From a stereo pair to a geocoded DSM — on one desk.

Automatic Terrain Extraction

The full stereo → DSM pipeline, server-side

Seven stages run as an async job you poll — from RPC extraction through a math-correct triangulation that fixes height by minimizing the two-image reprojection residual.

1. RPC extraction — both images opened via GDAL; their rational polynomial sensor models read from the inline GeoTIFF domain or .rpc/.rpb sidecar.
2. Epipolar grid — the rectifier seeds the triangulator's ground-frame mapping.
3. Dense match — the pluggable matcher produces a per-pixel disparity map (NaN where unmatched).
4. RPC triangulation — each valid disparity pixel → (lon, lat, height) via a bounded height search with golden-section refinement.
5. DSM write — a tiled DEFLATE Float32 GeoTIFF with NaN NoData.
6. Register COG — the DSM becomes an ImageServer layer.
7. Register DemSource — the DSM also enters the terrain stack so viewshed / LOS / slope use it at once.

# probe the matcher sidecar
GET  /api/stereo/capabilities
     → { matcher:{name,available}, pipeline:[…] }

# start a stereo → DSM job
POST /api/stereo/dsm
     { "leftCogId":"wv3-left",
       "rightCogId":"wv3-right",
       "options":{"HeightMax":9000} }
     → { "jobId":"a1b2c3…" }

# poll → produced DSM
GET  /api/stereo/jobs/{jobId}
     → { "status":"succeeded",
         "dsmCogId":"ate-dsm-a1b2",
         "demSourceId":"ate-dsm-a1b2",
         "validPixelFraction":0.83 }