Deep learning object detection to estimate the nectar sugar mass of flowering vegetation

Floral resources are a key driver of pollinator abundance and diversity, yet their quantification in the field laboratory is laborious requires specialist skills. Using dataset 25,000 labelled tags fieldwork-realistic quality, convolutional neural network (Faster R-CNN) was trained to detect nectar-producing floral units 25 taxa surveyors’ quadrat images native, weed-rich grassland United Kingdom. unit detection on test set 50 model-unseen comparable vegetation returned precision 90%, recall 86% F1 score (the harmonic mean recall) 88%. Model performance consistent across range this habitat. Comparison nectar sugar mass estimates made by CNN three human surveyors similar means standard deviations. Over half model fell within absolute those surveyors. The optimal number image samples determined be same for as average surveyor. For sampling protocol 10–15 replicates, application deep learning could cut pollinator-plant survey time per stand from hours minutes. restricted single view quadrat, with no scope manual examination or specimen collection, though contrast its object deterministic definition standardized. As agri-environment schemes move prescriptive results-based, approach provides an independent barometer management which usable both landowner scheme administrator. can adapted visual estimations other ecological such winter bird food, pollen volume, insect infestation tree flowering/fruiting, adjustment classification threshold may show acceptable taxonomic differentiation presence–absence surveys.