WV-parel sticky traps — interim analysis

Project 0043_wf-wvparel · export all_2026-09-28_0124 (catalogue 2026-09-28 01:24) · generated 2026-09-28

Data & design

Scope: sticky traps from the main design only (plasdras-pilot excluded). Traps marked missing are dropped. Each card = one trap position × round (r1–r3) × year, with 2 days in the field. Positions 1–5 are 1, 20, 40, 60 and 80 m from the ditch edge.

Two parallel analyses:

• Manual track (2024–25): boxes drawn by people and labelled by people. geen-dier excluded; unknown animals are counted but have no taxon.
• Model track (2024–26): cls-hier v4.1 labels on the boxes that exist. In 2024–25 these are the human boxes; in 2026 they come from detector od-card v5.0 plus 108 QC boxes. Model geen-dier (including the gate) excluded.

Cards per location × beheer × year

Unbalanced design: Lollum has no ontwikkeling plots, and Tjerkwerd has two. Location is a fixed effect, so beheer contrasts compare within locations.

Composition: class / order / family

Diptera make up 55–80% of the catch in every group; Hymenoptera come second. The full mean-per-card tables for class, order and family (both tracks, beheer × year) are in out/tables/mean_per_card_by_taxon_beheer_year.csv.

Abundance, biomass and body size

Estimates are marginal means from the GLMM (averaged over location, round and distance). The beheer × year and beheer × distance panels come from models that include that interaction.

Biomass and length classes are model-track only. Every body length, on human and detector boxes alike, comes from the body-length model; a “manual” biomass would repeat the model-track numbers with a slightly different set of animals, so it is not reported. Biomass is estimated per object from the model body length with the general insect length–dry-mass relation of Rogers et al. (1976), W(mg) = 0.0305 · L(mm)^2.62. It is a relative index: treat the ratios as meaningful and the absolute mg values as rough.

Biodiversity (family level)

The unit is the family, with two exceptions (†): Ichneumonoidea and Chalcidoidea are each one superfamily-level unit. Every object resolved to at least that superfamily counts in it, except the families the model does separate (Braconidae, Mymaridae, Torymidae). Family richness and Hill q1 per card count only objects resolved to a unit. Human labels reach a unit for only ~27–34% of animals, so manual-track diversity reflects annotation effort much more than the insects themselves.

Rarefaction / extrapolation curves

Hill numbers (q0 = richness, q1 = exp(Shannon), q2 = inverse Simpson) as the number of family-resolved animals grows (iNEXT, 100 bootstraps). Solid lines are rarefied, the dot is the observed sample and dashes are extrapolated to twice the sample. q1 and q2 level off almost immediately, so they are well estimated. q0 is still rising at the observed sample: rare families keep being added, and richness comparisons only hold at a common coverage (below).

Diversity at a common coverage (98%)

Why 98%: the smallest group (manual, 2024, ontwikkeling: 1,393 animals, 13 families seen once) reaches only 99.1% observed coverage. At a 99% target it sat right at its own coverage, so bootstrap replicates flipped between rarefaction and extrapolation and the q0 interval blew up (roughly 26–85 families). At 98% every group is well inside rarefaction.

With the model labels: q0 (richness) hardly differs between beheer types. q1 and q2 (weighted towards common families) are highest in parel in 2024 and 2025. In 2026 parel and ontwikkeling are equal, and gangbaar is clearly the least even (dominated by a few families).

PERMANOVA on Hellinger-transformed family counts. Plot × year level: marginal tests for locatie, beheer and year. Distance: trap × year totals, permuted within plot-year.

All effects at a glance

Ratios are estimated from the GLMM, each relative to its reference level. Blue = higher, red = lower; a cell is shaded only when that term is significant after Benjamini–Hochberg correction (across responses within a track). Hover a cell for the 95% CI. “≈0” / “n.e.” = zero counts in one level (e.g. Megaloptera at 80 m, or manual family labels that were not used in 2024).

† Analysed at superfamily level (see Biodiversity and the manual-vs-model section). Families: top 10 of each track. Biomass and length classes: model track only. Manual-track family counts inherit annotation depth, which differs between years.

Manual vs model labels (2024–25)

Both labels sit on the same boxes, so this compares classification only. Test cards were never seen in classifier training and give the clean estimate. Train/val cards were used for training and are shown for contrast. The human label is the reference at the rank the annotator chose; anything the model adds below that cannot be checked directly.

Agreement when both answered the same rank

Test performance is as high as on train/val (98–99% at every rank). The model generalises to new cards from the same sites and years; there is no sign of overfitting to the training cards.

Counts per card

Family richness per card: humans 7.7 vs model 16.2 families on test cards (Spearman 0.66). Humans record fewer families because most of their labels stop above family, not because the model invents families.

How deep does each go?

Is “model deeper” right?

Three indirect checks, since there is no human label below the annotator’s rank:

1. Calibration. On test objects that people did label to family, model family calls are 98.5% correct, and still ≥88% in every score bin. Applying that score-to-accuracy curve to the 3,570 unverifiable deeper family calls on test cards gives an expected ~95% correct. This is an upper bound: people stopped early on exactly the harder objects.
2. Score distribution. Deeper calls have lower scores than verifiable ones (right panel below), which fits them being harder.
3. Training support. For each family the model assigns in a deeper call, the table counts how many family-level human labels existed on this project’s training cards. Families with no support here (Ichneumonidae, Chalcididae, Mymaridae, Agromyzidae, Anthomyiidae) are learnt only from other datasets, or are a default child of the superfamily, and cannot be checked locally.

* Test precision is computed on objects that both labelled to family; empty where people never used that family.

Most common deeper assignments (test cards)

Methods & caveats

• Models: glmmTMB, y ~ beheer + year + ronde + distance + locatie + (1|event) + (1|plot) + (1|trap). Event = year × round (shared weather and phenology); trap = plot × position (repeated over rounds and years). Counts use a negative binomial model (NB2, falling back to NB1 or Poisson if it does not converge), biomass a Tweedie model, richness a Poisson model and log Hill q1 a Gaussian model. Terms are tested by likelihood ratio; beheer × year and beheer × distance are tested by adding each to the main model. p-values are BH-adjusted per term across responses within a track.
• Year is tested against round-to-round (event) variation. With 3 rounds per year this test is conservative: large year differences can still be “not significant”.
• 2024–25 vs 2026 boxes: 2026 counts come from the detector, 2024–25 from human boxes (used as is, as agreed). Part of any 2026 difference may come from detection rather than insects.
• Rounds r1–r3 are treated as the same period in every year, as agreed. Deployment dates are not in the data for 2024–25.
• Taxonomic resolution: counts at class, order or family use only objects resolved to that level. Family-level work uses the model labels as the main result and the human labels as a check.
• 4 of 434 likelihood-ratio tests could not be computed (sparse taxa, non-positive-definite Hessian); they show “–”.

Scripts: analysis/R/00_prepare.R → 01_ecology.R → 02_manual_vs_model.R → 03_report.R. All tables in analysis/out/tables/ (mvm_* = manual vs model), figures in analysis/out/figures/.