Return to page

AI for Environmental and Social Good: From Pixels to Science to Conservation by Dan Rubenstein

 

 

 

Read the Full Transcript

 

Sri Ambati:

 

It's a great privilege to invite Dan Rubenstein to the stage. Dan has done incredible work over the decades preserving biodiversity, understanding wildlife, endangered species, probably the leading expert in Grevy's Zebras and amongst other incredible species over the years. Dan and Tanya, Tanya are co-founders and Jason of a Wild Me Wild Me, how we met about seven, eight years ago. Leland brought Tanya into my office because we are on the poor side of the highway from Google in Mountain View. So Tanya was there. Visiting fundraising from Google did not work that well. When she came into my office in the next 60 minutes, we cut a check for her that was probably much larger than what she was expecting from Google. 

 

Introduction Of Dan Rubenstein

 

And turns out that was a first check for that project. Years later, we are here to say it's the UNICEF AI Project of the year for data for wildlife conservation. We saw Tanya there present and immediately get bombarded by three more talks to talk to the NSW and the museum of art. I think, suffice to say the Wild Me project has almost democratized how you do conservation and bring biodiversity. Of course, our job was to just give them the initial start, all the hard work done is in the field.  Now, Dan runs a great course in Princeton. Of course, I was hoping to partner with him even personally and do some more work to bring AI to really create, give rights to the people without rights or the animals, without rights. 

 

That's how we see AI for conservation is where we can amplify and protect the rights of. I've seen Dan talking to me earlier on how the government of Kenya is now even trying to have a good discussion with conservationists while they're doing urban planning. I think the animal rights movements and the wildlife movement started in the early days of when Jane and Dan and others were fighting for bringing rights to the species. I think today it has grown much more and hopefully in the course of the stock, we'll see some of that and we'll walk a mile in Dan's shoes. Without further ado, let me welcome Dan to the stage. Thank you. 

 

AI Becomes Involved In Conservation

 

Dan Rubenstein:

 

Thank you very much Sri for that wonderful introduction. It really is truly wonderful being here listening to what we've listened to this morning about AI and how it's organizing data. I particularly like the life cycle picture forAT&T where AI comes into different places. So as I'm talking, think about that life cycle because I'm going to call on all of you at the end of my talk to talk about how you and AI can actually be involved with people and conservation. You don't directly have to deal with protecting the wildlife, but you do have to affect the decision making of the people. And that's about the information flow that you were just talking about. If AT&T wants to be the best at providing information and connecting people to people, that will change the use of the landscape, which will have indirect huge impacts on conserving biodiversity ecosystem services and wildlife. 

 

So think about that as I'm talking and we'll come back to that at the end. I'd like to talk today about AI and environmental and social good and we'll talk about how we use images pixels to do science and to do conservation. And as Sri said, I work with a team, I'm the biologist in this picture. The others are all computer scientists, whether they work on data, whether they work on computer vision, they all develop algorithms. That's where Sri and Tanya got together and started talking about how H2O.ai can both contribute to the science and to the financial stability of Wild Me and Wildbook, which Sri very kindly did and stabilized the company as it was starting to have its huge impact. The state of the planet's biodiversity can be looked at through the Living Planet report in 2018. And the theme was aiming higher. 

 

The State Of The Planet's Biodiversity

 

The reason for that is populations of mammals, birds, fish, reptiles, amphibians on average have declined by 60% between 1970 and 2014, the most recent year when the data was compiled. That's a huge amount of species and population sizes shrinking. The earth is estimated to have lost about 50% of its shallow water corals in the last 30 years. And 20% of the Amazon has disappeared in 50 years and it's accelerated in the last few years. Under the previous president. Globally, nature provides services worth about 125 trillion dollars a year while also helping ensure the supply of fresh air, clean water, food, energy, medicines and much more. And yesterday was clean water day and as you know a few days ago we put the 8 billionth person on the planet. The problems are escalating and the question is how do we turn the corner? What I'm going to do is talk about how our research fundamental science by working with people and bringing them into the process, which democratizes science, will transform the landscape because they take ownership of their behavior if they believe that they're having an impact. 

