Learning by doing: applying artificial intelligence to biodiversity with BugBox
- Dr. Kelton Welch
- 5d
- 4 min read
By: Dr. Kelton Welch
11/28/2025
When conducting an analysis in the natural world, it’s important to keep your focus on the downstream task. As scientists, we all feel the pressure to endlessly hone and optimize our tools and methods. Method optimization is important, but it should not come at the expense of completing the research work we have to do. Our primary mission as ecologists is to provide society with much-needed data on biodiversity and the environment, so we can better manage natural resources and minimize negative environmental impacts. So it behoves us to maintain our focus on collecting that data even as we work to develop a new tool, like artificial intelligence (AI).
AI holds enormous promise for conducting biodiversity experiments at scale, especially with the rapid increase in available expertise and resources for model development. However, even with all these resources, there is still a great deal of groundwork that needs to be laid before AI has the subject-matter mastery to correctly classify even a fraction of the world’s biodiversity. We don’t have the time to wait for all this groundwork to be laid: the time for biodiversity science to make a difference in the world is now. Furthermore, even if we could lay all that groundwork, I’m not sure we should: that groundwork can come with its own major environmental costs, after all. So, there are plenty of reasons to expect any AI tools we use to be deficient in some ways. But that’s okay, because the entire history of science is a story of inadequate people doing great things with inadequate tools and inadequate data. We shouldn’t be afraid of that, but we should be aware of it, even as we keep our eyes on our downstream tasks.
At Ecdysis Foundation, our downstream task is our 1000 Farms Initiative, which involves providing environmental and agronomic data for farmers across North America, and hopefully demonstrating how farmers are using regenerative practices to protect biodiversity while growing nutritious food profitably. If we want to include entomology data with all the other data streams, we need a way to process it as fast as we process the other data streams. AI offered us that. So, we built a web platform called BugBox to rapidly identify arthropod specimens using AI. We trained our model on our own scientific samples, which means it learned the most relevant species from North American agroecosystems.
We needed to know how reliably our model could represent arthropod diversity. We knew that it wouldn’t be able to accurately identify every species in these fields, but we predicted that the diversity metrics it generated would still uncover useful trends. We set aside some of our entomological data for an evaluation of our AI, and we used me as a point of comparison for the AI’s performance. I reviewed every specimen BugBox identified for six different datasets in different cropping systems, and we retrained BugBox repeatedly on the corrected identifications. As we accumulated training data, the AI’s accuracy improved from 36% on the first dataset to 64% on the sixth one. But each time we calculated a biodiversity index, the same linear patterns emerged: the AI consistently found high diversity in the same samples where I found high diversity. So the AI was consistently telling the same story as I was, even though it wasn’t as accurate as I was.
We were intentionally naïve about the way we conducted this work: we simply chose a recommended AI model (MetaFormer), trained it on some of our reference images using some recommended settings, and started uploading batches of specimen images. We learned as we went by keeping our eyes on our downstream task, and now we have a much better idea of how to use AI and what it can do for us. Granted, biodiversity indices offer only a coarse way to measure biodiversity; but coarse analysis on a large scale can move the needle for policymakers and growers looking to support and implement regenerative practices. Without AI, the entomology wouldn’t have had a voice in the big conversations we’re able to have with these groups. And now that we’ve laid the foundation, every new dataset we collect contributes to biodiversity science while helping us optimize our tools.
I find this study reassuring, because I routinely feel inadequate as a scientist, but this work gives me hope that we can use AI tools for good purposes, even if the models — and our skills with them — are inadequate. My co-authors and I hope AI can have a positive role to play in the future of sustainability and biodiversity science, and BugBox is a manifestation of that hope. We invite other scientists to try BugBox out: contact us if you are interested.
Author Bio:
Kelton Welch is the (inadequate) taxonomist and collection manager at Ecdysis Foundation, who finds it difficult to look up from a computer screen full of insect photos to remember to go eat his lunch.
Links:
Images:
Two images provided: a collage of BugBox specimen photos for the backdrop, and a map of the 1700 sites visited for the 1000 Farms Initiative through 2025, a subset of which were included in this study.
Map Caption: The 1700 sites visited for the 1000 Farms initiative 2022-2025. Artificial intelligence is essential for processing data at this scale.
Image Credits (both): Ecdysis Foundation
Comments