Google’s AI Approach to Climate Change Flashcards
In today’s episode, I’m talking with Yossi Matias, Vice President of Engineering & Research at Google, and the founding Managing Director of Google Center in Israel. Yossi leads teams that are developing AI solutions to address climate mitigation and adaptation.
Google was actually one of the first companies to become carbon neutral back in 2007 and made a bold commitment. Last year we set a goal to achieve net zero emissions across operations and value chain by 2030. And this is something that touches many aspects of our organization. Now, also AI, we are an AI first company, and AI is actually obviously something that we’re looking into in all our products and function as operations. I’m now part of the research organization where we’re developing and advancing AI and part of what I’m doing is also advancing AI for the benefit of society in areas such as climate crisis, health AI, conversational AI.
March 2023
02/03/24
I’ve been leading Crisis Response at Google for a few years, but actually my involvement in Crisis Response goes back 12 years ago when we had one of the biggest wildfires in Israel for many decades near our Haifa office in the Carmel Mountains. I remember seeing that huge mushroom like smoke in the sky and looking for information on the internet to see what I should be doing. And I couldn’t find any useful information until eventually calling the mayor’s office and getting some actual useful information.
And this became the genesis (birth, start) of my efforts to look into how to provide actionable information to people during crisis because it turns out that the same phenomena occurs whenever there’s a crisis, natural disaster or other types of crisis, people are actually coming to google to look for information about what’s going on, how to remain safe, and there is an opportunity for us actually help people with this kind of information.
Soon after, I discovered is that on one of the most devastating natural disasters, namely floods, the information was just not there. Floods are estimated to have an impact on thousands of lives annually, actually thousands of people lose their lives and many millions impacted. The helpful information would be to alert people before floods occur. And this information was just not available. So this was the genesis to try and ask ourselves, “Is there anything we can do with AI in order to have better information that could help alert people about floods?”
Fast forward, it turns out that with advancing the science and technology and using the most advanced AI that we can have, along with advancing the science of hydrology combination with machine learning, we can actually predict floods pretty accurately. Just last year we sent out over 115 million notifications to people at risk in India, Bangladesh alerting them on floods coming their way eight hours or more before they arrive. And last November we actually expanded something we call the FloodHub to provide flood forecasting information to about 20 countries. So I found out that AI can actually be very helpful to address such a difficult and important problem such as floods.
And in parallel we started looking into other problems. So if the genesis of my interest in this space was wildfires, then now we already can use AI to actually have pretty good wildfire detection, boundary detection, so that we can actually provide near real time information to people and authorities about where the boundaries of wildfire is based on satellite imagery, again using AI. And of course the promise looking forward is to see if we can even have earlier kinds of warnings about fires as they earlier when they start.
Now when we think about the climate crisis with global warming, what we see is the increase in frequency and intensity of the likes of floods and wildfires. So obviously this shows the opportunity to use AI to actually address such disasters.
So you’re taking data from weather forecasts, prevailing winds and wind forecasts for wildfires, and maybe how dry things are using satellite imagery. On the flood side you’re using topography, rain forecast, and noting where people live. And you’re combining all this to create a predictive model to say, “Where are the wildfires going or where are the floods going?” And then you’re pushing this out through tools like this SOS Alert and other tools to communicate both with residents as well as with governments. Do I have that right?
Yes, it’s a pretty succinct description. I would say that AI plays multiple roles. So for example, for floods, you can think about multiple problems that need to be addressed. One is to understand the situation of where the water is. And sometimes we get this from partnering countries. Sometimes we can actually learn it directly from satellite information.
But there’s also other information that is critical to have such predictions such as understanding the lay of the land the 3D topography of the rivers so that one can actually apply those physics, hydraulic simulations in order to predict where the water is going to go and the extent to which they are going to go because we really would like to have predictions which are quite accurate about the water level in order to be effective in warning people.
And there’s the question of dissemination. Where here it’s again a combination of either using the mobile phones of people, we have also some partnerships with organizations on the ground that sometimes use some low tech technologies to notify people about things about water coming their way. This needs to be quite accurate with high confidence, otherwise people would not take action. So you don’t have too many chances to warn people if you can’t be trusted about these warnings. So accuracy of prediction is actually quite important.
Let’s talk about accuracy. That’s one of the questions that was coming to my mind. I’m glad you mentioned it, it seems to me it’s all different ways to measure the predictive accuracy of algorithms. A simple way of thinking about this often is a type I error and a type II error. A type I error being when you predicted a problem that you suggested people should evacuate and then it actually didn’t materialize. Type II error being you didn’t predict it and it materialized anyway.
It’s a great question. So what we’re trying to do is actually provide information in both directions to show on the map, here’s a polygon that with very high confidence is going to show where the flood is going to be. Now there are trade offs, so when you show high confidence there’s going to be floods, you are going to have very few false negatives. But by definition, have more false positives. So then it’s a judgment call of how to take these two areas and what actions to be taken.
