Project 01 – AI for Beekeeping – Initial Classifier
The Problem
To help with food security, we chose to start where most food does – with bees and beekeeping. An estimated 30% of human food directly relies on bees for pollination (and that is just the direct effect. Cows have to eat something too).
After reviewing the literature, we asked local beekeeping experts in West Michigan about the leading threat to their bees. Our research had shown us paper after paper on Colony Collapse, varroa mite population growth, and countless other pests threatening hives. We were certain the top culprit would be on our list.
We were wrong.
It was “the new guy”.
Sitting in a meeting with a local beekeeping group, we understood why. With all the potential invaders, conditions, and factors that can go right (but not too right) or wrong (but not too wrong), it’s amazing our little honey-making friends accomplish anything. In every meeting we have attended so far, someone holds up a phone with a picture, asking, “What is this?”
The wonderful part is, someone in the room knows. They’ve been keeping bees for 20years, 33 years, or even just three seasons, and they know exactly what to do.
This brings us to the “new guy”, the person out in the sticks all alone with their shiny new bee box and high hopes for honey. The one that doesn’t know, and doesn’t ask. The one that truly believes in their heart of hearts that everything is fine, when it isn’t. The one, quiet and alone, tending a corrupted and toxic hive, unknowingly toiling day and night to spread disease and corruption to every other hive in 25 miles.
He’s the problem.
How we can help
Don’t misunderstand me. Our solitary beekeeper isn’t a bad person; they simply lack the necessary knowledge. They need a guide—one they can consult for answers whenever questions arise. With out a doubt, a guide like AI could fill this role perfectly.
While Artificial Intelligence is an unparalleled source for knowledge propagation, its implementation isn’t straightforward. Before AI can help, someone must first teach the computer, a process that requires months of manual work and technical tasks, all performed by frightfully expensive experts. Such tasks could easily amount to nearly half a million dollars just to produce a dataset capable of seeing and classifying a hive’s normal contents, such as brood and honey. Only after this foundational work is completed can we even begin to address any “special” ailments.
But what a dataset! It would act as a bridge, allowing the human inspection our hives so dearly rely on at computer speed with a hard-drive’s memory. The ability to truly see, sort, and store the beekeeper’s observations indefinitely opens up so many possibilities. Building AI for beekeepers would have transformational guidance at their fingertips: How much honey, exactly, was there? Were there more or fewer brood than last time? Will the bees make it through the winter? Is the queen okay? How does this compare to last year, or the year before? Was the full super in the second hive, or the third?
How we are going to do it
Using AI for beekeeping is a daunting task, but with such great benefits, we have decided on this dataset as our pilot program. Once completed, we will release it here on our website, free to the world. With this one project, we can help every apiary research group and citizen scientist catch up on years of foundational work, a step many of them lack the manpower, funding, or organizational scope to do on their own. Once complete, it becomes the jumping-off point for all sorts of other projects.
AIs differ from other programs. They are not “coded”, they are “grown”. We feed them data and expertly prune them until they achieve a final product capable of amazing things. To achieve this with complex visual data like pictures, we will need a great deal of data for any respectable level of accuracy. Additionally, holding to the highest ethical standards, we will source the pictures ourselves. We won’t take any data from others for our use.
With a target of 100,000 fully annotated pictures, a “pipeline”—a block of code that automates a step of the process—is the most economical approach. A fully manual process invites human error, takes forever, and to be honest, with the critical state of bees today, we don’t have the time.