The biodiversity data gap must be closed before we start talking about using artificial intelligence (AI) in conservation work, as we don’t have the infrastructure for training the machine-learning models yet, a data expert told Carbon Pulse.

When it comes to nature, it is important to boost our knowledge before we can develop reliable AI models, according to Raviv Turner, co-founder of the Nature Tech Collective and member of the Taskforce on Nature-related Financial Disclosures (TNFD) nature data working group.

“We need to collect the dots before we can connect the dots,” Turner told Carbon Pulse in an interview. “I don’t know any AI models that can run on missing, sparse, and outdated data.”

AI is increasingly attracting the attention of the emerging nature market since AI models are seen as a powerful tool to measure biodiversity uplifts and deliver biodiversity credits.

Industry experts recently said that developments in AI are speeding up the time it takes to process large amounts of scientific data and could potentially help scientists make better decisions about how to preserve biodiversity.

In April, the Bezos Earth Fund launched $100 million in grants to fund AI-powered solutions to climate change and nature loss, demonstrating the increasing interest in AI among investors.

However, Turner, who has been working with AI for two decades, defines himself as “sceptical” about its application to biodiversity.

“The biodiversity space isn’t ready for AI yet, since there is an entire nature data infrastructure that needs to be built first,” he said.

“We need grants for tagging and building training datasets, listening to bird recordings, mapping DNA species, flying pricey light detection and ranging (Lidar) drones to 3D scan sensitive biodiversity hotspots.”

Lack of biodiversity data is also slowing down the development of biodiversity credit market, he noted, since patchy information exposes companies to greenwashing risks.

According to a study led by Yingtong Zhu of the National University of Singapore, commonly used tools to measure corporate impacts on biodiversity require urgent updates as they do not align well with the Kunming-Montreal Global Biodiversity Framework (GBF).

In light of that, the Nature Tech Collective is leading a coalition of researchers and NGOs, including Conservation International, Wildlabs, One Earth, and Conservation X Labs, to carry out a gap analysis on biodiversity data.

The coalition will present these recommendations during COP16 in Cali, Colombia later this year, and is planning a call to arms.

“Many of the philanthropic funds just want to go and support projects related to megafauna conservation because it has more appeal,” Turner said.

“But what about the invertebrates, the below-ground biodiversity, the fungi? There is so much more work to do which is important for both nature and human health.”

MAIN PROBLEMS

According to Turner, effective machine-learning models for carbon emissions are already in place, allowing scientists to analyse carbon forests and comprehend canopy height and density from space.

“It happens because we have enough data on forests from space that can train a machine,” he said. “We just don’t have those for biodiversity.”

While we cannot see biodiversity from space, there are interesting projects that are currently working on collecting biodiversity images through smart camera traps, creating images that AI models can process.

“For example, AI can filter through many hours of recording and noise in the jungle and auto tag apex predators, such as jaguars, to monitor their presence,” he said.

However, this method faces two key challenges. The first is that those cameras run on lithium batteries, which, due to the high humidity in the jungle, start to leak at some point.

“I’ve been to installations where people didn’t have immediate access to spare parts and had to travel many miles to go buy new batteries,” he said. “It is also incredibly difficult to install those cameras in hostile natural habitats.”

The other issue is related to automatic tags.

Taxonomists have named approximately 1.7 milion species so far, while a study in 2011 predicted there are some 8.7 mln species on Earth.

Since we have relatively little knowledge of the existing species, we cannot rely on automatic tags when we explore new habitats, according to Turner.

This is the case with the bio-acoustic technologies that leverage data collected by audio sensors. Placing small acoustic devices across isolated points in a landscape allows ecologists to record the multitude of sounds of the many animals living there, including birds.

But while there is an increasing interest in this technology as a tool to monitor and conserve already-known species, it becomes complicated when it comes to new ones.

“For instance, the problem is that we have a really small percentage of birds already tagged,” Turner said.

“If you’re working in a new area where you have never listened to birds before, you don’t have enough data to train the machine. And this is just an example where AI really doesn’t work. ”

The same issues occur with the environmental DNA (eDNA) collection. Firms such as NatureMetrics are making cutting-edge efforts to implement non-invasive biodiversity monitoring, but without the preliminary knowledge of species and indexes, AI models are useless, Turner said.

“If we don’t have the eDNA dataset mapped, this technology can just tell you the taxonomy of the family,” he said.

Examples of generative machine-learning models, such as Chat GTP, are trained on a huge amount of data across many years – and still have unresolved issues and problematic biases.

“The more training data we have, the higher the level of confidence, the better we can trust the result,” Turner said.

“We just don’t have that training data sets for biodiversity.”

FILL THE GAP

According to figures reported during this week’s NatCap symposium at Stanford University by Christopher J. Schell, a professor at the University of Berkley, countries with the least amount of biodiversity data are the hardest hit by the nature crisis.

While countries with higher incomes generally have greater biodiversity observations, the ones with some of the most severe biodiversity losses are lagging behind.

According to the 2020 Living Planet report, Latin America and the Caribbean experienced the largest decrease in biodiversity at 94%, followed by Africa at 65%. Europe and North America had regional declines of around 24% and 33%, respectively.

“Nature is so much more complicated than carbon because it is local,” Turner said. “This means that we probably need to start collecting more data on the sensitive biodiversity hotspots first, and these are in some of the poorest countries.”

“Countries [in] the Amazon Basin and the Congo Basin … can’t afford to collect those data because it’s expensive, and in those places we are flying blind.”

According to a separate report by the UN Environment Programme’s World Conservation Monitoring Centre, most non-EU countries assessed a lack of resources to produce the biodiversity indicators required under the GBF.

Meanwhile, Turner said the first step to closing those gaps is working with Indigenous Peoples and local communities, training them on some of those technologies, and involving them in conservation projects aimed at collecting data we don’t have yet.

Initiatives such as Savimbo’s work in the Colombian Amazon and Fundacion Pachamama’s efforts in Ecuador’s rainforest are currently running preservation projects in partnership with or led by Indigenous Peoples.

Also, Turner advised it is essential to involve nature engineers as well as data analysts and scientists who can understand and clean the information collected in those efforts and tag them properly.

“I’d love to see more of the NGOs and more of the different philanthropic funds just funding biodiversity data, bounties, and initiatives of this kind before we can go back and talk about AI,” he said.

By Giada Ferraglioni – giada@carbon-pulse.com

*** Click here to sign up to our twice-weekly biodiversity newsletter ***