Bees are important pollinators responsible for creating and maintaining the ecosystems that many animals and humans rely on. With many of its species on the decline, the Cheadle Center for Biodiversity and Ecological Restoration is leading the Big-Bee Project Network (http://big-bee.net) in an effort to digitize bee traits for further research in bee ecology. Since researchers lack data on anatomical traits that may make bees either vulnerable or resilient to human-induced environmental changes, it is important to find new methods for efficiently collecting these data.

One bee trait identified to prioritize was a bee’s hairiness because they’re important for pollination, sensing nearby danger, and staying warm in cooler temperatures and cooler in hotter climates. Through the use of deep learning and computer vision, a new pipeline was created to measure the surface density of hair from bee images. By optimizing two pre-trained U-Nets on lateral-view images of bees, the model became capable of pixel-based classification of where the bee is in a photo and where the hair is. These models were trained and evaluated on hand-crafted datasets of 550 images, which were remasked overtime for better performance.

While the main challenges came from differentiating texture of hair and texture of eyes/wings, another deep learning model succeeded at removing this noise before being fed into the hair detection pipeline. This multi-step pipeline improves previous findings by further preprocessing input images and is able to target the important signals of the image (hair). It also opens up the opportunity of using traditional edge detection methods for optimizing this problem.

This research is sponsored by the McNair/Edison STEM Summer Research Program and the National Science Foundation (#2102006).

The Cheadle Center for Biodiversity and Ecological Restoration (CCBER) is working with the UCSB Data Science capstone team to continue the research and understanding of bee ecology through image and trait digitalization.

In efforts to help researchers involved in the Big-Bee project, we want to automate the process it takes to handle and process the digital images of bees. Given images of bees with a QR/Datamatrix code, we sought to develop scripts that can rename the image files based on the decoded box and extract EXIF metadata. In the latter half of our project, we shift our focus toward the hairy characteristics of bees. Given data of high definition lateral bee images, we want to create models that can quantify how hairy a bee is. 

Results

Two scripts, the QR Code Scanner & MetaData Extractors, are published on GitHub and are currently used by two academic institutions. More trial testing is required to discover bugs and further improvements for the scripts. Using density as our metric, our U-net and transfer learning method yields an accuracy of 97.68% for bee masking and 87.2% for hair masking. Using entropy as our metric, we were able to compare the average entropy values for different bees binned in the following manner: low, middle, and high hairiness. The box-plot distribution of the bins and t-tests show that there was a positive correlation between bee hairiness and average entropy value. Furthermore, we discover that entropy analysis works best for specimens that have low reflectivity and minimal skin texture. 

QR Scanner: https://github.com/booleank/Bee-ScannerMetadata

Extractor: https://github.com/harperklauke/Metadata-Extractor