Laney Seigner 0:00
I'm going to start recording and yeah, get into some intros. I'm really excited to introduce our panelists for today's guest talk. We have three wonderful people joining us to share more about climate, AI, and machine learning applications. Aaron Davitt received his BA degree in ecology from the University of Denver, USA, in 2006, a MA in geology from City College of New York, USA, in 2011, and recently received his doctorate from The City University of New York Graduate Center last month. His previous research focused on the application of optical-infrared (Landsat-8), thermal infrared (ECOSTRESS) and microwave (Sentinel-1A/B and SMAP) remote sensing data to inform on crop growth and flood monitoring, and modeling of crop canopies. His current position is at WattTime as the principal analyst, remote sensing, where he helps identify satellite datasets for use in the Climate Trace coalition efforts to improve greenhouse gas monitoring. Welcome, Aaron.
And Gopal is - oh sorry just some back our noise. Gopal, do you want to give us a wave so we see you up there in the in the - there's Gopal. He is the CTO and head of product at Sust Global, a focused on geospatial analysis, or analytics for climate adaptation. And most recently, he led the analytics engineering team at Planet Labs. An aerospace and data analytics company that operates history's largest commercial fleet of Earth observation satellites. Planet Analytics serves a range of customers from city planning teams in governments, and the World Bank to defense and intelligence functions across the world. Using agile engineering execution from concept to scalable, high quality products, he has been an invited speaker at global industry conferences like Google Cloud Next, and leading technical conferences in the machine learning space, such as ICML and CVPR. Gopal holds a master's in electrical engineering from University of Southern California, and completed the Ignite Program, connecting technologists with new commercial ventures at the Stanford Graduate School of Business.
And Anthony, let's see who is looking at my screen here. There's Anthony's beautiful background up there is head of product and growth Climate.Ai. And he has over 10 years experience commercializing technologies at the intersection of science and technology, and multispectral imaging and analytics. He is the co founder of the Stanford alumni in food and ag group, and a mentor with tech stars Farm to Fork Accelerator. And he is a Stanford alum and first place winner in an International Science and Engineering fair. And yeah, we have we were just so lucky to have this group of panelists here today. So I was not planning any particular order for each panelists is going to share briefly in 10 to 12 minutes, their work, and they're what they're up to. And then we're going to leave plenty of time about 25, 30 minutes for open Q&A and discussion. So let's start with you, Aaron.
Aaron Davitt 3:30
All right. Well, thank you for having me. Let me get my share screen going. Can everyone see the screen here?
Laney Seigner 3:48
Aaron Davitt 3:50
Great. Okay. And if there's an issue, just yell, because we all know how problematic Zoom can be, so. All right, yes, thank you for the introduction. My name is Aaron Davitt. I am principal analyst of remote sensing at Watttime, which is part of a bigger coalition called ClimateTRACE, which stands for tracking real time atmosphere carbon emissions. So what I'm going to do is give you a sense of what ClimateTRACE does, is this coalition but also the role I play in it.
All right, sorry. There we go. All right. So what is ClimateTRACE? It is a initiative launched by a coalition of organizations that will use artificial intelligence, machine satellite imagery, machine learning other technologies to monitor greenhouse gas emissions worldwide. Our mission is to make anthropogenic carbon emissions transparent, accessible and actionable. So world's governments, NGOs, private sector, and citizens through collaborative with like minded organizations. So this coalition ClimateTRACE was started with Watttime, Google AI and Carbon Tracker. So it's a bottom up creation of like minded organizations and individuals that want to do good. And each member here that you can see what the images here specialize in a specific sector. And when I mean sector, I'm talking about the transportation sector, construction, agriculture, and so on. So each of these members here monitored greenhouse gas emissions from that specific sector.
So climate trace aims to launch a publicly accessible greenhouse gas emissions database in the summer 2021 that is global, comprehensive, real time empirical and extensible. But the question is, why create this database? Well, just some of you may or may not know. Currently, in many countries and individual sectors, the state of the art tool to monitor greenhouse gas emissions, is still to ask emitters to self report their own emissions, and then that information is mainly compiled for the results. Consequently, many governments companies and scientists must rely on data that can be years out of date, sometimes subjects that deliberate under reporting, and often only provides a high level summary of information that lasts specific details. So the goal or the mission of climate trace is to cover all major greenhouse gas emitters, sectors globally. And this is meaningful for a couple of reasons. First off, like I said, many countries, especially in developing world do not have access to technology or financial resources to be able to install greenhouse gas monitoring systems. And two, for many sectors, the greenhouse gas measurements are not sufficiently precise and given the contextual nature of emissions from sectors. So ClimateTRACE aims to address this by providing data that will be available for all countries regions, and by improving on the state of the art for emissions miring specific sectors.
So here's the idea of the tools and technologies we will use to monitor greenhouse gas emissions. So the ClimateTRACE system incorporates different types of satellite imagery collected by various space agencies and organizations around the world, then the AI machine learning can be trained to spot indicators of pollution and greenhouse gas emissions in the satellite imagery. This will be fed into a web based database that will be accessible to the public. It will be sortable in a variety of key breakdowns and clean geographies, sector, by source, etc. People will be able to generate reports and data sheets and data and conduct analysis from there. So, although each coalition member covers a specific sector, all models will be based on some form of remote sensing data and ground truth data, analyzed though AI machine learning models to estimate greenhouse gas emissions, and it will be extended to regions without ground truth data. Now, just so you know, we're not doing academic project. This is not just, you know, crunching numbers and that's it, but we're trying to make a real difference with this tool, this information here.
And we are working from a certain impact objective, as you can see here, and when I'm talking about impact objective for example, I'm talking about building trust in the Paris Climate Treaty, to have objective data on greenhouse gas emissions, but also helping develop the developing countries have the tools needed to monitor greenhouse gas emissions, to create some type of meaningful impact with these examples here. Now, right here, as you can see, is my specific job the red box outlines the satellite data here, this is the where I come in, I work on a satellite data side of the project.
