Ops evidence

Proof from the pipeline

These are real assets copied from the agri-edge engineering tree — CAD airframe renders and evaluation imagery. Captions state what each file is. Metrics match HANDOFF: sheep AP@.50 0.905; imageweeds species ~0.457 / binary weed 0.673; weed_crop is pipeline proof only.

Airframe

CAD renders of the planned under-25 kg RPAS layout.

3 assets
Airframe — isometric

airframe

Evidence

Airframe — isometric

CAD isometric of the planned under-25 kg RPAS airframe. Companion computer (Pi 5 + Hailo AI HAT) rides alongside a Pixhawk/Cube autopilot; the Pi never commands the aircraft.

Airframe — top view

airframe

Evidence

Airframe — top view

Top orthographic view of the same airframe layout. Payload bay sized for downward camera, reflectance panel workflow, and store-and-forward spool storage.

Airframe — side view

airframe

Evidence

Airframe — side view

Side orthographic view showing landing gear clearance and belly camera mounting. A belly camera follows the belly, not the wing — roll/pitch handedness matters for georeferencing.

Detection comparison

Classical ExG green-on-brown versus YOLOX on the same frame.

1 asset
Green-on-brown vs YOLOX

detection

Evidence

Green-on-brown vs YOLOX

Side-by-side comparison: classical ExG green-on-brown (OpenWeedLocator, MIT) versus YOLOX-s inference on the same frame. Green-on-brown runs ~62 ms with no NPU; YOLOX ~270 ms on CPU. Different confidence semantics — set thresholds explicitly when switching.

Livestock

Ground truth and ONNX predictions from aerial_sheep_yolox_s (AP@.50 0.905).

2 assets
Sheep — ground truth

livestock

Evidence

Sheep — ground truth

Ground-truth annotation from the aerial sheep evaluation set. Model aerial_sheep_yolox_s scored AP@.50 0.905 on 5,603 boxes. Livestock targets are large enough that this model runs usefully on a bare Pi with no accelerator.

Sheep — model prediction

livestock

Evidence

Sheep — model prediction

Example detection output from aerial_sheep_yolox_s through the deployed ONNX engine (letterbox + host-side decode). Same path the Pi runs — not a training-only torch path.

Weeds

ImageWeeds species ~0.457 / binary weed 0.673; weed_crop is pipeline proof only.

2 assets
Weed binary — prediction

weeds

Evidence

Weed binary — prediction

Example weed detection output. imageweeds_yolox_s: species AP@.50 ~0.457; collapsed binary weed AP@.50 0.673. Species ID still needs locally annotated field imagery before validated_locally can become true.

Weed species — prediction

weeds

Evidence

Weed species — prediction

Five prairie species head (horseweed, ragweed, kochia usable; redrootpigweed and waterhemp weak — treat those labels as “a weed,” not an ID). weed_crop_yolox_s is pipeline proof only (seedling/tray imagery — not field aerial).