The Nature of AI: Solving the Planet's Data Gap with Drew Purves

0:08Sometimes you think that the real change can come in the long run from these moments of awakening where people almost overnight can change their relationship with nature. If AI can help to empower that, that might in the long run be the most powerful role of AI.

0:26Welcome back to Google Deep Minds, the podcast. I'm Professor Hannah Fry. Now, we focus a lot in this podcast on AI and its interactions with humans. But there is another story that's unfolding, one in which AI could help protect our planet. Think oceans and forests and deserts and fragile ecosystems and the millions of animal species that are crawling, swimming and flying across this earth. AI has the potential to tackle one of the biggest challenges of our time, the damage to nature and ecosystems. But with a problem that's this vast, where do you even begin? Well, Drew Purvis, Nature Lead at Google DeepMind, is my guide through the rich terrain of AI for nature.

1:09He has got two decades of experience in ecological research and has been at Google DeepMind for almost 10 years now. Drew, thank you so much for joining me. Oh, thanks. It's great to be here. I think most people probably agree by now that the environment is an important aspect of our future, something that deserves preserving and looking after. What's holding us back in this area? What's making it a difficult problem? Well, the short answer to that often is lack of information. So you're right, you know, there's this groundswell now of agreement about the importance of biodiversity, ecosystems and nature.

1:43And then we have signs of action in different sectors, in the private sector, in the government sector. You know, if you look at some of the numbers, for example, I mean, there are 189 countries around the world that have signed up to the 30 by 30 plan, which is to protect 30 % of ecosystems on land and in the oceans by the year 2030. There's so much to feel good about. But often when it comes down to actually taking action on the ground to either protect or restore biodiversity and ecosystems, it's just a lack of basic information. So what are the big questions that AI could help answer in this space then?

2:14I mean, there are a number of them, obviously. But for example, if you're thinking about a protection scenario, protecting biodiversity, you need to know where the biodiversity is. It might be focal species or it might be biodiversity hotspots and so on. It might be a particular endemic species. If it's a restoration scenario, you want to find particular places around the world that have the highest potential for ecological restoration. But, for example, you might need to know which species of trees to plant or which species of animals to reintroduce. So down at this local level, it's because biodiversity is so place specific, you need that locally relevant information for communities or whatever local action is happening to guide the action down there.

2:52And often it's just missing. Is that the big goal then of Google DeepMind using AI in nature to really fill in the information gaps in the same way that we have done for the human world? Well, here at Google DeepMind, we're growing a portfolio of work around AI for nature. And there are at least three key categories of AI for nature to explore. So the first of those is AI for data. And that can mean bringing in data from the field, from things like cameras, microphones and so on, or identifying and bringing in data from the literature. because there's a huge amount of data there. The second category, though, is taking all of that data and combining it with lots of other sources of data, like satellite data and so on, to create that derived information that decision makers need to protect or enhance nature.

3:35And then the third category, which is easy to miss, is that all the information in the world is fine, but human beings can get overwhelmed by that amount of data. So then there's this key role for active deployment of AI to help decision making, to help people make sense of all that data and act. And that's, I guess, in part because, you know, This is sort of quite a new area, right? I mean, this hasn't been around for as long as some of the other applications of AI. It's not as obvious of how to sort of solve biodiversity as it might be to solve disease, for instance. I think that's true. It does feel like a growth area in AI.

4:08It's kind of interesting, actually, that ecology obviously has a long history. But interestingly, mathematical ecology and statistical ecology and even machine learning ecology has a long history, too. Fisher, for instance, the famous fisherian statistics, a lot of that was developed in an ecological context. I think that's because ecology is inherently quite challenging. There's no obvious way to sort of record and frame ecology. You've got a vast variety of species. You've got processes operating at different scales. The temporal signals are always really noisy. And so it's interesting. It's always been a little bit of a challenging domain that has led to innovation in statistics and machine learning.

