Line-transect distance sampling lab

Walk a survey line, record how far each animal was, and correct for the ones you missed.

Presets

Playback

Survey

Camera

Layers

Quality

Model check

The science

What is modelled

A wildlife survey cannot count every animal, because the ones far from the observer are more likely to be missed. In line-transect distance sampling the observer walks a straight line, records the perpendicular distance x to every animal seen, and uses how detections fall off with distance to estimate what fraction was missed. Here the animals are randomly placed (a Poisson process of density D per km²) in a strip 300 m wide, the observer walks the line, and each animal inside the truncation distance w is detected with probability g(x).

Equations

  • Detection function (half-normal): g(x) = exp(-x² / 2σ²), so g(0) = 1 (everything on the line is seen). The alternative “hazard-rate” truth g(x) = 1 - exp(-(x/σ)-3) has a flat shoulder near the line.
  • Effective strip half-width μ = ∫0w g(x) dx; for the half-normal μ = σ √(π/2) erf(w / σ√2). Detection probability inside the strip is Pa = μ / w.
  • Density estimate D̂ = n / (2 L μ̂), where n is the number of detections and L the line length. The naive n / (2 L w) pretends Pa = 1 and is biased low.
  • σ is fitted by maximum likelihood to the truncated half-normal density f(x) = g(x)/μ: maximise -Σx²/(2σ²) - n ln μ(σ) (golden-section search).
  • Counts of detected animals are Poisson, so n has variance equal to its mean.

Checks run in the page

  • Closed-form erf formula for μ against a numerical Simpson integral.
  • Maximum likelihood on 20,000 simulated detections recovers the effective half-width μ (and σ, which is only weakly determined when σ is much larger than w).
  • Mean of 300 replicate surveys against the true density (the replicates use a transect long enough to give about 600 detections each, half-normal truth).
  • Variance-to-mean ratio of the detections over replicates is close to 1 (Poisson).

Honest simplifications

  • Animals are stationary and detected exactly at the moment the observer is abeam of them; real surveys must deal with animals that move before being seen, rounding of distances and detections that are not independent (groups, clusters).
  • g(0) = 1 is assumed: every animal on the line is seen. If that fails the estimate is biased low.
  • Animals are drawn about four times life size so they can be seen from the air. The animal models are generic and stand for no particular species. The 300 m strip and the 60 m/s walking speed are conveniences of the animation, not realistic effort.
  • The estimator always fits a half-normal. With the hazard-rate truth this is the wrong model and the bias reading shows how much that costs. Real analyses compare several candidate detection functions with AIC and add adjustment terms.
  • Variance, confidence intervals and cluster size are not computed. Nothing here is advice for designing a real survey.

What to try

  • “Thick bush” has a short σ: far animals vanish, the naive estimate is far too low, but D̂ stays near the true value.
  • “Rare species, long walk” shows why a low density needs a long line: the replicate histogram is wide.
  • “Shouldered detection (hazard-rate)” makes the half-normal fit slightly wrong and shows a bias.
  • Shrink w: you waste detections but the estimate stays unbiased.
Loading model...
Loading Blender model...
Detected distances and detection shape
Log-likelihood of sigma
Density estimates over 300 surveys
Made with 🟠 Blender · Three.js