Which modelling choices are supported when marine survey data are limited? An evidence map of 95 studies and a benchmark on structured subsamples of a trawl survey

Ebenezer Afrifa-Yamoah , Alexandre C Siquerira, Yaw Kwaafo Awuah-Mensah, Francky Fouedjio, Ute Mueller

  1. Marine management decisions increasingly rest on species distribution
  models fitted to sparse, clustered survey records, while the guidance
  for fitting them comes from data-rich settings. We asked which
  modelling choices have been tested under the conditions marine
  practitioners face, and whether the ordering that guidance assumes,
  spatial structure before algorithm choice, holds at the sample sizes
  at which it is applied.

  2. We screened 330 records and retained 95 peer-reviewed marine
  applications that treated spatial autocorrelation explicitly, coding
  the contrasts tested, the sample sizes reported and the metric family.
  We then benchmarked, on the NWFSC West Coast Groundfish Bottom       Trawl Survey, subsamples of three demersal species at 30 to 500 records
  crossed with dispersed and clustered coverage and with interpolation
  and extrapolation targets, fitting random forests and generalised
  additive models with and without spatial terms and a Gaussian Markov
  random field, evaluated against the held-out survey and by random and
  spatial block cross-validation.

  3. No marine study crosses spatial treatment against algorithm family on
  the same data, and one reports performance below 100 records. In the
  benchmark, spatial structure helped or harmed according to how it was
  added: a coordinate smooth in a GAM reduced held-out AUC by 0.06 to
  0.15 below 200 records and at every size under clustered sampling, a
  Matérn random field recovered most of that loss, and coordinates in a
  random forest helped only with dispersed coverage and an interpolation
  target. Dispersed coverage was worth as much as the gain from 30 to
  500 records. Random cross-validation erred by up to 0.13 AUC in either
  direction according to coverage, and spatial blocking moved the
  estimate by less than 0.03. Likelihood models were overconfident at 30
  to 50 records; a random forest with coordinates was nearly calibrated.

 4. Spatial against non-spatial is not a contrast on which to base a
  recommendation; coverage is the first-order decision; internal
  validation does not reveal extrapolation loss when the sample is a
  band; and at the smallest sizes model choice affects the probability
  scale more than the ranking. The corpus and benchmark yield a
  practitioner guide graded by the evidence behind each element.