 

Data Deficiencies

 

The problem is that despite all of these problems I just listed, we often don't even have basic data. So we can see in certain parts of Africa, elephant numbers are known, they go up, they go down. But in certain parts of Africa, central Africa, we don't even know how many elephants there are and we don't know the history of the population cycles at all. So we don't know whether that decline is an aberration from a peak and it will turn around or it's going to cause an extinction of a whole region of elephants. We're just data deficient. Killer whales, killer whales are all over the Pacific Northwest. How many are there? How do you make a decision whether the different populations are to be treated differently or the same. Some hunt differently. Knowledge transfers through the societies very differently. That's my area of specialty, which is animal behavior. 

 

But if you're data deficient, how do you get the public to invest money as taxpayers through their governments to lead to policy changes? And so we were hearing at the very end of the last talk about the AT&T insights in AI helping government organizations make changes. That's terrific, but a lot of times we don't have the data to even base the analysis on. And when you do have the data and you can say a species like the polar bear is vulnerable because the ice is melting and they can't get around and they can't hunt and they can't raise their babies, the problem is we don't know the population trend because our data is spotty over time and over space. And so the Red Data List will say it's a vulnerable species, but it could be endangered because we don't know what the cycles really were and we can't forecast what the likely population will be as climate change continues. So these are the problems that we face as biologists working on a landscape. Then when there is data, how good is it? So the whale shark, which is what got Wild Me started, you can see I've highlighted here the effective population size, which is the number of genetically different individuals in the population. So it's always smaller than the census. Population size is 103,572. That's a pretty good number, but it's listed as endangered. But look at the error bars, 70% variance.

 

You, as a government official spending taxpayers money, would you spend money to preserve a species where you don't even know how many are there? That number of 103,000 could be as low as 27,000 as high as 180,000. These are the decision makers who don't trust these numbers. And we often say follow the science. Let's use evidence-based solutions. Well, if the evidence is weak and people don't believe it in their gut, they're not going to use it. And so stasis occurs, no change occurs, no investment occurs. So this is the problem about conservation. It's about information quality and access and how it fits with people's internal beliefs, goals and desires. So what about the population dynamics? We can count those numbers. So we can talk about how fast does the elephant population of Africa decline or is the snow leopard numbers stable? There's very few. We don't know because we don't have a good time series. 

 

But we can ask other questions. If we know individuals, we can ask how far do whales travel? Because we can follow the fates of individuals and do all whales behave in the same way? And if they do, that provides one type of option in management. If they don't, then your decision making on how to intervene is going to be more complicated  and more difficult to get people to agree on how to behave. How many turtle hatchlings survive? We know as you read in the press, that hatchlings come up. If there's city lights behind the beaches in Florida, they go the wrong way. So many of them die at that point. So you want to get them into the water. But it turns out that most of them die anyway in the water. It's coming back to the land with the lay the next generation of eggs. 

 

That's critical. And that's the problem of fisheries with their nets. And that's why turtle excluding devices are now required for shrimpers so that the turtle, if it gets in there, flips the net and can escape before it suffocates. If you ask any fishermen, what do you think of that? They say, Hey, tell me if you had a hole in your net, would you like it? Okay, say I don't like it. But they're willing to go with it because they're internal clock says that preserving the species is good. 

 

Where Do We Get The Right Data

 

So we can ask really important questions, but where do we get the right data? So for example, Vulcan paid 8 million over two years to find out how many elephants there were in Africa. And you can see that in Central Africa the actual counts versus the probable numbers are wildly different. Whereas in the rest of the world in Africa, the numbers are actually pretty good, but they had to do it over two years because you wanted to see change and you could see change in the right hand panel that in Kenya and northern Tanzania, they're green, the populations are slightly increasing, yay. 

 

But look at central Tanzania and further south in East Africa, the red, the populations are declining. So there's a huge dynamic. Why the difference? But it took 8 million dollars to get those answers. Is there a better way? That's not sustainable. You're not going to get NGOs providing that kind of money for every endangered species. And the cost can also be reputational more than financial. You see this elephant here has a GPS collar on. The GPS collar is very important because for a small number of individuals you can get very fine grain data on their movement patterns, habitat choice and interaction with people to look at human wildlife conflict. But that is an invasive strategy. You're having to put the animal asleep, and be able to put the collar on. The collar can be too tight if the animal fattens and grows. And as a consequence you have things like the Orca killed by satellite tag leads to criticisms of science practices. 