So I want to make it clear with my experience, I am not a data scientist. I am a remote sensing scientists, and some of my previous work focused on floating crop monitoring, which requires combining satellite imagery modeling ground data, a lot of the time with my doctoral research, focus on this, but also how can satellite and imagery be used individually and in combination to monitor land surface conditions? So taking this and by extension, the climate trace? I'm plying, trying to understand and inform the data, inform to the data scientists "how can satellite imagery be used to monitor greenhouse gas emissions?". So really, my role at Watttime is to leverage my remote sensing expertise to bridge the work of the coalition members and collaborators that have the highest impact on global greenhouse gas monitoring and reporting. With that I work with the data scientists policymakers, ground data experts and like in a holistic approach to monitor missions. And the end images here, kind of give you an idea of what my role entails. So here's an image of a power plant in Arizona. And what I do is identify different satellite imagery, as you can see here with ECOSTRESS and ResourceSat2 so this is a sensor mounted on the space station. And this is an Indian satellite that's been in orbit for a couple years now. ECOSTRESS measures land surface temperature, and ResourceSat2 measures different bands of the electromagnetic spectrum. In this case, we have red, green, and shortwave infrared imagery here. So the question is, how can we use these satellite images to spot indicators of pollution and greenhouse gas emissions from this coal power plant? Also, can we these information gleaned from these satellite images be transferred to other locations? And can we use other satellite platforms to inform us emissions coming from this coal power plant or some indicators of emissions at this coal power plant? In addition to identifying satellite imagery that can be useful, I have to interact with the coalition members, cross different eight organizations to understand their remote sensing needs and provide some advice on what to use. I do education efforts to explain the work and application of what we're trying to like with this presentation. This panel here, I field inquiries from the data scientists on how to improve algorithms relate to remote sensing data, I prefer perform my own. Sorry, question. Okay, I'll keep on going perform analysis. Also make informed decisions on purchases of satellite remote sensing data, because not all of its free out there. And like I said, communicate this information to data scientist. So looking back at my PhD and teaching career, I lean into the skills I developed over time and apply them to my current role. So here's another example of what I do. Here's some specific sector coverage for the two coalition members. We have here Blue Sky Analytics based in India, and they look at farm and forest fires. And we have Hypervine here based in the UK that looks like look at emissions from construction and mining. And on the Venn diagram here provides an example of what I tried to do. So Blue Sky Analytics looks like crop burning Watttime looks at plume emissions from power plants, that's my organization, and Hypervine's looking at identifying mining type and emissions. So the question is, how can we use the same sensor to answer all three of these needs? And then I relay that information to data scientists and the AI folks, and also the coalition members so we can decide, okay, we can focus on one satellite to answer all our needs.
So some key takeaways. The climate trades coalition requires different expertise to monitor greenhouse gas emissions. So being a data scientist is only one member of the team. So I'm a scientist, remote sensing scientists, engineers, and more all contribute to the team. So the skills required for my position being a remote sensing scientist is being familiar with various remote sensing platforms. And the application context is important. Understanding how the data will be used, what the data is measuring, and what data is needed to be useful. Also experienced with combining remote sensing, modeling and groundout to answer a question that's very useful experience, communicating the data and application especially in the context of the mission, and also being able to work with a team and translating across different crews, between data scientists, policymakers, and then like my side remote sensing folks. So thank you for your time, I appreciate you listening. So if you want more information, feel free to email me at Aaron@watttime.org. And feel free to check out these websites here. This is the watchtime website. Here's the ClimateTRACE website to provide more information, but ClimateTRACE is doing and you can follow us on Twitter. But also here's an article that was sent to Medium back in July talking the launch of this coalition. So if you want some more information, the Medium article is a great resource to use. And also on the ClimateTRACE website feel free to check for job postings, I everything we are looking for data engineers, but there might be other positions that might be of interest to you. So thank you.
Laney Seigner 14:46
Thanks so much, Aaron. That's great. That's really super informative and wonderful to hear about. I'm sure if anyone has any burning questions that they want to bring up now go ahead. We can maybe just take one if there's if there's a big one. Otherwise, we'll go into the other two panelists and have an open Q&A at the end.
I thought to ask a quick question if that's okay. Do you, I'm kind of curious about the like, the kind of litigious side of all of this, like, you guys, are you guys worried at all about kind of getting...sorry, I wrote notes, and I'm losing them Are you afraid of like these kind of big fossil fuel companies trying to like suppress your findings? Also, I'm just curious if you guys find any kind of illegal activity through this or like, activity that shouldn't be occurring, or should be more regulated. But is is occurring?
Aaron Davitt 15:46
Yeah. I'll, I'll answer your second question first, before I get to first one. So yeah, one of the things is we were trying to be an independent monitor of greenhouse gas emissions. So we're trying to be well, we are being objective with this. So one of the things is tracking legal polluters or illegal activity, that is one of the goals, but that's only a small part, too. Because initially, when I enter this role was like, yeah, we're going to get those people that are being, harming the environment. But it's also we realized, too, there are also a lot of groups out there, companies and all that are actually doing good. They're actually doing good job reporting on emissions. So this is a way to support that like, Yeah, actually, you know, they're right in their estimates, and that helps policy to giving them some support that way, so a little bit going after, like the legal but more about supporting, like the good actors in this. And the first one is in terms of suppressing, we haven't encountered that. But also, there seems to be a shift with some of the industry, people we've talked to to do a better job with tracking greenhouse gas emissions, because it seems to reach a certain point where it's, it's economically important to be much more transparent, but also considering the greenhouse gas emissions.
Laney Seigner 17:15
Thanks for that question and your response, Aaron, and we're gonna move on to Gopal and Anthony. But, Aaron, if you want to look at a few of these questions in the chat and answer some of them while they present, that's one option, and then any that don't get answered, or that are longer explanations, we can come back to in the open Q&A. So let's see Gopal, I will turn it over to you. And you should be able to share your screen and yeah, get right into things.
