Ingest and photogrammetry

Frames and a flight log go in. Georeferenced, calibrated raster layers come out. This is the stage that is normally rented per hectare, and the one where the open source alternatives are strongest.

Geotagging

See Positioning. The lever arm is the step everybody skips and it is systematic, so it does not average out.

Radiometric calibration, for multispectral

Run this before anything else touches the pixels.

1. Dark current and vignetting, per band, from the camera's calibration data.
2. Panel calibration: the reflectance panel imaged before and after the flight gives the digital-number to reflectance scaling.
3. Downwelling correction: the light sensor's per-frame irradiance handles cloud passing during the flight.
4. Band alignment: separate lenses see slightly different scenes. Register the bands with feature matching before computing any index.

The reference implementation everybody learns from is MicaSense's own imageprocessing notebooks. Read them, then write your own, because the discipline has to fit your workflow.

Without this, a "change layer" is largely a record of the weather. See disease detection.

Photogrammetry

The AGPL question matters. Running unmodified ODM as an internal batch job that produces files you then serve is the ordinary, uncontroversial path. Modifying ODM and exposing it to customers over a network is where the network clause bites. Decide this before the integration exists. See the licence table.

Getting a good reconstruction

Runtime

Roughly 600 frames from 25 ha at 2.5 cm/px is 30 to 60 minutes on a mid-range GPU box. Plan the pipeline as a queue with retries, not as a synchronous request: a stitch that fails at minute 40 must not lose the flight.

Raster derivatives

All of it is GDAL and numpy, and none of it needs a licence.

You often do not need a mosaic

A full photogrammetric stitch is expensive and, for several capabilities, unnecessary. If every frame has an accurate pose, you can project its corners onto the ground and place detections in field coordinates directly, then deduplicate across overlap.

Skipping the mosaic where it is not needed is the single largest processing cost saving available, and it also removes hours of latency between landing and a usable result.

Thermal alignment

Thermal frames over a uniform canopy have almost no features to match, so they mosaic badly on their own. Fly an RGB camera alongside, solve the geometry from the RGB frames, and apply that solution to the thermal frames using the fixed transform between the two. It is the most useful trick in the whole thermal pipeline.

Publishing a stitch to the site

The web derivative, not the master:

gdal_translate -of WEBP -outsize 3000 0 -co QUALITY=82 ortho.tif ortho-web.webp

Then attach it to the flight record, with the footprint in the field's normalised space:

buddy imagery:attach ./ortho-web.webp --bounds="-0.04,-0.03,1.05,1.02" --resolution 4

The footprint comes out of the GeoTIFF rather than being guessed: read the geotransform, express the image's corners in the field's own bounding box, and that is your minX,minY,maxX,maxY. See The field map for why the equals sign is required and why the bounds matter.

Quality gates

Write these as automated checks that fail the flight rather than as things somebody notices later:

A flight that fails a gate is reflown. That is far cheaper than a prescription built on a bad map, and the customer never sees it.