Classifier funnel
Build and classifier facts
How the numbers are produced
Capture. Three Tactacam Reveal X Pro cellular trail cameras are mounted on host palms below the crown, PIR-triggered, with no-glow infrared at night. Photos are bulk-exported from the Reveal web app and ingested by hash, so re-exports never double count. Each frame carries a camera id and HST timestamp; temperature, wind and pressure come from the Reveal metadata when present and from Open-Meteo hourly reanalysis otherwise.
Night. Adult CRB are nocturnal, so all nightly figures run noon to noon and are named by the evening date. Hour-of-night charts start at noon for the same reason.
Classification. Frames are scored 0–1 by the active model recorded in the facts panel. The current model is a YOLOv8m detector fine-tuned from the University of Hawaiʻi CRB-Cam weights (Paryavi & Jenkins, 2025) on beetle images composited onto this site's own frames, followed by a temporal rule: a box that persists across consecutive frames from the same camera is boosted, a box with no neighbour is damped. Thresholds are calibrated per model from its validation F1 curve. Scores at or above τhigh are probable; scores between τlow and τhigh are pending human review; below τlow is rejected. A human label always overrides a score: confirmed means a reviewer marked the frame CRB. Precision and recall are computed on labelled frames with predicted positive = score ≥ τhigh. Until frames with real beetles are confirmed, the model's validation numbers refer to synthetic composites and the only measured quantity on real frames is the false-alarm rate.
Sampling caveat. Until a scheduled time-lapse stream exists, compare nights by confirmed detections and read trigger volume as a wind indicator, not an activity indicator.
Green-waste piles. Each pile has a first- and last-observed date from drone orthomosaics, ground-truth walks or KOCA records. Persistence class is the span between them: under 48 h (the larval-viability guidance), 48–72 h (inside KOCA's internal policy window), over 72 h. Compliance is the share of KOCA-jurisdiction piles not exceeding 72 h.
Damage. Crown-damage points come from the KOCA April 15 map, the micro-drone reconnaissance and repeat flights. Where treatment occurs nearby they are tagged before/after so efficacy can be described without the pilot conducting treatment.
Confidentiality. The public build aggregates piles and damage to zone centroids, withholds imagery and omits any feature whose landowner has not consented, per Section 12.4 of the pilot plan. Camera coordinates are rounded in the public build.
Known gaps
Exports
Every chart has a Table view and a CSV button. The underlying build files are linked here for the baseline assessment.