4:41And so we very much expect that in the work we're doing now, trying to apply this deep learning, for instance, the recent AI revolution to ecology that also will be a way to innovate. OK, well, let's dig into some of this stuff then. So in particular, in that middle layer, right, of like taking in data and sort of deriving something from it, because, I mean, a lot of this involves mapping, right? Like sort of that fundamental stage. Tell us a little bit about mapping the biosphere. So many of the decisions in environment and ecology come down to place. And that's because in ecology and biodiversity, each place is so different, different species, different habitats and often different issues that you need to deal with you know whether it's agriculture or wildfires etc so that missing information i mentioned before most of the time it's missing geo information and so very simply you can think of that as like maps you're right and so we need to map habitats that's kind of the substrate over which ecology plays because substrate is largely defined by plants so habitats and plants are closely related so you know whether that's forest grasslands etc but of course then we want to map all the species ideally that's a very challenging problem just because there are millions of them and a lot of them are very small and can't be seen and so on.

5:49Of course there's maps for now but also you might want to understand historical change. Those baselines of change are very important and of course being able to project maps into the future which is a whole different challenge. So but all roads lead to maps in ecology. All roads lead to maps. Okay that's interesting. I'm sort of surprised that this doesn't exist already though. I mean can you not just get this from Google Earth? It's a really interesting question because we've got so used to it. Of course it's been amazing the geographic information we have through things like Google Earth and Google Maps.

6:16You know, real human achievement to bring all that together and make it available. But it's very human centric so far, that data. So of course, we've got roads and you've got every shopping centre, every post office, everything like that. But actually, it's much less developed in terms of mapping the natural world. It's focused more on people than the environment. Yeah, that's right. And a lot of the information that we want about the natural world has never been recorded like that on paper necessarily. You know, where all the different species of trees are or where the habitats are or where the boundaries between you know the rough grassland and wetland are so we need to actually create that information for the first time it sort of seems a bit bonkers to me that we're in what 2025 and this hasn't been done already i mean there are attempts at this they exist i mean landsat was developed in the 1970s right of like taking satellite images and categorizing different regions i think there was some stuff before that for military purposes using satellite images but is this just a different scale to those things It amazes me too.

7:10Several things amaze me. One thing that amazes me is that the satellites even exist. I mean, that's incredible, right? It just blows my mind. And they've been actually for many decades. We've basically had these floating high-res digital cameras looking down at the Earth for decades. That's absolutely incredible. It's also incredible the amount of geographic information we actually do have. There are amazing data sets or amazing maps of all kinds of things that are relevant to the environment. And yet at the same time, sometimes you come in with the most basic questions like where are the forests?

7:37Do we not know where the forests are? That's right. And honestly, it surprises me too. We don't have gold standard accepted global maps for most of the habitats that we need to understand, including forests. So there is no absolutely universally accepted gold standard for forest, non-forest. Now, having said that, there are some pretty good ones nowadays appearing, but there's still room for improvement. On the other hand, what we're looking at is even if you had that to guide decision making, you often need to distinguish different kinds of forest. For example, the most basic split, natural forest from planted forest.

8:07And there definitely isn't a gold standard accepted map for that. And some of the work we're doing at the moment is trying to do our best to provide the best yet available map of that. We call it the Natural Forest of the World project. Why are forest maps useful? What can you use them for? Forests are incredibly important, harbouring biodiversity and carbon. And for that reason, there's a lot of government policy and a lot of international regulation around forest. And forests are a major focus of concern for conservation groups. and in all cases they need the best maps they can so they can know where to protect if they're looking to protect forests where to potentially restore forests where to monitor for problems like forest diseases and on and on can i see it have you got it yes i have yeah i've got it right here on the laptop so firstly what we produced is a map for the whole globe at 10 meter resolution so each 10 meter pixel is classified and we give it a probability that it's natural forest and that's what we're viewing here.

