This is pretty fun – researchers at Northwestern at Pitt installed cameras in rain gardens and used machine learning to analyze over 40,000 photos of pollinators visiting them. This also appears to be a successful interdisciplinary collaboration between civil and environmental engineering, electrical and computer engineering, and geology and environmental science academic departments, which I can tell you from personal experience can require overcoming surprising and substantial resistance. The article is not open access, sadly. This is why my next job may be happily cleaning toilets at a university, if it gets me gym, pool, and library access.
The effects of urbanization and climate change have resulted in a global decline in pollinators. Adaptive strategies such as green stormwater infrastructure (GSI), and rain gardens (RGs) specifically, have the potential to create habitat for pollinators in urban environments in addition to other benefits such as hydrological control. Although many studies detail the habitat provision of natural areas, few explore quantitatively the pollinator activity promoted by urban GSI and compare it to a native system. The aim of this proof-of-concept study is to demonstrate an innovative methodology to continuously track pollinator activity using camera traps and employ advanced data analytics to interrogate the videos and identify patterns in pollinator activity. We monitor pollinator activity diurnally across multiple urban RGs and pollinator gardens (PGs) in comparison to a restored prairie using 10 trail cameras with motion detection and IR spectra technology over a month (July 2023). We compare the pollinator activity at patches of a high Coefficient of Conservatism (C-value) native species (Echinacea purpurea; Asteracaeae) to that of a lower C-value native species (Asclepias syriaca; Apocynaceae). Data analysis including use of an unsupervised machine learning algorithm, suggest that RGs located near PGs with high C-value and high-density plant assemblages can promote orders of magnitude greater pollinator activity than isolated systems, or even, a restored prairie. Once acquired, field images were labeled for deep learning model detection. We trained a YOLOv8n model with 42,271 images over a course of 5 runs to detect pollinator classes. These results support the findings that urban landscapes can provide habitats to support biodiversity and offer valuable insights into design strategies that can potentially boost the pollination services of small, patchy GSI systems such as urban RGs when they are located with connectivity and in proximity to PGs. Finally, our trained YOLOv8n model ran with a total precision of 90%, total recall of 79%, and a total maP50 of 85% providing an accurate methodology and foundation to efficiently interrogate and quantify pollinator activity in the field.