Output formats

Source: write_outputs in Model_Traning/v5/infer/predict.py. The same writer is used by the CLI, the FastAPI service and the Space.

Files

File	Type	Units / content	When
`ndsm_m.npy`	Float32 array H × W	metres above ground; NaN = no data	always
`ndsm_m.tif`	Float32 GeoTIFF	same, with CRS + geotransform	georeferenced input
`ndsm_std_m.npy` / `.tif`	Float32	per-pixel uncertainty σ (m) from Head B	always (TIFF if georeferenced)
`dtm_m.npy` / `.tif`	Float32	bare-earth elevation above datum	`absolute`
`dsm_m.npy` / `.tif`	Float32	absolute surface elevation above datum	`absolute`
`seg.npy` / `seg.png`	uint8	class id 0–6 (classes)	always
`rgb.png`	8-bit RGB	stretched input on the output grid	always
`ndsm16.png`	16-bit grey PNG	quantised height (encoding below)	always
`terrain.glb`	glTF binary	textured height mesh	`mesh=true` (default)
`terrain.obj` + `.mtl` + `texture.png`	Wavefront OBJ	same mesh	`mesh=true`
`shadow_img.png`	binary PNG	detected shadow mask	always
`meta.json`	JSON	stats, scene, preprocessing, files	always
`report.html`	HTML	self-contained report	on request
`validation.json`, `gt_ndsm_m.npy`, `err_ref_m.npy`	JSON / arrays	reference comparison	after `/api/reference`

GeoTIFF DSM/DTM outputs carry a compound CRS: the scene’s horizontal EPSG plus the vertical datum of the DEM used.

`ndsm16.png` encoding

height_m = height_min_m + (png16 / 65535) * (height_max_m - height_min_m)

height_min_m and height_max_m are stored in meta.json. The quantisation step is (max − min) / 65535: about 0.2 mm for a 12 m scene and 1.7 mm for a 112 m scene.

`meta.json`

Abridged from a real Cartosat-2E job:

{
  "stem": "input",
  "product": "nDSM",
  "units": "metres_above_ground",
  "height_min_m": 0.0000013,
  "height_max_m": 11.812,
  "height_mean_m": 0.253,
  "height_median_m": 0.0039,
  "frac_below_1m": 0.923,
  "size_px": [1024, 1024],
  "scene": {
    "width": 1024, "height": 1024,
    "gsd_m": 0.6, "gsd_source": "geotiff",
    "georeferenced": true,
    "crs": "EPSG:32644", "crs_epsg": 32644,
    "transform": [0.6, 0.0, 213806.4, 0.0, -0.6, 1727287.8],
    "valid_frac": 1.0,
    "source": { "rgb_bands": [3, 2, 1], "pansharpened": false }
  },
  "preproc": {
    "encoder_model_id": "facebook/dinov3-vitl16-pretrain-sat493m",
    "mean": [0.43, 0.411, 0.296], "std": [0.213, 0.156, 0.143],
    "canonical_gsd_m": 0.5, "tile_size": 512, "patch": 16,
    "radiometric_stretch": true, "stretch_lo_pct": 2.0, "stretch_hi_pct": 98.0,
    "target": "nDSM_agl_metres"
  },
  "vertical_datum": "EGM2008",
  "files": ["rgb.png", "ndsm16.png", "ndsm_m.npy", "dsm_m.npy", "dtm_m.npy", "…"],
  "ndsm16_encode": "height_m = height_min_m + (png16/65535)*(height_max_m-height_min_m)",
  "uncertainty": { "std_median_m": 0.251, "…": "…" }
}

Key	Meaning
`product`	`nDSM`, or `rDSM` when the GSD was assumed
`scene.gsd_source`	`geotiff`, `user` or `assumed`
`scene.transform`	GDAL-order affine geotransform
`frac_below_1m`	share of pixels below 1 m. The web app warns when a scene has very little flat ground
`vertical_datum`	datum of `dsm_m` / `dtm_m`

Web app exports

The browser can re-export any loaded result:

Group	Formats
3D object	GLB, OBJ (zip with MTL + texture), PLY, STL, the server’s `terrain.glb`
Heatmap	PNG, JPG
Elevation data	GeoTIFF Float32, GeoTIFF nDSM and DTM (after anchoring), NumPy `.npy`, 16-bit PNG
Project	`.dwproj` (reopen later, including settings)