9:02I should probably do a little audio description for the people that are listening rather than watching. We have essentially in a browser what looks superficially, well exactly like Google Earth, but overlaid on top are these teal coloured pixels that align but not exactly with what you can see to be trees from the satellite images. That's right. So you can illustrate that for instance in, and this is important but also difficult to tell the difference between natural forest and planted forest because the natural forest will typically be much higher in biodiversity and often carbon if it's somewhat old growth so this will be an area that you would tend to want to protect more and equally if you detect forest loss then the loss of natural forests would typically be much more concerning than for instance the loss of planted forests many of which are planted in order to cut them down like a timber so they can actually be sustainable but in this case in the southeast us for example there are a lot of loblolly pine plantations for example so these are a type of planted forest which is planted to give timber very quickly after about five years okay and so you can see there we've got a mix of areas that we classify as natural and areas that we classify as planted forests yeah and if you zoom in and out you know most of the time of course they're inaccuracies but we're over 90 accurate when we test it so you can see here as we fade up our predictions there then if we look at those remaining parts you go in you'll find these neat rows and that's telling us it's planted for so to be able to do this at scale yeah so okay i noticed in this example it's sort of natural forest not natural forest but I also noticed that you've got a confidence threshold there.

10:28So are you sort of classifying this with a probability rather than just saying yes it's natural forest? Yes we are and it's something we're quite passionate about really because there is a tendency if you see a map that is classified in a black and white like way to believe that and even if it's relatively accurate most of the time it's really important to be aware of the inaccuracies and so that's why we much prefer to present things as this uncertainty map. Now of course one thing you'd often do is to think okay I'll choose some threshold then which I will use for my purposes to distinguish forest and non-forest but depending what you're after you may choose different thresholds so if you were really looking to make sure you're protecting all of the remaining natural forest in an area you would set your confidence threshold it's actually setting it low so that you pick all of that up but of course you'll also pick up a whole load of planted forest but you sort of don't mind along the way whereas on the other hand if you had very limited resources let's say and you wanted to verify that the loss of natural forest had occurred and you can only afford to visit certain places you may put a very high threshold on it to make sure that you're really visiting the places where you're most confident about and so on so we're sort of providing that to downstream users in this uncertainty form we're open sourcing all the maps we're open sourcing actually the data and also the new models we developed open sourcing out to the community what it means for downstream users is that they can get to their own high quality remote sense maps with much less compute than before, lower data requirements, and all importantly, lower skills requirements as well.

11:56So if we do it right, something like that can really help to democratise remote sensing on the outside. In terms of what's going on behind the scenes, I mean, you're using a vision transformer, right? How does that work? Overall, we're doing is bringing in this massive satellite information. So these are enormous images for a start, the sheer number of pixels and then for each pixel it's not just rgb like it would be with an image but it's often many different bands infrared etc and those bands are going up through time then you've got multiple satellites each satellite is giving you records every few days you have a lot of missing data from things like clouds so you need to take this enormous amount of data and somehow crunch it all down to pull out the thing that you actually want to know like is it a forest and the model in the middle that we're using is this vision transformer model but i mean And transformers are sort of, people usually associate them with large language models, right?

12:47Like, you know, paying attention to different bits of sentences is more important than others. How does it feed in when it comes to maps? That's right. So transformers were mostly developed to work on language. And this idea of these attention heads and so on, they were then adapted into vision transformers to do similar things on images, attending to different parts of images. What we've done here is then expand that out to a special vision transformer that's set up to deal with the challenges of satellite data. So it's actually now a multimodal, temporal, spatial vision transformer. But it's still a form of vision transformer.

13:17It's a great story about how you can get this exchange of methods between different areas from language modelling into vision and in this case into remote sensing and out there into environmental policy. So it's paying, when it's working out, whether it's forest or not forest, it's paying attention, as it were, to sometimes the infrared data, sometimes the image, sometimes the sort of the topology of the area, that kind of thing. That's right. The transformer architecture is an extremely expressive architecture that gives the model a huge amount of freedom to choose what it attends to and how it then combines that information downstream into its prediction.

13:51And that's particularly valuable in an area like remote sensing, where you've got such a wide variety of modalities coming in and all the data sets are so huge. And so it's an area that really benefits from this kind of extra sort of expressivity and flexibility of the model. But that expressivity in the end comes down to forest, not forest. That's what's funny, right? All of that is crunched down into just that, in this case, that single map. Well, OK, I'm just picking up on something that you said in your answer there, because you were talking about how you have these satellite images over time.