Gopal Erinjippurath 17:43
All right. Thank you for like, organizing this. We're gonna try and share screen. Is it showing up? Yeah. It is, I see you nodding. Wonderful. All right. Good morning. Good evening, everyone seems like we have folks joining from all over the world here. So super excited to be presenting to you all, and sharing some some of our hard earned wisdom on some of these climate related topics to all of you and excited to engage with this community. Thank you, Aaron, for your presentation, it touched on a few of the things I do in my past life around sensor modalities and bringing data together towards having meaningful insights. So I want to tell you a little bit about planetary scale location insights and how they're connected to climate adaptation seems like all of you have come together with a common purpose around climate change and your interest in climate change. So I wanted to share some knowledge there. And you know, hear your thoughts to this process, too. So I serve as the CTO and head of product for a new venture Sust Global, where we focus on global sustainability of operations in the context of the ever evolving climate, and our company's been founded with the mission, to democratize access to climate data, and make the world's climate data universally accessible, and actionable. So, you know, in just just to give you a quick background, I've spent the better part of the last 15 years working on different aspects of spatial temporal data. I spent a lot of time in imagery as part of the founding team building Dolby Vision, which is like a product in the cinematic space, and since then transitioned into more environmental data, working at Harvestings Planet in the last few years. I was the head of analytics and insights at Planet there, the we were responsible for bringing out some of the data driven insights from the rich collection of Earth observation data sets that Planet collects. And now I work on this new venture where I have the fortune to share some of this work and past, in some interesting venues in the past.
I just want to talk about climate adaptation real quick, and how we look at this the space. So you probably heard a lot about the climate models. And in the broader context, you can think of it in the context of space and time. And most climate models have uncertainty in some form. And that kind of affects decision making based on climate, as you'll see in from a few examples, I want to, I want to share with you that temporal uncertainty and positional uncertainty, and when you look at most global climate models, they cover the historic, they cover the future, they go from today to the end of the century. And the challenge, there is often the positional uncertainty, which is how accurately or like a specific region, can you understand the effects of climate. And what we have learned in the last few years is when it comes to policy when it comes to look at like decision making that happens at the local level, whether it's climate adaptation, or climate mitigation. And climate models go very global, and which is great, but they're often challenging to make actionable decisions based on. So in order to bridge these two worlds between climate models and business decision making, you kind of need to be able to bring together the knowledge of the environment and the changing climate, at asset level. And that's what we call asset level geospatial analytics. So bridging the void between climate models and actionable decisions.
So I want to talk about, you know, take a step back and talk about imagery analytics for location insights, I feel like many of you folks are like, you know, in regions where there's been increasing amount of flooding the India, Bangladesh, parts of Europe there, you know, there's there's 1 in 100, severe year floods, severities that are impacting regions in successive years, which normally shouldn't happen.
So I want to just talk about how how you can do some analysis at the asset level in this context. So you know, talking about the work we do the planet, like here is an indication of the Dow satellite, it's me, planet, we have upwards of 150 of them in orbit. Some of you are curious about resolution is probably four to six meters, depending on the latitude of your app. And in terms of ground sampling distance, with the orbiting out to the earth like close to five kilometers. And that, through having many satellites, we're able to create fairly rich mosaics, which have fairly good coverage at like a monthly or an annual cadence. So think of it as like, through these satellite data captures on Earth observation, you're able to get to something like a Google Earth every week, or every month. So through that we can find changing patterns of human settlement is a work we did around bridging the gap between data science and the geospatial data. And what's actionable intel. This is a map of all the roads and buildings across the world in 2019, which was not using running some machine learning algorithms on top of the monthly and seasonal mosaics from climate data. So one of the things we looked at in terms of urban growth is over the course of, you know, study period of 18 months, how to cities really involve and which parts of that city fall under flood risk regions? And that's kind of the data that civil governments, city planning officials often often craving for, because not everywhere is there like a consistent way of doing census and understanding patterns of growth. So this kind of the rough workflow, you go from imagery to monthly build, you know, building maps, and then from the building maps over successive months, you can identify changes of building growth, which is indicative of changes in human settlement, and how they intersect with flood risk. And what are the hotspots which are vulnerable to floods. So that's a high level idea. We looked at five cities. And but I just wanted to show you examples of the imagery related to this. So one of the cities we studied was Bangui, in the Central African Republic. And this is kind of what the mosaic looks like. So if you look at October 2017, this is what the city looks like from overhead imagery at like four to six meter resolution. And I'm going to just flip through a few a few images in successive months and show you kind of what roughly we can see us a patterns of growth. So this is what it looks like in the beginning of 2018. This is April 2018 is like mid 2018. And you can see over the course of months, you kind of see an expansion of the city. And then if you run the machine learning or AI capability on top of this imagery to detect buildings, this kind of what it looks like in the first month, this is kind of what it looks like in the last month of the study that's 18 months apart. So this is the beginning of the study, this is at the end of the study. And through that you can identify the hotspots of built up area. And the hue of the yellow here is an indicator of the increased amount of building density in a specific region around the city. So this is indicative of the kind of patterns of growth in most African cities, because the growth is more lateral rather than vertical, like what you see in like Dubai, or like San Francisco or any of the developed cities. In Africa, it's more lateral, and the city expands kind of outwards, and you can see that in the cities are one of the fastest growing cities in Africa. And if you overlay a flood this map, this kind of what it looks like. So what you see in red, are indicative of areas with very high flood risk. So ideally, you don't want a lot of human settlement there. But then this kind of makes us begs the question, which is where are we be building and are we building consciously. And that's kind of the question, you can answer this. So you can identify hotspots like, like the one highlighted here. So then you can quantify that risk, which is between the period of the study how much of growth has existed, and what percentage of that growth is in Iris areas, in specific cities. And that's kind of the capability we were serving city planning teams across the world, including the UN building development team.
