Domain-specific detection models

It is tempting to build one detection engine and point it at four problems. It is also the reason so many monitoring platforms perform well in a demonstration and poorly in a deployment.

Different questions, different errors

A corridor model has to find individual tree crowns and separate them from each other, because the unit of action is one tree. Its expensive error is a missed crown — the one that grows into the conductor. A forest division model has to classify area, because the unit of action is a patch of ground. Its expensive error is a false clearing that sends a patrol into difficult terrain for nothing.

Those two models want opposite things from their thresholds. Tuning for the first makes the second noisy. Tuning for the second makes the first blind. Averaging the two produces a model that is mediocre at both and can be defended in neither review.

Scale is not a hyperparameter you can borrow

Detection models are trained at a particular ground sample distance, and their learned features are tied to how large the target appears in pixels. Take a model tuned for one scale and run it at another and recall does not degrade gracefully — it falls off a cliff, silently, while the model continues to return confident detections on whatever it does still find.

This is the single most common cause of a system that worked in a pilot and does not work after a change of imagery source. Nothing errors. The count falls, and the result is indistinguishable from an improvement on the ground.

Ground truth is domain-local too

A model is only as good as what it was scored against, and a benchmark from a different forest type, a different sensor and a different season is not a score — it is a hope. Reviewed ground truth from the deployment's own area is the only thing that tells you what the system is doing, and it is the part of the work nobody puts in a proposal because it is unglamorous and continuous.

What this looks like when it is done properly

• One detector per domain, tuned on that domain, with its own thresholds and its own accepted failure modes.
• Thresholds set from the asymmetry of the errors, not from whatever maximises a symmetric score.
• A confidence value that means something operationally, so low-confidence detections can be routed differently rather than shown identically.
• Re-scoring against fresh reviewed data on a schedule, because the ground and the imagery both change.

None of this is unusual. It is more work than a single model and a demonstration, which is why the shortcut continues to be taken.

An improvement in detection accuracy delivers little if it applies to a question the deployment does not ask.