14:20So then rather than just mapping where the forests are or aren't, does that mean that you can look at how the forests are changing? Yeah, that's a very important point. Yeah, indeed. So the interesting thing with remote sensing and satellite data is because the satellites have been up for years, If you can do this mapping, you know, say you take one year's worth of satellite data, say for this year, and we can see where the forests are this year, we can automatically redo that from the past to estimate the picture of change. And this is really important. There's a project called Global Forest Watch, which Google has supported before, that does this.

14:50And so you can each year the latest map comes out and you can look at the changes in forest cover through time. More recently, we've created this stack of maps that we call deforestation drivers. and what this is actually looking at is it's taking each unit of forest loss and categorising the cause of that loss in terms of logging, agricultural expansion and so on and so on. But we've applied that for the last 20 years so we have in collaboration with our colleagues at the World Resources Institute at WRI we've worked up this global map of the causes of deforestation for each year from the year 2000 to today.

15:24And what do you find when you do that? It's really interesting to look at these patterns. Go on show me show me I know you've got it I No, you've got it. I want to see it. So the first thing you can see when you look across the world is that there's forest loss everywhere. Yeah, quite a lot in Northern Europe. Yes, for example. So it's so easy to think that all of these problems are just occurring in the tropics or in the global south. Right, exactly. And the second thing, though, is the causes of the loss does change between different regions. So if you have a look at, you know, zoom in on Brazil here and South America, you can see, you know, again, you can go right in.

15:52But lots of permanent agriculture. That's right. Yeah. So this sort of darker yellow colour. And you'll find, you know, when you fade in and out, you'll often find that this is associated with clearings and so on. So you can see how that's working. I mean, you can literally see how it picked out the edges of fields effectively. That's right. Because all that fine-grained data is actually in the satellite signals, right? There have been previous maps, but this one is 10 times finer resolution. And that's what really enables you to pull out these local patterns. And that's where a lot of the decision-making happens, is at local scales.

16:19And so understanding things at local scales is really important. Putting all of this together, if you're sort of long into the future, right, Is there like a real-time aspect to this? Could you prevent illegal logging from happening before it did? It's a really good question because the work that we've been talking about up to now tends to be taking satellite data from an extended period to estimate how things were either right now or looking into the past. Retrospectively. Right. Whereas this idea of real-time alerts and then also longer-term future forecasting is some of the real growth areas for us.

16:52There are already actually existing organisations that are using satellites to try to pick up deforestation alerts live. So that actually exists. One of the challenges, though, is that in order to pick up most of the real deforestation, they've had to use a methodology that picks up a lot of false positives at the same time. So it's actually really challenging. So there's so-called deforestation alerts. It's an incredible achievement. But at the same time, it feels like that could be improved. So obviously, if we could get it down to the point where the alerts are much more accurate, then it would be much more easy for people to act on them.

17:23I guess this goes beyond just satellite data then. I mean, you have to start working in other types of data, wouldn't you? Yes. Well, I think so. I mean, the interesting thing is that you can produce a model of change and you can train that on past data and you can do that very scientifically and you do all of your held out data and your validation and all that clever machine learning stuff. The problem is, on the other hand, that the whole thing around climate change and environmental change, the future is different to the past. So how can we be confident that we can reapply that model into the future?

17:52Like this isn't like predicting the weather a week in advance. The whole underlying model is different. That's right. And when you look at, say, this process of deforestation that we've been talking about, I mean, that is a human driven process. And humans can change behaviour very, very quickly in responses to things like changes in the market. You know, perhaps the price of soybeans goes up and then that encourages more rapid deforestation. On the other hand, government regulation and policies or local regulation and policies and so on. So traditional machine learning approaches are going to really struggle to deal with that because they can only deal with numbers and they can only deal with the past.

18:25And I think that's one of the reasons why we're really excited about hybridizing, in a sense, the best of the scientific tradition of simulation modeling with the machine learning, data-driven approach to forecasting, with actually things like the large language models like Gemini, which potentially could pick up on, you know, changes in government policy, news, media, perhaps integrate together things like agricultural prices and so on. to help us at least to slightly adjust and identify some of the sort of uncertainty scenarios around those future predictions. If that's vegetation, though, if that's like trees and hedgerows and shrubs, what about, I mean, we also care quite a lot about the creatures that are living inside them.