But what we also learned three years ago, I went with my team to see things how they are in the field and what we really learned is the importance of the water level. So we got into this place that our predictions show that it’s flooded and it was flooded except that what we’ve seen is kids playing with the balls in the water and this actually highlighted that we need to have the water level is critical because if it’s very shallow then it’s inconvenience, or fun depending on what’s going on. But if it’s deep, that’s actually where it’s life threatening.
The other thing that I wonder about this, is how this is interfacing with governments because governments are usually thought to be the ones who are issuing these warnings, the evacuation warnings or the boil water warnings and things like that. And you mentioned you’re working with governments, but if folks are finding this on an SOS Alert coming directly from Google, if there’s a conflict between Google’s advice about whether to evacuate or not or how much water’s coming and their government agency’s advice, how are you managing the issues there, where the government might be upset with you, if you’re telling folks to evacuate and they’re not, or they’re telling people to evacuate and there’s no mention from Google, how do you manage that conflict?
So actually there’s no conflict because these efforts are actually great examples of partnering between technology and governments. There’s a very nice op-ed that was written just before COP26 by the Bangladesh government official highlighting that the innovation of the type of partnerships that we have on flood forecasting team is kind of a role model to the kind of innovation that we’d like to see for climate crisis. This is actually an example where we’re partnering not only with governments as well as with academia, with NGOs, with organizations on the ground.
So let’s talk about other opportunities for Google to use its AI technologies to address maybe longer term concerns on the adaptation resilience side. I know you’ve done some work on urban planning with agriculture, with land use change. What’s another story you could tell us, pivoting now from the short term crisis response to a longer term issue?
Let me give one example of a project that is showing a lot of promise in doing that. So obviously a lot of the carbon emission is coming out of cars and much of it is actually happening in the cities. It turns out that a lot of the carbon emission happens actually in intersections where cars are waiting at traffic lights. And not only because they’re waiting, but often actually they are starting and stopping and starting and stopping. And it turns out actually this has a significant contribution to the carbon emission of cars in cities.
So we set up this initiative that we call Green Light where we’re looking into if we can do a better job in how traffic lights are scheduled so that we can reduce the number of those events of start/stop, which are contributing to carbon emission. We do that without actually putting any sensors whatsoever because that would actually be a pretty expensive and slow process, but rather just by using information that we have from apps or waste of traffic flow through those intersections.
And it turns out that we can do quite well. We had 10 cities in pilot in 2022 from Rio to Bangalore, Seattle, Budapest, Manchester, multiple cities in Israel and more where we try to apply, and the way it works is just analyzing the traffic that we can see and applying AI to try and get a better optimization of the traffic slide schedule and suggesting to partnering cities to try that schedule. So example in Hamburg, Germany: cars made 25% fewer stops resulting in approximately 10% fewer emissions at intersections. So that’s an example where we can actually use technology to address a phenomena that is contributing to carbon emission. One interesting aspect of this is when you think about it is the relative ease of deployment, no sensors required, pretty simple change. It actually integrates into the workflow of existing cities and still it can actually make a difference in the carbon emissions.
Currently we have traffic lights, the traffic patterns that we have cars through maps and Waze in the aggregate form. So it’s the same type of data that was used by cities during COVID when they were looking for mobility reports to understand where cars are going or which direction or essentially the same information that they used to show you that there’s a traffic jam ahead of you and how long you should expect the traffic is going to get you to get from where you are to your destination.
Essentially that information is all about an aggregate number of how many cars are there. So this information is just information that can be used for optimization. Of course the technology itself is how to actually do the scheduling in the right way and in a way that actually scales effectively. And that’s where machine learning plays a role in solving the problem effectively.
The other thing that’s super interesting to me about this Green Light program is like it’s quite complicated to change one light pattern and then model how that traffic will be resolved at that one corner because that might be in initially what you’re trying to optimize, but then well that traffic then moves to the next corner and you have to sort of optimize the system and not just one street corner at a time because these things all have interactions and you’re doing all that in the background of your optimization algorithm.
I mean in some cases what we really need is something that is real time analysis. This is an example actually that typically the partners are pretty consistent, so it’s more like doing the optimization. But you’re right, I mean this is kind of an optimization problem that those starting computer science would possibly apply various techniques to that for years now. And now with machine learning technologies and new ways to actually compute optimizations of this sort, we can do a much better job and a much better scale.
Another example I can perhaps highlight is a project called Tree Canopy: cities around the world are really concerned about is what they call heat islands, which have implications on particular areas in the city that are subject to higher temperatures and also air pollution, which has impact on health. We can provide the information to cities so they can make decisions of where to plant trees such to lower the street level temperature, make streets greener, and improve air quality. So to understand the tree location that is going to make the biggest impact, that’s a design principle that is helpful to actually make the right kind of investments. Obviously this is kind of a longer term investment, and this is a project that, for example, just last year it was expanded to hundreds of cities. It’s another example of combining data with optimization. So it’s another example of combining data with optimization where here you’re optimizing, it sounds like heat island effect mitigation and perhaps also pollution mitigation.