So that one case study, I want to tell you about another interesting one around wildfires. So no conversation in Northern California, especially in the parts of the world where I'm at, which is in San Francisco, is just kind of like relevant if you don't talk about wildfires, because that's what we've seen the most of in this in this year. So just take a step back and look at climate models. This kind of what an earth system model predicts in the context of warming temperature between the beginning of the Industrial Revolution, I think that's how they pick 1850 and the close of the century. And you probably you've seen some version of this figure in in any textbook around climate change. But what this tells you across different scenarios, what does the average lift in terms of warming temperatures look like across these scenarios, which are called SSPs, and representative concentration pathways, so to get to the kind of model, the social and economical dynamics across the world, and how that connects into climate change in warming in in different parts of the world. So this is kind of like across the world, the warming graph, and you can see the variability based on the scenario. So if we as like a society, do the right thing, the right thing is essentially think about climate change and have the right adaptation and mitigation measures, we have the opportunity to make this like a less warming scenario than if you did nothing. So that's one thing you can take away. But interestingly, like this is kind of like a collection of mortgage portfolios across the portfolio assets across the US. And you can see many of them are based in California. And if you wanted to study the asset level risk across each of these portfolio elements, you can then go into those climate bargains, look at the representative model outputs, and then ensemble those results, and then come up with a heat map. So it's relevant because in California in the last year, you've had the last multiple years, you've had so many wildfires, it's kept the fire and rescue team incredibly at stretched, and has created a tremendous amount of economic as well as social damage. And in, you've probably seen some pictures in like September, there's another sky look like that at 11 pm in San Francisco, it was just like orange and such, and it's a sight that many of us have never seen. And it's an indicator of like the wildfires that were happening in the neighborhood and the smoke flowing closer to the bay. And if you looked at the risk of the different assets in that portfolio, and you sorted them by their risk exposure, then you see that the riskiest assets out pretty much in Southern California and parts of like the northeastern us and what have you seen yellow like that close to no risk and situations. So, to that you can aggregate the vulnerability across those portfolios across different types and look at risk groups over assets and then you can get to estimated loss distributions across different types of assets. So the key idea here you've gone from a portfolio to, essentially understanding across different kinds of assets where the risk is. And that helps actionable insight in terms of funding mechanisms and in terms of new kinds of buildup in terms of loan underwriting. And that's kind of a decision making around climate adaptation often for science. So that's kind of like the workflow, we're trying to get to climate adaptation measures. And I just want to take back, to go back to asset level geospatial analytics, you know, everyone thinks it's all about data, it isn't. So in order to make meaningful functional capabilities that can influence the world today, in terms of the decisions that we all make, in a combination of spatial inference, machine learning and data interpretable metrics that can be communicated to a broader audience, and in the right kinds of interaction design. So those are the things we're focused on at Sust Global. And, yeah, I'm happy to field any questions, feel free to reach out, you're always looking for spark talent to join us. We like an elite team of scientists and commercial folks. So tons of open roles we're actively hiring for. So how do you want to do this...Laney do you...?
Laney Seigner 31:19
Yeah, thanks so much for your talk. And yeah, I especially appreciated that point at the end, like actually having data be actionable is a is a step that is really important and some folks in my PhD program we're working on that really had a lot of work cut out for them to take the climate science models and communicate it in a way that whoever they were talking to, you was really going to be able to use it. So that's that's a great point. And yeah, let's take one question for Gopal. If there's one that someone wants to jump in with, and then feel free to keep writing other questions in the chat, and then Gopal, similarly to Aaron, you can just respond to a few of them in writing in the chat as we get to the last panelist. So anyone want to just ask a first question to Gopal.
Aaron Davitt 32:11
I can ask a question. If no one else.
Laney Seigner 32:14
Go for it.
Aaron Davitt 32:16
I was just curious to know, you had a flood- you're looking at buildings are going out laterally. And you're using a flood risk area to determine vulnerabilities, buildings, always wondering if that flood risk map is that a static map or does it get updated regularly?
Gopal Erinjippurath 32:34
Yeah, yeah, great question Aaron, I think like there are a ton of providers of like this maps. And you know, that's like that's kind of input into loan underwriting, or like insurance underwriting in many regions. So depends on the region. So some areas are updated. This one, for example, that we use, was updated every year. So that's kind of the periodicity of updates. And again, now with the changing climate, we probably have real time inputs into that. So I would say it varies, but you can expect the manual update.
Aaron Davitt 33:07
Gopal Erinjippurath 33:10
Any other questions before we dive into Anthony's presentation?
Laney Seigner 33:18
And feel free to keep writing them in the chat if it hasn't. Oh, yeah. Bruno, do you want to go ahead and ask that? And then we'll move on?
Yeah, I'm not sure my question is completely clear. But basically, when when you do that analysis on on, on different assets, and I'm thinking about real estate here, you will identify assets at the highest risk. So should, and you there are a number of reporting schemes that exist. So I'm thinking of Crespi because I know it. But can that information be used by asset managers to be able to report some of their assets, as a twist, where you've had assets?
Gopal Erinjippurath 34:01
Yeah, absolutely grew. I think that's like a target market, the active investing, which is the aggressive kind of reporting, and asset managers really need these kind of tools. So that's kind of one of the verticals in this area.
Laney Seigner 34:19
Okay, lovely. Well, let's turn it over to Anthony. And yeah, and then we'll open it up for a general discussion.
Anthony Atlas 34:27
Great, and I just want to ask one question, how, how technical is the audience, just like the backgrounds group are coming from? Is it a pretty big cross section? Or?
Laney Seigner 34:35
I think, yeah, it's a pretty big cross section. And some people have very technical backgrounds and very data science focused in their careers and others are coming at the climate problem from from other aspects. I mean, I would guess for this talk, it's maybe more of the technical folks in our cohort, but I haven't like closely analyzed everyone who's here, but yeah, it's pretty diverse. And so I mean, any like acronyms, we'll just, let this define them, you know, people can, people can jump in with clarifying questions if they're, if they have any. But yeah.