 

So the public loses face in scientists by using invasive techniques. Is there a way to get quality data that doesn't use invasive techniques? And the answer is yes. There are images out there through photographs. We as scientists can take pictures but everybody in the public can take pictures. Everybody is taking pictures, they're putting them up on Flickr, they're in iNaturalist, they are there. There's tons of data. Some of it has a GPS location so you know where the animal is. It usually has a timestamp in the camera. If people set their camera then you can find out when that picture was taken. So you get a spatial temporal map of the species that you're interested in. Scientists can learn to scrape that data. Okay, that's the sort of thing that we're talking about here. That was the first piece of the life cycle analysis that has to be cleaned. 

 

Machine Learning Used To Get A Clean Data Set

 

A lot of the data's not so good so you have to throw some of it out. But can you use machine learning then to process those millions of images to get a clean data set that you can then ask scientific questions about? And I think the answer is yes, but that's where machine learning comes in. Finding the hotspots is in our case is the way we go about it. We started out with IBEIS, Information Based Ecological Information System. When we partnered with Wild Me we created Wildbook and ours is looking at species that are distinctively marked. They're naturally barcoded. And so as a consequence that you can start to look for the features, the hotspots using the sift algorithm to identify every single individual and then compare it to other images. And so these are all the species and some of them are lab animals like Xenopus frogs that are used for developmental biology rather than mark them and toe clip them and change their behavior, you can use the natural spot pattern in the laboratory cause you can always have a camera over the tank. 

 

From Bounding Boxes, To Species, To Individuals

 

So we use the natural markings and look for hotspots and then we've built algorithms to be able to compare individuals. How do you find Zippy the zebra from all of these images of all of these different species? Well, the first thing you do is you put bounding boxes around and we used to put the bounding boxes around by click, click humans doing that. Now we do it with machine learning and we then train convoluted neural networks to be able to find the right species. So who is this zebra? We can now compare it using our hot spotter analysis and voila, by the stripes the key features on the stripes you can get the algorithm to do the hard work. And the interesting thing is that in general there's no information when the stripes look like that. But when they crisscross there's lots of hotspots. 

 

Once you have the data on who, when and where you can start to follow the fates of individuals, the where is critically important. So you can start to build home ranges and then you can add metadata, the type of habitat it lives in and you can then start asking questions about population dynamics. And now as you just saw in that video we can use curve rank to look at the curvature in the notches in the tail flukes of whales. So we don't have to just work on distinctively marked species. So we have a new algorithm and so the software will be able to do algorithm A and then add algorithm B and look at the two as a composite to try to figure out the identity. In this particular case there's 2 million photos of these particular whale sharks. It's in Flukebook and from 32,000 pictures there were 67,000 different individuals discovered. 

 

So we now know the number of individuals, we probably know the location because the cameras are set with the GPS and we can start to put together where they range. Do they move by season? Do they stay put? Are there subpopulations inside the species that have to be thought of as similar or different? This is how a scientist will dissect the problem and try to make recommendations to managers and conservation organizations to try to protect endangered species. And now there are so many different Wildbooks out there. We've got the wow book for zebras, Wildbook for Lynx, Wildbook for Jaguars. We have skunk Wildbooks, we have Flukebook that I just talked about. We've got MantaMatcher. There's a whole variety of different groups using our algorithms to look at the identity, the location and the metadata associated with what they as scientists get to try to tell the story about this species, how it uses the landscape and how it's being impacted by people also using the landscape. And it's that tension that has to be solved and information we ideally think will be the way to solve it. So if you have good data, data, people trust, then you might be able to start to change attitudes and if you can change attitudes, you might actually get changes in action.

 

Example Of AI And Grevy's Zebra

 

So let's look at an example. I studied the Grevy's Zebra, this beautiful animal here. There's two species of zebras in East Africa, the Grevy's Zebra which is big with the big round Mickey mouse here is the white belly and the thin stripes, the plain zebra is more abundant. There's about a million in Africa and they have the big fat stripes that touch under the belly and they have pointy ears and they're small and dumpy. These guys are regal, they're more horse-like than the plain zebra. How many are there? This is a problem. 

