Sleuthing Plant Species

Lauren Gillespie, PhD ’25, remembers the day growing up near Fort Worth, Texas, when the historic ranch next door was bulldozed for development. 

“I swear, every single living thing had been bulldozed,” she says, from the creek where she and her siblings would catch minnows and bother frogs, to the Ashe junipers and Texas live oaks. “Every single square inch of top soil had been removed, every patch of grass, everything truly.”

She recognizes it as a pivotal moment in her life. “That was a large motivator for me, because I saw land use change right before my eyes.” 

“It’s great that maybe a hundred families got to have their picket fence and one-story house, but what about all the life that was there before? Where had it gone? It was in a pile rotting under the Texas sun.” 

As a recipient of a TomKat Graduate Fellowship for Translational Research, Gillespie has been determined to develop better tools for estimating biodiversity in ecosystems—of capturing a snapshot of what’s growing in a given place to enable new ways of protecting them from extinction. 

Biological diversity, or biodiversity, is a term that encompasses the variety of life from the soil microbiome, to fish in the sea, to the largest redwoods on California’s coast. Gillespie is looking specifically at plant biodiversity, which can be a proxy for the resilience of an ecosystem, as well as a landscape’s ability to sequester carbon dioxide. 

Her research is a blend of machine learning predictions calculated from millions of images and data points, and field research that is counted hour by sun-scorched hour. With one advisor in a computer science lab and another in a plant biology lab, she is tempering what’s next in artificial intelligence with what’s trusted by ecologists today. 

As a daughter of Texas, where over 93 percent of land is held privately [1] and people embrace the slogan Don’t mess with Texas, from the outset Gillespie has searched for a way to cast measurements from the sky.

In 2024, she published her findings in Proceedings of the National Academy of Sciences. By combining satellite and aerial imagery with the trove of citizen science data available in California through apps like iNaturalist, she and her collaborators demonstrated it’s possible to train a neural network model to outperform traditional approaches, allowing them to categorize plant communities with high accuracy and to map species within a few meters' resolution. 

Using Deepbiosphere, the name of their deep learning model, it turns out that changing landscapes have distinct signatures that can be detected and coded, identifying large losses such as those from severe wildfires and lumber harvesting to less perceptible alterations in habitat fragmentation and the countless subtler shifts in flora.

In the future, her research could enable global monitoring tools to automatically track plant biodiversity from satellites—allowing governments to pinpoint not just where, for example, rainforest is being logged illegally, but what species are being lost in the process.

Here in the Golden State, the applications are immediately useful. After completing her doctorate this year, Gillespie will become a postdoctoral scholar at MIT and team up with the California Department of Fish and Wildlife, among other entities, to advance California’s goal of protecting 30 percent of the state’s landmass by 2030 for conservation and climate resilience.

A decade ago, botanists and cartographers began detailing a map of California native plant species for this purpose, a project that will take another 5 years to complete, at least. 

What artificial intelligence provides is speed.

“We do not have time to wait for the higher-quality maps,” says Gillespie, with urgency in her voice at the dual specters of land use change and global warming. What ecosystems might be lost between now and the year 2030?

Yet for California legislators to use Deepbiosphere to identify and protect vulnerable habitats, the model must first be proven consistently accurate. That is Gillespie’s current ambition: To ground-truth the algorithm’s predictions of satellite data, with her own hands and feet when necessary. 

Brazil-bound 

Last February, Gillespie boarded a plane for the Universidade Federal de Minas Gerais, 270 miles inland from Rio de Janeiro. She had won a Fulbright Scholarship to study alongside Danilo M. Neves, a famous Brazilian ecologist. What better place to further her research than in Brazil, the most biodiverse country on the planet?

Notably, Brazil also has among the strongest environmental protections written into its constitution[2] , although legal enforcement can fall short of those ideals. And as the fifth-largest country in the world by land mass, with 3.286 million square miles, Brazil has a lot of ground to cover.

“They’re still discovering new bird and frog and plant species every year,” says Gillespie.

Originally, she had envisioned developing a version of Deepbiosphere bespoke for Brazilian habitat, but she soon discovered that the country has far less citizen science data than does California. 

“Because you need resources. People need a smartphone to go and take these photos and so Brazil has 0.5 percent the data density of California,” she says. “Upon realizing that, we decided we can’t realistically train this model from scratch in Brazil. But what if we could use some cutting-edge machine learning tools to take the California model and tweak it just slightly for Brazil with what little data we do have?”

Ten months later, Gillespie and her Brazilian counterparts had developed a protocol tailored for the Cerrado savannah. All told, they cataloged eight sites from two states and five municipalities, documenting over 1,000 trees, almost 100 soil samples, and hectares of drone footage. 

“Now, being from computer science land, where we are used to talking about millions of images, eight [sites] does not feel like a lot,” she says. “But considering the amount of data availability otherwise, it is still a significant amount of data captured.”

Her time at the Universidade Federal de Minas Gerais is proving fruitful for her Brazilian labmates, too. Colleague Kleiperry Freitas will use these soil samples to understand how soil diversity has helped shape the Cerrado’s rich multitude of trees. Meanwhile, labmate Catherine Rios Santos recently won a prestigious scholarship to travel to Germany and train machine learning models to identify individual trees automatically from the drone data they collected together—estimating in minutes what would normally take an entire day in the field. 

“I’m very thankful for TomKat’s support,” says Gillespie, noting how the center provided funding to support her endeavors in Brazil, as well as take a specialized botany course in South Africa. 

“They’ve been able to help this research flourish.”

Biodiversity credits

The Anthropocene Epoch is the widely referenced, if unofficial, unit of geologic time describing humanity’s impact on the climate and ecosystems. According to the United Nations, human activity has already altered more than 70 percent of Earth’s ice-free land.[3]

Some of these changes are obvious and intentional: The clearing of prairie to build tract homes that Gillespie witnessed as a teenager. Some are insidious, accidental: The blue live oak that died in front of her house from a tree fungal disease that spreads faster in warmer winters. 

“Me and my siblings, the three of us together with our arms outstretched couldn’t encircle the entire trunk because it was so big.”

As she has mapped the landscapes of North and South America, Gillespie wants technology to be a fulcrum for positive change. Learning about how carbon credits operate from Stanford’s climate tech community, she wonders now if economies could take a similar approach for rewarding biodiversity and issue credits for its preservation.

After all, the carbon offset market is a way of valuing nature as a service, rather than a commodity. She hopes one day her research could serve a similar purpose: That biodiversity credits could change the bounty hunt, by making nature worth more alive than dead. 

“We’re here hopefully to make it better for the next generation,” she says. 

This article is part of the TomKat Center Spotlight series designed to highlight the impact and trajectory of the work of faculty and students who received funding through our Innovation Transfer Program, TomKat Solutions, and Graduate Fellowships. Stanford University does not endorse any non-Stanford entities, programs, products, or services listed in the article. 

[1]Source: Texas Parks and Wildlife Department 
[2]Source: Brazilian Constitution 
[3]Source: United Nations