Counting Is Not Detecting: Resolution, Cadence and Decision Error in Automated Sticky-Trap Insect Monitoring

Automated insect monitoring on yellow sticky traps is a mature technology: convolutional detectors reach high average precision and commercial camera traps are sold as a service. Yet the metric these systems are optimised for, detection F1, is not the quantity a grower acts on, which is a count compared against an economic threshold. We separate the two and show they disagree. Using the 284 annotated traps of the Yellow Sticky Traps benchmark (8114 objects), we first measure the trap itself: the yellow substrate suppresses the blue channel to a median background value of 3, while whitefly show a median blue excess of 190, so the substrate acts as a colour filter that makes blue a nearbackground-free detection channel. A two-stage detector built on this observation—blueexcess blob extraction followed by a nine-feature classifier, both cheap enough for an edge device—reaches recall 0.972 and precision 0.895 under trap-level grouped cross-validation. Sweeping image resolution then produces the paper’s first result: F1 is maximised at the lowest resolution tested (0.956 at 4.3 px per whitefly), where counting has already collapsed (MAE 8.1, R2 0.56), whereas counting is best at an intermediate resolution of 8.6 px per whitefly (MAE 2.8, R2 0.92, bias 0.27). Full resolution is worse than both, because more pixels surface more non-target clutter than the classifier rejects, inflating counts by 6.2 insects per trap. Optimising F1 therefore selects the wrong camera. We also find that the crowding failure mode emphasised in the literature is confined to traps carrying more than about 100 insects, a regime in which the threshold has long been crossed and the decision is no longer in doubt. Finally, embedding the measured counting error in a seasonal population and decision model over 720 zone-seasons shows that sampling cadence dominates counting accuracy by an order of magnitude: moving the same detector from weekly to daily reading cuts excess crop damage from 445 % to 21 % above an oracle, while the gap between a crude and an accurate detector at equal cadence appears not as delay but as a false-alarm rate of 33 % against 2 %. The practical implication is that cheap cameras read often beat expensive cameras read rarely.