field}
  B -- yes --> U[Use supplied GSD]
  B -- no --> C{GeoTIFF with<br/>geotransform?}
  C -- yes --> G[GSD = pixel size<br/>in metres; reprojected<br/>if geographic CRS]
  C -- no --> D[Assume 0.5 m]
  U --> R[Resample to 0.5 m]
  G --> R
  D --> R
  R --> M[Model → nDSM in metres]
  M --> P{Georeferenced?}
  P -- yes --> Q[nDSM + optional<br/>absolute DSM]
  P -- no --> S[rDSM<br/>scale = assumed/supplied GSD]

GSD source	Recorded as (meta.json → scene.gsd_source)	Product label
User value	user	nDSM (or rDSM if not georeferenced)
GeoTIFF transform	geotiff	nDSM / DSM
Assumed 0.5 m	assumed	rDSM

Resampling

• Inputs finer than 0.5 m (for example 0.3 m WorldView) are downsampled to 0.5 m, predicted, and the height map is resampled back to the input grid.
• Inputs coarser than 0.5 m (for example 0.6 m Cartosat PAN or 1.0 m products) are upsampled to 0.5 m. The model cannot recover detail that the image does not contain, but heights stay metric.
• Very large images can be limited with max_side. The service switches to windowed processing above 40 MP instead.

What a wrong GSD does

Height cues such as shadow length, roof size and parallax scale with the true ground distance. If an image is really 1.0 m/px but is processed as 0.5 m/px, every structure looks twice as big in pixel terms and is predicted roughly twice as tall. The web app therefore:

• reads the GSD from GeoTIFFs automatically and shows its source,
• warns when no resolution is known and 0.5 m will be assumed,
• labels the result rDSM so it is never mistaken for a calibrated product.

Tip

For a screenshot from Google Earth or a web map, estimate the GSD from the scale bar (metres ÷ pixels), or measure a known object such as a tennis court (23.77 m long). This is the same method used to re-measure GAMUS.

Multi-band and Cartosat products

• RGB is taken from bands 3, 2, 1 of 4-band products by default. Override it with bands= (API) or --bands.
• An NRSC product delivered as BAND1..4.tif + BAND_META.txt can be uploaded as several files at once, or as a .zip.
• 16-bit data is mapped to 8-bit with the scene’s 2–98 % stretch, the same as in training.