 

And so we went out and decided to get real estimates based on where people could find them because normally the government pays millions of dollars to fly roots and count the zebras under the wings of the airplanes. And they're very careful not to refly the same roots because they don't want to double count. And so they'll send out the pilots to do this block on day one and then this block on day two because the chance of moving is very small. So it's a very rigid system and very, very expensive to do for a whole nation. It's millions of dollars. Instead we said how about doing something different? Let's get members of the public, especially the middle class from Nairobi, to leave the city and go out onto the landscape with the people that actually live with these zebras and treat them as vermin because they eat the grass that belongs in the belly of their cow, in their livestock from their perspective. 

 

Let's get everybody out there with one common goal. Snap pictures. We can use our stripe recognition software to start to identify who's who and see how many there are. So we took 350 members of the public conservancy members like this woman here. You notice she's one of my scouts and she's got a GPS, she's got a checklist, she gets paid $3 a day which is an infinite wage and she makes $36 a month. And I can tell you on the side if she's young, she'll convert that $36 into $250 by taking the money and buying sugar, tobacco, and phone cards in town and selling it retail back in the community. So our seeding them, paying them as scouts gives them wealth so that they can send their kids to school and to the clinics if they get sick. So, community conservancy members are critically important. 

 

The Great Grevy's Rally

 

We also use rangers and scouts from conservancies, national parks and reserves. We got involved with the county governments, we asked them for money to run the rally and some of the counties did. But more importantly, members, the workers in the offices said I'll come out for the weekend, I don't need to be paid, I just want to do this. We then use the Kenya Wildlife Service, an independent scientist at zoos and like me at universities with our students. So it's a collective of different publics all participating, doing one thing, driving around, taking pictures, So Red Box is the range where the Grevy's Zebras is 25,000 square kilometers. And what we did is we trained them and their first year we did it in 2016 40,000 images. We had the process in 2018 we had 60,000 images to process and in 2020 we had 90,000 images to process. 

 

This is where machine learning becomes important because most of the pictures are garbage. The animal is too small, it's only a tiny portion of the animal. So there's not enough information, it'll always be considered unique and it can never match which inflates your denominator. Bad news when you make an estimate. So we have to clean up the data that machine learning has to be taught and trained to be able to do that. And so here you've got the picture of the various publics. The white guy in the middle here is the US ambassador to Kenya. At the time he was gung-ho, he came up from the embassy and took part in the process. Now why all this public effort to get a precise estimate is difficult and counts by aerial surveys are incredibly difficult and they're very expensive. So when I started in the late seventies to study zebras in general, there were about 14,000 Grevy's Zebras on the planet. 

 

Most of'em in Kenya noticed the numbers declining. By 1988 there were 4,278 and by 2011 there were 2,546. In theory, if I put the error bars on plus or minus a thousand the government doesn't do anything when the errors are that great, they don't believe the numbers in the first place. And so we wanted to use people to do a better job at getting the estimates and try to shrink those error bars. And so what we did is we trained trainers who went back to their communities to train the ordinary citizens. The National Science Foundation gave us these cool cameras which we got a grant for that are GPS enabled and we then now give them out to the school kids we use for all sorts of projects but every two years we use them, we collect them back to do the great Grevy's rally. 

 

Drive, Photograph, Analyze

 

So we have 150 cameras that go out. Now we have 350 people. So that means we have to connect everybody's camera to the GPS location because most of you won't have a GPS camera that can have a long lens to get a close up. So we get high quality stripes. Now no one ever sets the time or the date stamp on their camera, but our GPS enabled cameras are set and everybody that's in the same car gets a QR code and we say 3-2-1 snap. They take the picture at the same time and from that moment on, those cameras are synced with the GPS enabled cameras. So every picture will be who, when and where. 