Anthony Atlas 35:06
Ok great. All right, excellent. So let me let me share my screen here. And so I'm going to give, I guess the way I thought of doing this is just to give to sort of take you through like a specific example, like working with one customer segment, for us. So. So Climate.Ai, I mean, the work that we do, it could obviously, it could impact many different industries, like from agriculture, through timber, and natural resource management. As people have mentioned, sort of, you know, real estate and energy, and there's a lot of other groups. So, as a company, we decided to focus on agriculture, after doing a bit of an exploration and then within agriculture, I'm going to use a presentation that we often leave behind with customers in the seed category, just to give you as much of a kind of concrete maybe, you know, understanding of sort of where, where the, where the rubber meets the road, I guess. But seed companies are really big sector for us, we are also working a lot with processors. So if you think about people who are maybe a processor for potatoes, right, who's gonna work with a ton of growers in some region to produce potatoes, and then it's going to turn those into, you know, chips and french fries, etc. So that's another, you know, potatoes can be one, but we have rice as a few other categories like that. And then we also have a bit of work with, with ag invested in the finance side of it, so people who are buying land or also lending into it, and are thinking about their their risks that way.
Unknown Speaker 36:36
Okay, so I'll just jump, so just a quick background, though. So the company, ClimateAi was started about four years ago at a Stanford. Max and Himanshu are the founders, we have, Himanshu's from India, he led modeling out there for the Indian government. And then Max is grew up in Ecuador, and he's certainly does industrial scale, pineapple farming. And then he went to Harvard did applied math, and then they both met at the GSB, doing kind of a joint program with the GSB, and also the Earth, the Earth System. So and I kind of mentioned that just to show, it's like, there's a really diverse group that we've got different backgrounds and that are, people are bringing in. For us our expertise, and where we've gotten a lot of value is being at that intersection of you know, we have climate scientists, a lot of people with PhDs, postdocs, etc. in climate science, then we've got data science, and we've also got the ag side and by bringing all of those together under one roof, and then going super deep into ag, that's really where we've been able to make a difference for customers and not staying like a superficial sort of higher level, and I think that you, that might be something you've seen a lot of areas of people working in climate is like, you know, they're going for a specific segment, if they can really understand that problem, and then figure out how to apply it there, I think they see a lot of value. Because there are tools and people that are trying to just do things at a much higher level or provide a sort of broad, broad stroke kind of tool that could work with any type of asset or something, I think that that ends up being pretty problematic, from what from what we've seen.
Anthony Atlas 38:07
So the the company that the the key insight, and what was interesting when it got started was basically climate forecasting. And I'm sure some of the other panelists, you know, the sounds like they use climate forecasts and some of the models that are coming out from the different government centers. So right now, it's something that is basically only done by government agencies. So the people who do the climate forecasting, they're typically, you know, you've got NOAA in the US, you've got the European Center, they have teams of hundreds of scientists, and they're basically modeling all the physics of the interactions on the atmosphere and the oceans and the, you know, land surface models and things like that. And so they're really complex models had to be run on supercomputers and so they only release data, you know, new results probably once a month, because it takes a few weeks to run the model, and it cost millions of dollars and then they've got a new set of results. So the interesting attack like approach to this, that climate I took was basically applying machine learning and looking at climate as an image and then starting to bring a lot of like computer vision techniques into it. So when I say that, I mean, you could literally picture the, you know, the surface of the planet with like grid cells drawn all over it, and those are your pixels now. And then instead of the fundamental components being, you know, red, green, and blue, their temperature, pressure and wind speed, and then your, you know, at that point, you're off to the races starting to, you know, instead of detecting a pedestrian that's walking out in front of the car and figuring out is it going to go left or right. In this case, it's a heat wave that's, you know, starting up in one area and then might impact another area. And so I think we're the first private company actually have our own climate forecasting engine that we can run like this and we were able to beat the European Center at forecasting things like El Nino, and some others. And then the real beauty of it though, is that it's lightweight. We can run on a laptop and two GPU. So the flexibility is actually what's sort of opened up everything in trusting that we're that we're doing here.
And so just to give you a high level, I mean, sort of the old way of doing things is like people just looking at a weather forecast. And they are traditionally using historical averages for most of their planning. So this means they're taking a three or five year average, to stick with that potato example, right? The processor might say, Okay, well, I've got orders, I need this many potatoes to fulfill my contracts. So, you know, based on the average, you know, number of, you know, pounds of, you know, potatoes that are produced per acre, this is how many acres I need of the different varieties, and we're going to grow them, okay. And that that's kind of how they make a decision. And the problem is that these historical averages have gotten increasingly bad at, you know, predicting the future and what's going to come. And so they're the variability is just going way up as a result of the shifting climate. And that's really where Climate.Ai focuses is trying to help people understand how weather like fundamentally is impacting their business and how it's like driving the variability in the result and isolating that piece of it. And then creating forecasts that are actually looking at the impact. So we're really like impact modeling for the business. So instead of it just being some forecast, it's what does that actually mean, for the crop and the specific varieties that I'm growing on these farms in these regions. And then given the forecast we have today, and you know, with all the uncertainty in there, how do I optimize the decisions that I need to make at this moment, using the best available information rather than just defaulting to a historical average. And so we work at a range of timescales, everything from the short term weather kind of side of it of days, you know, looking at a week out kind of thing, all the way through months, and like across the season, and then all the way through, you know, decadal forecasts that are looking out, you know, much farther into the future. And I think I hit most, our, so again, we're integrating business data, you know, from the customers that we work with, we obviously have the weather data, and then all the modeling to create these forecasts, and we can do it globally, we can do it from a regional level down to a specific farmer specific asset as people were talking about earlier. And our goal always is to bring the lowest uncertainty that is scientifically possible to forecast the parameters that the the customer cares about.
