2017 ESA Annual Meeting (August 6 -- 11)

PS 45-171 - Finding flowers fast: Using a drone and machine learning to create bee foraging maps

Wednesday, August 9, 2017
Exhibit Hall, Oregon Convention Center
Cassandra Burgess1, Kangni Wang2, Yekaterina Kharitonova2 and Matina Donaldson-Matasci3, (1)Department of Engineering, Harvey Mudd College, Claremont, CA, (2)Department of Computer Science, Harvey Mudd College, Claremont, CA, (3)Department of Biology, Harvey Mudd College, Claremont, CA
Background/Question/Methods

Honey bees, key pollinators in many agricultural and natural ecosystems, forage on an enormous scale. Each hive can have over 10,000 bees, and foragers will travel up to 10 km from their hive in search of food if necessary. Poor nutrition, due to reduced diversity and abundance of floral resources, is thought to be a driving factor in recent honey bee declines. However, the huge foraging range of honey bees limits the ability of researchers to track those floral resources using conventional time-intensive ground surveys. With such limited information on the resources available, it is difficult to make direct inferences about the impact of floral diversity and abundance on honey bee foraging and health. Utilizing recent advancements in machine learning and recreational drone technology, we have developed an automated system to map flower species and location over large areas. Maps are created by flying a drone in a grid pattern over the area of interest while taking photographs every 2 seconds. These photographs are then stitched together to generate a map. An algorithm based on computer vision and machine learning was developed to locate flowers within those images and classify them by species.

Results/Conclusions

A k-nearest neighbors algorithm was used to distinguish flowers from their surroundings throughout the map. These flowers were then classified as one of four species using a random forest algorithm. Testing of the algorithm was conducted by determining precision and recall for each species using manually labeled images. Using this approach, the algorithm was able to distinguish flowers from non-flowers 100% of the time (precision=1.0, recall=1.0). It consistently identified both Penstemon spectabilis (precision=0.89, recall=0.82) and Acmispon glaber (precision=0.96, recall=0.99), and also identified Marrubium vulgare (precision=0.85, recall=0.58) and Salvia apiana (precision=0.75, recall=0.89) much of the time. This promising methodology for mapping flower locations will allow researchers to obtain larger-scale and more accurate representations of the resources available to a honey bee colony. With detailed maps of resources of interest to bees, researchers will gain insight into how the floral resource environment can shape the foraging behavior of honey bees and other pollinators, and how it impacts their nutrition and health.