pollinator_visits.xlsx

Plant × hour pollinator visitation dataset derived from time-lapse image detections and meteorological covariates (2023–2025).
Insect visitation was quantified from time-lapse photographs acquired every 5 min and processed with an internally developed YOLOv8-based detection model (predpolli; INRAE forge) to detect and classify insects as bees (e.g., Apis spp., Megachile spp., Halictidae) and bumblebees (e.g., Bombus terrestris, B. lapidarius, B. pascuorum). Model performance was assessed against expert annotation on 5,000 randomly selected images for bees and bumblebees (precision/recall/F1: bees 0.906/0.954/0.929; bumblebees 0.750/1.000/0.857). Each photograph was time-stamped and linked to hourly meteorological data recorded by a nearby weather station (INRAE 31035002; <500 m). Detections were aggregated at the plant × date × clock-hour level (Europe/Paris time) by summing counts across all images collected for a given plant within the same hour; n_photos reports the sampling effort.
This table reports the plant × hour dataset prior to late-flowering filtering. Columns include plant identifiers (Code_Plant), time stamp (Time, time_rounded, Date), aggregated counts (total_bees, total_bumblebee), sampling effort (n_photos), trial/environment (Trial), allelic class (Allele_HaCWINV2), NIL pair (Pair), and hourly meteorological covariates: PAR_H (photosynthetically active radiation, J·cm⁻²), I5 (actinothermic index at 50 cm, °C), U (relative humidity, %), and VT (wind-run distance, km). For statistical analyses, this dataset was subsequently restricted to daytime observations (06:00–21:59) and to the first 99% of each plant’s cumulative visitation curve, thereby excluding late-season images with very low visitation and capitula bearing very few florets (typically <10).