 

And so we then just tell them go drive. We break that area down into 45 counting blocks. We give everybody a block that they have to drive. We don't tell them how to drive because we don't want to bias their behavior. Drive as much as you want, drive as often as you want, crisscrosses any way you want, just take lots of pictures. So they drive, photograph and analyze and you can see how HotSpotter defines the hotspots and how it does the degree of similarity in the matches. So HotSpotters doing the work, we then take these known individuals on Saturday and do it again on Sunday. And because we do two consecutive days, we can use a very simple way to estimate the population size. The Lincoln Peterson index. And the reason for that is no births will occur in two days. 

 

No deaths, no immigration and no emigration. It's a closed population. And by having people take tons and tons of pictures, we don't get to every animal being equally capsurable but we get highly capsurable. So we come very close to the five assumptions to let us use the simple equation which is this, the number seen on day one time is the number seen on day two divided by the resighting scene on day two. So the way that works is if you see 10 animals on day one and 10 animals on day two and you resight all 10 on day two that you saw on day one, it's 10 times 10 is a hundred divided by 10, your population is 10. If you saw a 10 on day one and 10 on day two, but only one on day two was seen on day one, it's 10 times 10 is a hundred divided by one. 

 

Present Results

 

Your population estimate is a hundred very simple index. And we could do that nationally, we could do that by county, we could do that by individual properties, which is a reward for the people that open their land. Okay? And so then we present these results to the people but also nationally to the government. So we always have the black and white ball, we all get dressed up either in native clothing or in our fanciest dudes. And I go and I present the results either to the cabinet secretary or the head of the Kenya Wildlife Service who managed the wildlife and controlled money for tourism and wildlife protection. And this is the sort of pattern we get. So this is the 2016 and 2018 heat maps of where the Grevy's Zebra were seen. And you can see they're more or less in the same place in the south and Laikipia, but in the north those hotspots move around a little bit. 

 

The biology up there is a little bit different. 2016 towards the west, they're not there because people didn't follow our orders. They treated it like aerial counts. They didn't red drive in the same areas the next day. So we undercounted. But you can see the hotspots on the second year were a little bit different from the lighter hotspots in the first year. So we understand the dynamics that some are sedentary and others are moving around more. So that's important to know. We also can look at by the county that Samburu, which is where the Grevy's Zebra evolved, is no longer the center of its largest population. They're in Laikipia for the south where the plain zebra is. Why are they there? These guys are arid adapted, they don't need to drink but every three to five days the plain zebra needs to drink every day they're there because the population, the human population, is growing by 3% in Samburu. 

 

Accordingly, the livestock, the sheep and the goats and the cows and the camel are also increasing and the pressure on the landscape is so great they don't let them drink. By hiring scouts, sharing the data with the communities, the communities are going aha, maybe we're the problem. And they change their behavior and now they let them drink with the animals. Okay? But you can see Laikipia now is the hotspot, it's where most of the animals are. And you can see we can provide the results to each community, each reserve and each private landowner. So they take ownership for their numbers and if they take ownership for their numbers, they might take ownership for protecting this iconic historic species. Now that little red circle there is the center Laikipia where the Impala research center is, which is where Princeton has its research center with the Smithsonian Kenya Wildlife Service in Kenya National Museum. 

 

And the three properties inside that circle are Ol Jogi, Loisaba, and Mpala. That's 750 Grevy's Zebra out of the 1200 in Laikipia fully 60% are there. That's the largest population in all of Kenya. Okay? Now what's fascinating is in 2016 our estimate is 2,350 plus or minus 93. That's a tiny error. And then in 20, in 2018 the numbers had gone up to 2,816. Again with a small error of 163. Despite two years of drought. Drought should be killing the animals as it is now we're now in a worse drought. There's been four rainy seasons missed. And you can see by county level that 2018 the numbers had gone up in every one of the counties and they've gone up nationally. But the good news is sort of tempered by the fact that the demographics, the age structure was varying. Now we've done mathematical models that show that if the recruits in the population, which are the infants, the yearlings and the two year olds, the ones that aren't breeding reach 30% of the total population, then they will replace the adults that are dying and you'll have a stabilized population. 

