Reliable density depends on the sampling and observation models, not detector accuracy alone.
Make the area denominator explicit
Density divides abundance by an area, but the flight footprint and observable area are not necessarily identical. Define how canopy, image edges, overlapping frames and missing plots affect the denominator. If detection differs across habitats, an apparent density difference may be an observation difference. Preserve the strata needed to model that distinction.
Model three different errors
Missed detections, false positives and duplicate counts require separate treatment. A video track ID is not automatically a persistent biological identity across flights. Combine time, position uncertainty, plausible movement and species cues, and retain ambiguous matches for review. Deduplication rules should be versioned so a changed rule can be traced into the final count.
Correction requires defensible assumptions
A simple correction divides a cleaned count by detection probability, but animals may have unequal probabilities of being observed. Repeated-count models also depend on assumptions about closure, independence and observation error. Research on automated koala surveys has compared N-mixture and modified Horvitz–Thompson approaches rather than treating detector output as abundance. Choose and diagnose a model suited to the sampling process.
Report an estimate with its limits
Report the estimate, interval, period, area, effort, model and detection-validation method together. Evaluate uncertainty in the change itself, including paired sampling units and dependence between years; overlap of separate annual intervals is not a definitive test of change. A research interface should show the conclusion, the change interval, limits to comparability, and then supporting images and calculation history.
Separate sampling inclusion from detection
A plot’s probability of being sampled differs from an animal’s probability of being observed within that plot. The teaching expression N̂ = n / p assumes a defensible common detection probability; a detector confidence score cannot be substituted for p. Document the separate treatment of extrapolation, availability, false detections and movement duplicates. A correction for one mechanism does not automatically correct the others.
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.