19:02How do you bring that in? So habitats is one thing, but then most species are not visible from space and they don't form a habitat in that sense. Instead, it's all the insects, the fungi, the birds, etc. And so much of our concern around biodiversity and so much of the action around conservation is actually around those species. And so you look surprised earlier when you said there isn't an agreed map of forests. That's true. But that's also true for most of the species in the world. Now, again, I want to it's really important to realise everything we're doing in ecology and biodiversity builds on this huge foundation of organisations and hard work from the past.

19:36So there's an amazing organisation called the IUCN that does provide maps for 140 ,000 species worldwide. It's amazing. On the other hand, those maps are very coarse. They're driven by expert opinion. Those experts, they really know what they're talking about, but necessarily they're quite coarse maps. So 50 kilometre or so cells. And they're only refreshed about every 10 years. Oh, wow. Yeah, so there's an obvious potential role there for producing, in theory, much better maps using the latest in AI and remote sensing and et cetera. So we're sort of exploring that as well. Okay, but then how do you do it, though?

20:07I mean, because you haven't got the equivalent of satellites, you know, in space photographing insects. That's right. So you could view it as a sort of prediction problem. What you're saying is the signals I do have, like the satellite signals and other input data, can help me produce a sort of probabilistic estimate of how likely something would be there. And there's probably always quite a large residual uncertainty on that. So how do you go about doing that? Interestingly, at the absolutely coarsest level, it's kind of the same old, in a sense, machine learning approach. In other words, you take your training data and you take your input data like the satellites and you bring in a special model and you fit the model and you get the output and you evaluate it.

20:44You know, it's the classic stuff. But in this case, it's really challenging because the data we're bringing in mostly is citizen science data. So this is people out in the field noticing a bird. You've spotted something, yeah. That's right. There's an incredible platform called iNaturalist. It's on your phone now. And if you see something, you can take a snapshot of it and then you can upload that to iNaturalist. And iNaturalist actually have their own machine learning model so they can classify that image to species, which is fantastic. and this is now the main source of these on the ground observations of species globally is coming from this amazing phone app called iNaturalist so that's fantastic but actually that data is still very biased and patchy why so for example people typically are close to their home when they take a picture so where people go when they go they're typically on a nice day and also what they choose to look at and take pictures of so what you'll find is you'll get lots of pictures of brightly coloured birds close to cities on sunny days and what you don't get is many pictures of you know small brown mushrooms in remote places when it's raining and although i'm joking about that there's actually a much more serious element to this too which is when you look globally at the distribution of citizen science data there are major areas of the world with very little sub-saharan africa for example because actually i naturally depends on having a certain amount of leisure time really to do these sorts of activities yeah and also speed of internet connection and like quality of your smartphone camera right and this is a real irony then for two reasons in a way one is from the fundamental ecology if you weren't careful if you were naive about it you would conclude that all species like to live near to middle-class neighborhoods and on the edge of cities because that's where most the data is and then globally there's a real irony there which is we're lacking data in many of the places that have the most biodiversity in many of the places where the current threats to biodiversity are the greatest and where people's livelihoods depend most on biodiversity which is just kind of awful right and so the magic in a sense that we're looking for from deep learning in this space is to see if we can somehow generalize on this very patchy bias data that we do have to ecologically plausible distributions for for all of the species worldwide and that's a big challenge and at the moment i think you would say the jury's out as to whether it's possible but we are seeing some evidence but does the things that you've learned and that you now sort of know in adverted commas about the vegetation then feed into where you expect the species to be?

22:59I mean, does like one sort of help the other? That's a really good point that we know enough about ecology to know that actually things like climate and elevation and habitat are very strong predictors of species. So you're absolutely right. We've already talked about the fact that we can classify habitats from space down to a fine grain. So that itself, we would expect to be highly predictive of which species are there. So then bringing in all this observation data, then we can look at the overlap between these in the areas where we do have data, and hopefully pick up a general enough understanding to then be able to reapply that globally.

