Automation should make every result traceable to its imagery, models and review decisions.
Ingest a survey package, not loose images
Ingest imagery with survey identity, logs, calibration, timestamps, sampling units and missing-data reasons. Check integrity and prevent duplicate ingestion while preserving immutable originals. Bind inputs to software, configuration and model versions. This is an implementation proposal, not a claim that aircraft or live inference services are already connected to the guide.
Put quality gates around detection
Check corruption, blur, exposure and positioning before inference. Tile if needed, detect candidates, return them to original coordinates and reconcile overlaps. Preserve excluded areas and times. Silently dropping poor imagery would otherwise convert an observation failure into an apparent absence of wildlife.
Separate pseudo-labels from verified labels
Machine labels are candidates. Review uncertain cases and audit samples of confident predictions and negative images. Record who corrected a label and why before promoting it to a verified set. This reduces exhaustive inspection while preserving a way to detect systematic omissions and overconfident errors.
Make processing resumable and traceable
Manage ingestion, quality checks, inference, deduplication, review, estimation and reporting as resumable jobs. Use idempotent identifiers so retries do not multiply records. Before release, check unresolved observations, validation conditions and uncertainty. Link the report back to source evidence and calculation history.
Keep the review sampling design
In the Airffic design proposal, retain the sampling frame, selection method and sampling fraction for each review stratum. Error rates from a deliberately difficult review queue do not directly estimate survey-wide performance. Maintain an independent audit sample including areas with and without candidates, and return review effort to the cost ledger. Hard examples for retraining and samples for measuring performance serve different purposes.
Evidence & further reading
Check the original study conditions, species and location before using a result in a plan. Do not transfer findings universally across species and settings.