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