And oh, and just amount of my role here, if it wasn't clear, so I lead product and growth, it's basically, you know, focus on commercialization in new markets. So a lot of it is sort of the intersection of, you know, what's possible, understanding the customer need kind of trying to go zero to one, and working with the data science team on you know, what we can actually deliver, and then try and actually execute, ramp that up and then kind of bring that into, into new verticals.
And just to give you an idea of the like, sort of overall technology stack. So I mentioned that first piece, which was the first piece on the left, the top left, there was basically that was what we you know, the work that happened at Stanford that sort of put us on the map, we had this breakthrough was pretty interesting, because it was also doing transfer learning from all of the like, simulated, you know, thousands of years of, of climate data, and then transfer learning and applying that on the actual observed data, because we actually don't have that much historical climate data, you know, we've got 40 years of probably observed history. And if you're getting data once a month out of these global climate centers, that's just not that many data points to actually apply, you know, machine learning, to. So that was the first thing that we got into sort of modeling, looking at the impacts on specific crops, this is a little example of us, forecasting strawberry yields about a month or two months out with incredibly high accuracy. And we went from there, we did the long term piece. And so the long term, I think, what's sort of interesting to highlight there is that you have these few of these global climate models that are, you know, coming out of all the different agencies and, and they're not all created equal, right. I mean, everybody's building them and tuning them for different things that they care about, right, so that the Europeans are trying to focus on what's going to impact Europe, you know, broadly, right, India's got theirs, and they're focused on different oscillations and stuff that would impact India. And so they're all better at forecasting different things and in different places. And so unfortunately, when they make a lot of the long term reports, and they say, here's what's going to happen for the future. Typically, they just average all of these models, which is actually not, it doesn't really make that much sense when you start to think about it, because they're not all created equal. Right, some are better and but it's it's not PC also to like, say one country's model is not good at doing a certain thing. And so they end up in this weird thing where averaging is like the only sort of acceptable answer for this. And so what we do is like a dynamically, we dynamically weight them. So it's an intelligent weighting that's based on, you know, what did the forecast say, and then what was observed and then using that to actually assign weights across all of them to create a forecast that ultimately has a lot less uncertainty in it. And so now it can become a lot more actionable because I think as you're all sort of setting off on your, you know, the work in your careers here in climate. One of the big problems is that if the uncertainties are really huge for the future, you might have two totally different potential actions, you would take us a possible outcome. And that's not really that useful to a customer. So trying to narrow that and really hone in on, what's going to happen in in a specific area is really critical. And we did a similar thing with short term weather, basically, to get more accurate short term forecasting, which, again, it has proved quite successful. And then a couple other areas we use machine learning that are just might be interesting to the group, or we do a bit of crop stage monitoring, and like drawing up the growth curves, basically in season. And this is a combination of it's an unsupervised learning approach using imagery. So this is one of the first places that we actually use imagery the way people traditionally think of it. And so it's just taking satellite inputs from many different types of satellites, so it's very redundant, and then it can deal with clouds and things like that. And then using it to kind of fit a growth curve and estimate the the cycles, so you can do that in real time. And if you imagine in an ag space, being able to see at a glance, you know, what is the state of the crop across all of the areas where I'm growing it, I know the weather today, so I have some idea of maybe how that's impacting the yield or the quality or something like that. And then with a forecast coming up, that's not really targeted to their business, and is more accurate, because of the way Climate.Ai does it. You can bring that all together to really say, okay, well, here's what we think the timing is going to look like, when the crops gonna be ready and the order and the impact on the, like I said, yield quality, things like that. So so that's a kind of high level, a lot of the different tech pieces that we have. And I could dive into a lot more of these, but I want to make sure we leave some some time for questions. So I'll probably cut it there. But you know, happy to happy to talk more about sort of any of this and go farther.
Laney Seigner 46:40
Yeah, thanks so much, Anthony, that is fascinating, and a lot of food for thought. And we do have a lot of folks interested in the Food and Ag space related to climate. So I think there should be a lot of questions around that. Yes, zooming back in here. And one of my questions just in the sense of like using this higher, and higher data quality approach to planning and and like all all across the ag supply chain, if you're just getting used to that, like maybe 20 year mindset instead of a one year or five year planning timeframe, and you're in a place like the West Coast, where fires are just an increasing risk in the fall. Are you like, able to have conversations based on these data tools around like, oh, should I consider planting earlier or harvesting? Like, like ways that you know, farmers can mitigate that projected climate risk for their area? Whether it's harvesting, hopefully, before the fire is hit or like, is that something that feeds in at all? Just because it's been so recently on so many growers minds in the West?
Anthony Atlas 47:45
Yeah, I mean, I think specifically for us, it's, it's not that easy to work around five minutes, it's very hard to actually forecast, you know, fire, but you can look at kind of fuels and dryness and things like that. Um, I think that the biggest thing is that the fires and also the COVID situation have really put a spotlight on resiliency and the need to be able to maintain your business's, you know, productivity, irrespective of shocks to either the supply side or the demand side. And that's where they put that at the forefront. I think that's just sped up conversations were, broadly all of our work fits into that of like, how are you going to respond to potential shocks to the system and is your system really fragile or is it something that's actually quite robust? And so I think that that's more how we've seen it factor in and I try to plan on the weather and everything but fires specifically a lot of is just like, get the crop out of the ground as fast as you can if you if you have time. If you think it's if it's coming in there. Otherwise, I mean, with vineyards and stuff, though, grapes are getting damaged from smoke, they get smoke taint, right. And so yeah, if you could, if you could really accurately forecast, you know, the impacts of smoke taint on wine grapes, you can make a lot of money. Getting out there on the on the spot market contracting grapes and knowing which ones were you know, before the testing basically came out, because then the price goes up usually, so you can make a nice return. If anyone has a lead on that talk to us. We could we could floor a project.
Laney Seigner 49:09
Yeah, that would be certainly amazing and good to have. Okay, let's see Bruno, you want to go ahead and ask your question. And then we'll, we'll take a couple like one or two specific to Anthony, and then anyone who had a question for anyone or all of the panelists, feel free to put a plus one in the chat and I'll call on you.