 

So 30% is the critical. And notice in 2016 our behavior by hiring scouts changing the behavior of the local people has most of the county populations at 30% or near 30%. Notice what happened in 2018 with the drought. The recruits all died, the infants in the yearlings didn't make it. But there's something about the biology that's fascinating. Meru was the county where the number of lions were so high that in 2016 the numbers were way off, the babies were getting eaten by lions. We petitioned the Kenya Wildlife Service, please let us use immuno contraception to reduce the number of lions on the landscape. Let's give this endangered species a shot at protection. Now remember, lions are also at risk. So this means the government is having to weigh two endangered species and their decision was nope, we'd have to let nature run its course. 

 

We are not going to interfere. After they saw these results, they let us contracept the female lions. And what do you see happening? The number of babies in that county went up. Data made a difference. And I'm also optimistic about the future because after the drought, the rains came the overgraze landscape that normally looks like this actually looked like that. So if you don't destroy the rootstock of the grass, it will come back. And as a consequence of that, when a female can raise her young to the age of independence, which is one year of age when she stops suckling, she has to skip a year or two or three to recover her bodily strength to breed again. And that success of the baby limits the future reproductive potential. But the flip side of that is when you lose your baby, you start cycling immediately because you haven't really invested. 

 

You've invested in a fetus but you haven't invested in growing up a young adult. And so the females all cycle and this is what it looked like right after the drought. There's 10 females here with 10 youngsters less than three months of age. In other words, the explosion in breeding is likely to have compensated for the lost individuals. But we'll know that from looking at our data from the 2020 census, we haven't analyzed that yet in part because we're trying to improve the algorithms because it takes a lot of work to go through 90,000 images. We're trying to change the way we do it so we don't have to rebuild the data structure every time you identify a new animal. So it's a purely  computer problem and computer vision logic. So my computer vision, people are working on that, but once we have it, will we get the 2020 data? 

 

Policy Makers Respond

 

The consequences of this have been transformative. First off, all the counties governors, and that's like a state in the US, said I have an iconic historic species and the numbers are turning around, I better do something about it. So they came up with six goals that they promised to sign, which is to provide water, to provide money to the communities to be able to shepherd these species. But there's the deputy cabinet secretary who I'm handing the report to and I, and he right after I handed it to him says, I officially accept these numbers of 2,812 Grevy's Zebra as the official numbers for the government of Kenya. That's the first time that a non Kenyan group has ever had their numbers accepted and believed as true. And I asked him afterwards why? He said, well you organized the rally but the people, the Wenatchee, they got the data, the public of Kenya got that data. I have to accept that, I have to believe it. And so those numbers became real. 

 

AI Creates Concerns

 

But AI creates problems. First off, cartels and organized crime are into just moving stuff around. They don't care if it's ivory, whether it's women, whether it's illegal drugs, whether it's false prescription drugs that have no effect but they also move poached skins from the Grevy's Zebra. And so you start to see that all this AI, this pictures with who, when and where is creating a problem. And so you can see, please be careful when sharing photos on social media, they can lead to poaches of our rhino, turn off the geotag function and do not disclose where the photo was taken. Do you think this woman taking a picture of the rhino has actually turned that off on her phone? The odds are no, because she's going to want to talk to people and do other things. And so every time she takes that picture and sends it out, someone scraping the web and knows exactly where that rhino is, puts the rhinos at risk. 

 

And so it's not just on the popular media, you've actually got reports from the Yale School of Environment that is saying this is a serious problem. And then you've got people like Tanya who I work with, who are working with coding encryptors to try to be able to protect the species from the excesses of people that are enthusiastic and want to do good but are a little bit too lazy to turn off the GPS function on their phone. So that's number one. That's a first concern. Second concern is a concern that a lot of the AT&T people were talking about, which was data bias. Data bias comes in many ways. Here's the true population, but how do we get it? Well we take pictures during the daytime, not at night, can't drive around at night, can't see anything. We could use camera traps, but camera traps we think are different from driving. 

 

We're doing a project now with people from Caltech where we put camera traps along the roads. We used to do our censuses during the great Grevy's rally. You know people are savvy. If I see the tail or an ear of a Grevy's Zebra, I'll go off road and get a good picture of it. Camera trap can't do that. It has to be within 25 meters and it has to snap it. And at night the picture's going to be lower quality than during the daytime. So there may be a bias in our estimates based on the fact that we work harder than a camera trap can. So that's number one. Photo capture bias. Half the photos we have to throw out but you have to clean up the data. So that was number one on the AT&T lifecycle. Then there's the issue of social media bias. 