23:30There is another challenge here, which is you have these long tail distributions in ecology all over the place. You have a few common species and then lots of very rare ones. So we need to do the same thing across the taxonomic sort of tree of life. You know, that can we learn the key relationships for the more common species that are data rich, enabling us to make the best predictions we can for species actually fairly data poor. So, for example, if we bring in species trait data, we might find out that in this part of the tree of life, let's say, it's the larger bodied species that tend to live in the colder environments.

24:01And we might be able to discover laws like that in places where we do have lots of data, like Europe and the US, in a way one day where we can apply that with a certain level of confidence in areas where we have much less data, like the global south, you know. Not that we'd ever have a blind guess, but what it would really do is lower the data requirements in those areas. You know, once we've learned some of the more general rules, it would mean that actually a relatively small amount of observational data may be enough to sort of anchor the distribution of that species. Such an exciting space.

24:27So a few years ago, I think series two of this podcast, we spoke to some of your colleagues who are working with some data from Serengeti. On a surface level, some of their ambitions were similar to yours, which was to understand the types of species that were inhabiting that particular era. In fact, I have a little clip of Meredith Palmer, who's a conservation scientist, talking about this project. Have a listen to this. It blows my mind how incredible AI can be for solving these kinds of problems. I've had moments where I've looked at an image and had a computer identify a species I, you know, on first glance didn't even recognise was there.

25:05I do wonder about other types of data here. I mean, you're talking about data, about the landscape, about elevation, about the vegetation. but are there other elements of data like photographs for instance that you can also bring in here yes there are and that in general you could think of that as this trend towards multimodal models and multimodal ai and that particularly lights up i think in the world of ecology and biodiversity science because you've got this such a wide variety of organisms at different scales and so different modalities suit monitoring different kinds of species so they're talking there you've got images let's say humans take deliberately with a camera they're the ones that say iNaturalist specializes in.

25:43Often though say in the Serengeti projects and others we're talking about so-called camera traps these are cameras that are motion activated cameras and a lot of the time they've got infrared lights on them so they're actually taking infrared images at night so even images are you know there are different kinds of images you can take. You've then got aerial imagery let's say from drones or from planes you've got the remote sensing satellite imagery we're talking about then also you've got bioacoustics so-called bioacoustics is where you're deploying microphones in the field that might be familiar to a lot a lot of people for example around the apps that identify bird song merlin and ebird and similar because those those already exist i mean i've played with them a little bit right like you're outside you turn it on it tells you what birds you're listening to that's right but that's supervised learning is it that manages to do that yeah that's right so you need a sort of what we call a corpus of labeled data to start with so you need a picture with a label along with it you know this is a monarch butterfly or this is a crow crow squawking yeah exactly all of this sort of stuff and with enough of that you can then train a model that generalizes so that's what's called supervised learning that's right and that's powered most of these developments in each of the individual modalities you know whether that's images you know remote sensing etc yeah you've got a project though called perch right tell me about perch so perch is in this space of bioacoustics and the idea of deploying microphones afield is really attractive because there's almost always something making a sound and it's incredibly rich the information we can extract from sound actually and you can leave the microphones out for days on end to record all that data and of course sound is not necessarily just species but it can be behavior we might be able to distinguish juveniles from adults you might be able to pick up warning signals so insects make sounds reptiles make sounds you know birds make sound amphibians make sounds you know etc and you pick all of that up and you don't have to have line of sight to the organism either it can work at night so it's an incredibly promising technology for pulling out data from the field.

27:31But if it's already been done, I mean, if you do already have these apps that can tell you if you're listening to a crow, then what else is there to do? Perch is taking a slightly different approach. It's a foundational model for natural sound. And what that does is like other foundational models, it's not out of the box. It's not just providing you with the ability to identify lots of different birds, for example. I mean, it does do that. But what it's really designed to do is to allow people to rapidly create their own detectors for things that haven't been in the detector already. So, for instance, if you're working on particularly rare species, it won't be in one of these apps already.