For example, the city of Los Angeles piloted Tree Canopy and now it’s become a critical piece of the city’s long-term goal to increase tree shade by at least 50% by 2028, in areas of highest needs. So if you think about it, you have limited resources of course for investing in a city. And then the question is how do you invest it in tree planning so as to get the best result for the city.
Google Cloud has launched some accelerator programs to link Google’s climate products to entrepreneurs.
We have multiple programs where we are encouraging and supporting entrepreneurs and NGOs and startups worldwide. So first one kind of observation is that innovation can come from many directions and definitely I think there’s an opportunity for many to contribute towards solutions for the goals of addressing the climate crisis. So for example, at Google.org, which is our philanthropic arm, we announced the 30 million impact challenge for climate innovation. And we made a $25 million commitment for what we call AI for Global Goals, which is to use AI to accelerate progress against the sustainable development goals. And we also have a program that is called Google for Startups Accelerator, where we did dedicate one of our focuses to sustainability in something that we call Startups for Sustainable Development where we now already support about 400 or so startups, which are working on various aspects of sustainable development goals. And it can be anything from direct mitigation or adaptation or it can be like food security, which is another kind of reification of the climate crisis. And the idea here is that we can support startups by providing them with mentorship, with collaboration.
For example, we have an Africa collaboration between a research team in Accra, Ghana and a company called InstaDeep about using machine learning to identify areas of locust breeding. By identifying it early enough, the premise is that we can actually reduce the damage done by locusts, which has a severe damage of course, on food security.
Let me give you an example perhaps to one startup that is helping small businesses through one of our programs, a startup called Normative. Normative is providing an AI enabled business carbon calculator and it was supported by our Google.org, including fellows, which essentially are Google employees that are helping out the startup for a certain amount of time and it actually developed the world’s largest free resource to measure and reduce carbon emissions for small businesses. So the tool now measured more than 1.7 million metric tons of CO2 emissions, equivalent to annual emissions of about 200,000 people in their daily life. And it now helped over 2,600 small businesses calculate their emissions over more than 80 countries.
So the beautiful thing about when we think about the efforts that we’re doing at Google is that when we’re trying to solve a problem, then the problem doesn’t need to tie back to financial consideration. Once we made a decision that it’s part of our mission to help people in the context of the climate crisis, then the metrics by which we measure impact is how much progress we’re doing on that. For example, in climate mitigation, the ultimate metric is how much carbon can we actually save? And that’s actually the only guiding principle here. Similarly, in the context of providing crisis response information to people, then the metric, the relevant metric here is about how many people can we help and to what extent can we help them.
These tied directly to our mission of organizing the world’s information making it universally accessible and useful. Plus it tied to our efforts to use technology to help society in a strong way. So these are kind of the guiding principles on these efforts. As it relates to the climate crisis, here we made very bold commitments, which is about how to help out with carbon reduction. So in addition to net zero we also are having goals about how to contribute to help with climate mitigation. And these are the guiding metrics by which we measure the impact that we can do.
So if there’s no explicit financial returns in those considerations, how does Google decide how much resources to provide to this effort?
So the way I think about it is that given resources that we committed to put on that, how do I maximize the impact of what I can actually get out of that? So there are multiple ideas and multiple directions we can invest in. The question is really all the time, how do we invest so as to maximize the impact of what we’re doing with whatever resources that we can.
Let’s take a glimpse into the future with the advances of AI and the advances in data on climate of all sorts. What do you think are the most promising opportunities we haven’t seen yet, but that might be of five, 10 years down the road?
I think the most promising opportunities are those that we don’t know yet. And again, practically most of the examples they touched on today are things that perhaps we didn’t think about them a few years ago. And again, I like to think about flood forecasting as an example for a problem that even though it appeared to be too difficult to tackle, it turns out we can actually more or less solve it effectively and at scale. Similarly, Green Light is something that is a little surprising in its simplicity. Simplicity from a conceptual point of view, no sensors required, it’s just taking this information, make it better experience for drivers and reduced carbon emission.
The other of course is technologies. Various types of technology that can help us identify opportunities, providing better information to people, providing better ways for people and businesses to take action that are going to result with mitigation, with reducing the carbon emission, more efficient energy, better way to actually utilize more efficient energies. I would say that possibly some of the opportunities are opportunities that we are not thinking about them as yet.
For those of our listeners who are interested in getting into business and climate change issues and perhaps into this particular interface, what do you recommend they do to prepare themselves? And which types of companies are hiring? Are there global hubs or global cities where these types of opportunities are more likely to arise?
There are hundreds or thousands of startups and NGOs trying to tackle various aspects of that because it becomes a priority to businesses, then I know that the climate tech industry is also starting to emerge and there are venture capital firms who are investing in climate tech related initiatives and ventures.
My advice is basically to look for things that matter, to try and ask about any direction, how much impact is it going to have? Obviously depending on the objective, it can be either the business impact in the area of climate tech or it can be the impact derived by how to help out, address the problem in innovative ways on either climate adaptation or climate mitigation.