 

People will tell you, oh I've got zebras here, come here. Okay, if everyone goes there, then they're not checking out over here. And you're starting to bias your coverage of the landscape, which we told people. Drive whatever you want. We want their haphazard behavior because that comes close to random, which the models require. So social media can bias you and if you're scared that you're going to fall off a cliff with your car, you avoid that area. Okay? Right. Okay. We did have three cars break down. People wanted us to repair them and we said no, that's your contribution to conservation, thank you very much. But if you are fearful, you put a bias in where you drive, that changes the accuracy of your data. And then comes the model selection bias. We deliberately chose the great Grevy's rally two days in a row. So we could use simple, elegant mathematics. 

 

If you process the data over longer periods of time, it's an open population. Births can occur, movements can occur, the models are there, they're more sophisticated and it's harder to meet the assumptions. So the data has to be cleaned up in different ways. It has to be parsed and filtered. So that's where AI also can come in and as a consequence you get a population estimate. But all of that stuff in yellow could create an error in the estimate. So we have to be careful of that as scientists and that's where AI can help us with that. 

 

Next Steps

 

So where are we going? Okay, the great Grevy's rally has demonstrated the value of AI. My goal, along with Tonya and some of our European colleagues is depicted in this picture. This picture is our dream. Our goal is to use smart drones to explore and exploit the landscape. 

 

As it is, we drive around using tracks and our knowledge, but not everybody has knowledge. Wouldn't it be great if we could have a fleet of drones that are talking to each other, Hey we're over here. Come and get a bunch of pictures, then they go spread out again. So we've got engineers working on flocking in drones. That's number one to that trade off between exploring and exploiting. It's a trade off for everything. Every farmer, every fisherman, every scientist always has that problem. Then long range, how long does a drone stay up? 30 minutes has to come back down. You have to send another drone exactly at that same geo point. Why not have a blimp? That's solar charging station. We're giving away solar, we're giving away phone chargers here. The same logic can apply that the drones can come up, clip on charge, drop off and go. 

 

So they can do it 24/7. Okay, pose estimation. We often get behavior of an animal when we put a GPS collar on if it has an accelerometer because we can get its movements in plain X, Y, and Z. And we can use machine learning to train an algorithm to know that when the X, Y, and Z are this way, it's walking, standing, running, fighting, urinating, dust bathing, sleeping, whatever. But that's invasive. Again, you have to put the collar on. Could we use the body posture? You're sitting here, you're awake, you've got some arms crossed, some not arms crossed, some fingers touching. So we'd have to train the algorithm for people that are attentive coming in different ways. But we can do that. We've got a team of scientists that are also using machine learning to do pose estimation. And then we can use network analysis for sociality, which we already do somewhat. 

 

But if we know who the friends are and who associates with whom, we start to see how the structure of the population, how mom and kids work, how males work, how females work. These are all going to an effect. Your breeding, your ability to spread genes into the future and make babies the recruits that are going to replace the adults that are going to die. And when we can learn about habitat choice, because this is really important, urbanization and development are taking place everywhere in the world. While we watch climate change sweeping and changing the vegetation quality because of droughts and such and such, people are expanding right now. Nanyuki, which is the nearest city to Mpala, is 300,000 people by 2035 the government wants 3 million people there. Where are they going to live? Where's the food going to come from? All this land is going to be converted. 

 

And so this really leads to a fundamental problem. When we met with the Kenya Wildlife Service to talk about the future of Grevy's Zebra, I raised the point about how many people understand green infrastructure is different from black or hard infrastructure. Every developer, every engineer knows when there's a river you build a bridge to cross it or a tunnel to go under it. If there's a mountain you go around it or under it. We deal with hard infrastructure that we can see. Green infrastructure is what the animal sees that we don't see. Can we use unsupervised machine learning to make the invisible, the visible? Can we change the way we look at a landscape through the lens of the endangered species and change the way we respond? Right now, it's designed such that when you understand what a species needs you, you say let's protect that area. 