So my question is about so being able to see climate impact at a very local level. Could you also model what would happen if we did something? So for example, it was thinking about the Bay of Bengal around the coast. Could you identify the best places to restore mangrove around that around that coast using your model?
Yeah, that;s an interesting question. I think that as long as there's a good sense of you know, what is the impact of the benefit that comes from taking sort of certain corrective action, then you can say, okay, well, what would that likely mean for the, you know, overall risk in the portfolio of production or something like that, that you might have if you if you made these changes, right. So there's a lot of people are trying to say, hey, I work with all these producers or something, they're all doing different things. And we can identify if you're taking certain actions, and what does that actually mean for the resiliency, right, we can kind of score if you have a measure of something, to then score that and say, well, what does that do overall, for you the farm or that a portfolio level, the resiliency of your production, your supply chain there? And so I think the big, big question would be with the mangroves is just is it well understood, you know, sort of the the impacts that you get from that and sort of how to model that and then you could factor that in to basically, you're just doing optimization at that point I think so. So that's my high level answer, but I think the details are probably where it gets a little bit trickier.
Laney Seigner 50:57
Yeah, thanks for that response. And let's see anyone else? I'm sorry, if I missed a question from up above, or anything like that, but feel free to bring them back up or ask me was, yeah, let me know.
Unknown Speaker 51:08
So can you figure out which land specifically in a particular area is going to be least impacted by fires or climate impact? So like, I mean, could you identify some piece of land that somehow is going to miss like, when there's a big fire is not going to be impacted? As much somehow?
Unknown Speaker 51:28
Yeah, I mean, so extreme risks- so I mentioned this too much in my presentation, but like, forecasting, extreme risk is really critical, right? So the things you can look at and say, Okay, well, if the averages are changing by a small amount, you know, by a degree, let's say, what does that mean for the risk of an extreme heatwave that's over some threshold, and even small changes can be a very large impact on the risk of an extreme. And so we can use the forecast to then sort of disaggregate a for- like a longer term forecast, and then break it up and say, okay, what is the risk of this extreme event that you're interested in? However, there are certain extremes, like some of the stuff people care about, you know, a hurricane or a wildfire, I mean, they're, those are very, very hard to forecast, I think what's a lot easier to say is, you know, if we're looking at a long term trend, which areas are going to be increasingly suitable or less suitable for certain type of production? And what is the probability and the risk of experiencing certain extremes? And then when you look at that over time, and with enough years or enough locations, yes, you can do that with like, you know, your probabilities will be right. It'll be wrong at some spots, but you'll be right more often than not, but specifically, we're gonna say, could you avoid like buying a property in the spot, because it's like less likely to get hit by fire? I think probably at that point, you're a lot better going like super local with somebody really knows the area and figuring out how to just like, you know, looking at where the fire is going to track on the ridge and things like that, and probably like, probably those local features are the most important thing, that would be my guess. But we could definitely say, what's the risk that, you know, you're going to have temperatures and dryness, etc, that would, you know, facilitate that?
Laney Seigner 53:00
Yeah, great question. And thanks for that response. Ritesh, we'll go to you and then Keisha.
Thanks. Yeah, this so my question is for Aaron. Um, you kind of mentioned how your models can kind of applied to places without ground truth. And I'm curious, what are the challenges there? how well that works? If you run into any issues with that, I guess, sort of second part to that question is for anybody on the panel wants to take it. Like, machine learning in different domains has different biases, that we've kind of experienced, or what time how that affects our this facial recognition or voice, you know, voice speech to text, this kind of thing. So I'm curious, if you run into any like, interesting, weird biases based on data in this in this climate related domains? Because I haven't really, I'm not really sure how that would manifest itself in this domain. I'm curious if you run into anything like that, and how you know.
Aaron Davitt 54:07
So okay, I'm going to start with the second question and work back to the first one. So I'm not a I'm a remote sensing scientist not a data scientist but the some of the biases that the machine learning people have talked about is actually with satellite data, because some of the satellites acquired data at a certain time of the day. So a lot of models are trained on that specific time. So then, that can create a bias of like, like, mainly morning, like you were trying to predict emissions in the morning, but then what do you do when you're trying to get in the afternoon? So we're trying to remedy that with other sensors that can like then help train the models to count for this morning bias that is current. That's the extent of my bias with it so it feel free to use Tell me if you have any questions about that. The first one is had to do it with ground data like areas without ground data. So I mean, the whole point of low comparing where we have satellite data and granddad with a modeling is to create as robust tool as possible that can just be applied to somewhere else. But we're always looking for collaborators and people that are willing to share ground data in areas that we don't have, in order just to vet and to make sure our approach is correct with estimated greenhouse gas emissions. But from a satellite perspective, you can also treat satellite data, as instance, in situ ground data, depending on how you use it. So in some cases, we use that as a way to get a sense of what's going on and location without that specific ground data.
Sorry, can I that was my question. So and the reason why I was asking is because I've worked with air pollution data, some satellite data, some like model ground based model, model data. But I know that like, the the satellite models are, are trained on some of the ground based observations, but like, we can't apply those to Africa, necessarily. I mean, they get applied to Africa, but like, they don't, they don't work that well. So I was just wondering, like, if there was, I guess, something I didn't know, like, some some clever trick that people were doing in order to make that work.
Aaron Davitt 56:38
I was gonna ask you do you know, the secret handshake or a clever trick? But I mean, no, I mean, those are things like, you know, some some of those models you've seen in papers out there trained on, you know, United States or Australia or Europe, but then we switch over to some a country in Africa and or South America, you get errors associated with it. So there's some things that need counted for. Currently, we are working on that, and looking at those two, reaching out to collaborators too.