 

Then a transition area around it and everything else for development, that's called the man and the biosphere model of conservation. Let's flip it, let's say the green infrastructure that you can't see. Give it to the wild species and give that area total protection and put people in the area where the people aren't. So don't deny urbanization, just put urbanization in a different place so that it doesn't impinge. And this became crystal clear when we had our collared data and we're talking about the movement of Grevy's Zebra in front of Kenya Wildlife Service and a representative of the then President Kenyatta. And we showed there were 10 pathways to get from the grazing areas to the drinking areas and the birthing areas. And he said, we're not protecting 10 of them. We'll give you five, we'll urbanize the other five and you tell us which five. 

 

So this starts to get to some of the points that the AT&T people were talking about and machine learning, we've talked about democratizing AI, let's flip it. Can we use AI to democratize science? So one of the points that was raised was many different people's inclusion of different views and values and that's often ethnic and cultural differences can open up the number of lenses you use to ask the questions to shape the AI. So that's democratizing AI. But can you use AI to change people's attitudes, change awareness, make the invisible visible? And that's the challenge that I see. I heard that AT&T uses AI in part to manage people. So let's manage people not just for profit inside the company. Let's manage people to have better information so that they don't destroy the landscape that will come back to preserve conservation. And if the information AT&T provides to let them manage their landscape more efficiently, then AT&T would be democratizing science and democratizing society for conservation. 

 

Concluding Example

 

So let me end with an example. I also work with farmers and the reason I work with farmers is that I got involved with herders because these animals eat the grass that belong in the belly of the livestock. Well the crop people hate the pastoralists and the pastoralists hate the crop people because the crop people are upstream where the water comes off the mountain and they use all the water to grow their crops. So there's not a lot of water for the livestock or the wildlife. So there's a tension between these two different types of agriculturalists. And so I got interested in doing a project with drip irrigation to substitute for the wasteful irrigation that they use. 

 

It's called flood irrigation. They use cheap Chinese pumps at a thousand dollars with pipes and they flood the field to make lakes around all the crops. And it's about a hundred thousand liters every three days per acre, per hectare. So that's even worse.  and that's a huge amount of water. None of it gets downstream. So drip irrigation would reduce that. Okay? So we try to do that and we can create a third growing season by doing that. If we conserve water because they can put it into a tank and they can drip it out during the true dry season. The problem is most of the time they don't even bother to collect their crops. And the reason is no markets. The price at the markets that they go to because the middleman, the supplying chain intermediary goes to one market. There was an article in the New York Times a few weeks ago about the fact that big business, the international exporters use AI on their phones to get the price of the best market. 

 

And the small farmers aren't able to do that. They're just not able to have a click system that you talked about making it really simple for them to say which of the 10 markets I could sell to is going to give me the best price and then how do I get my crops there? So we'd need sort of an Uber system or a Lyft system. They'd need to be able to access distributors because right now they're slaves to one, they don't know which market they're going to, the conditions and the contracts are terrible. It's often cheaper to just let it rot and use it for compost next year than to try to sell it and pay all those movement costs for nothing. Better information connecting to many different publics, connecting the many different publics to each other would provide more income and more efficient use of the land, which would then release the land for the wildlife. 

 

That's called land sharing and land sparing. And, this would be a way in which AT&T which is probably not in Africa, I use Safaricom, but they could partner with other phone companies there to be able to provide vehicles and AI and tools for the rest of the life cycle. Some of these products in the shop that could be available in a click simple way for people that are stressed trying to farm. Because farming is labor intensive. They don't have time, they just want to know the answer and go for it. So those are sort of the ways in which I see a company like AT&T and H2O.ai working together to provide knowledge, information flows and technology through AI to change people's behavior which will indirectly benefit conservation. So this is the team. I showed the dreamers on the left, but then there's a lot of people behind us who also do in the trenches writing the code, doing the analysis. And our goal is AI and data science enable science, conservation, public engagement, and many different publics to connect communities. So they work together to provide environmental and social good. And my challenge is that companies in the US that have access to the tools to make the products that could change people's attitudes towards science, towards information and make them better decision makers will have huge effects on conservation. So thank you very much.