Right, and then how do you deal with like, I mean, it seems like you guys get data pretty, pretty quickly but like, do you have the same issue with aerosol optical depth? Like if it's cloudy, for example? I mean, that's just I know, that's, it'd be more of an issue for air pollution then than other things.
Aaron Davitt 57:22
Yeah. I mean, clouds are course on issue with optical infrared. But my background is radar, I use SAR like Sentinel-1 for most of my research. So this question is like, can you use that as a way to, indeed do some type of indication of emissions. And, of course, radar does not measure emissions, because of the wavelength and everything. But can you use it to look at you know, activity in some factory, based on machinery, things like that, or some type of other indicators that can be linked then to the missions that I mean, so we're, those there's other satellites there that could be that cuts through the clouds, but then you have to do an extra step of trying to link it to something meaningful.
Anthony Atlas 58:08
And I'll just add, as a two second thing, that we we spend a lot of time doing bias correction after the fact. So especially like, in our case, we use we do an unsupervised learning approach, the bias doesn't come in, you know, early on with the way that, you know, let's say you're labeling the data set or anything, but what you ultimately end up with is something that needs to be shifted around at the end. So you've got a good sort of objective measure. But then as far as, what is this thing that we're detecting actually mean as far as what you see on the ground, and then when we get that adjustment factor, then we can kind of shift everything and we can make a lot more accurate. And that applies, as well, to forecasts and temperature and things like that, when you want to take into account you know, or local variations in a microclimate or something.
Right. I mean, like, one another question I have is, is like, how much does it matter? Like so for some of the air pollution stuff, ike it doesn't it's frustrating that you don't have like detailed information on the ground for like, everywhere, but I, in some cases, it doesn't necessarily matter that the data is perfect. Like, I don't know, I'm guessing that for other things like climate, climate projections, like it must matter quite a bit.
Anthony Atlas 59:11
Yeah, I feel like that all that all depends on the customer, right? It's a you don't need to go overkill, but you need to get to a good enough to be meaningful, where you're going to have a real ROI cuz you're trying to get people to change their behavior, right? So you're trying to like what is that bar? So I know what that bar is for us and for all of our customer segments, but imagine Aaron, do you ever have people come back to you and they're just like, no, this is this is too you know, you're going to detail this is too much like we don't...
Aaron Davitt 59:34
I mean, we're working on multiple levels of detail but yeah, we're trying to figure out what detail to go at like is this like, just do like a city level set is like is that good enough? Or do we need to go really down fine detail so it depends on the impact we want to try to make with it.
Anthony Atlas 59:50
Sophistication right, of the probably the person you're working with I'm because that's I think it's very important for everyone to remember that like the world the wider world out there is using very rude tools to make very important decisions right now and so the state of the art might be a very low bar, that's a good place to start. There are people even for us that we could sell to that are far more sophisticated, but like, why would we start with them? Right? they'll, they'll be the most demanding customer. And, you know, you can work up towards that. But yeah, the world is not being run with like, you know, super advanced, insightful ways of doing all this stuff. It's it's kind of how it works.
Aaron Davitt 1:00:23
Like just the continued, some countries are just starting to get into remote sensing as like the last couple of years, like there's a whole UN/NASA training to get sustainability goals to actually get countries to use some of the free data from these from NASA to help inform, some countries still have, you have to order the map, which is you know, printed by hand, and it has to be geo referenced in everything. So there's a push trying to get countries to use more of this information. So that's like a starting point, with some of them.
Anthony Atlas 1:00:55
And I just saw a quick question here about locusts, I'll just flag that. Yes, we haven't done this, but it's come up. And we, you basically this, if you could tie it back to the climate and the weather, then you have a shot at doing this and so we do look at forecasting risks of pests and diseases and how that's going to change as the climate changes and what that looks like, both in season and longer term. And so I think the germination cycles are related to temperature and a few other things that are pretty critical here. And so you could look at how that risk factor would be going up and up over time and then what's the likelihood of having a big outbreak, which I know has been, you know, decimated places like Kenya and some others right now.
Laney Seigner 1:01:28
Yeah.Yeah. Yeah. That's cool to be able to use that for that. So we are this past time a little bit. So we so I'm just going to ask that if anyone has a question that didn't get answered, you can just share it with me maybe, or put it in the chat and all I'll save it. And if it's okay, with the three of you, panelists, for me to send you an email with like, any two or three lingering follow up questions, I'll just do that. But yeah, is that ok?
Anthony Atlas 1:01:53
Ya. And then the same thing, I think the other people said, you know, we're, we're hiring, we also take a lot of, you know, we take interns, we do all things like that. And we're kind of ramping up, I think we're hiring on base every single team right now. So if there's something that interests you, or maybe a collaboration, anything like that we do, we're also very, we like to work with others and try to share things that's kind of in our DNA. So yeah, please keep us in mind.
Aaron Davitt 1:02:14
Yes, data with ClimateTRACE and everything. collaborators, people that are interested in working, you know, feel free to reach out.
Laney Seigner 1:02:20
Yeah. And is global in both of your cases, or all of your cases, or is it mostly hiring in the US or collaboration in the US? Just because our cohort is global.
Aaron Davitt 1:02:31
We are remote right now, so I imagine we are global in terms of organizations in terms of specific Watttime, I think it needs to be within conus. But you know, I think everything you thing's negotiable during the COVID time.
Anthony Atlas 1:02:45
We're pretty global. We got team members in Europe and South America and Asia, things like that.
Gopal Erinjippurath 1:02:50
Cool. Yeah, same here remote and global. You have to be friendly to those two in the times of this pandemic.
Laney Seigner 1:02:58
Anthony Atlas 1:02:58
Yeah, we could have all different call on how to manage these timezone issues.
Can I just ask one question, I actually look at all your website, right. And I have no idea of what kind of software or technology stack you actually are using behind. And so I don't know, the you have a need for software engineer? Or is it more data engineer.
Anthony Atlas 1:03:26
Aaron Davitt 1:03:48
Yeah, for the most part as too. Python, cloud processing, Java, things like that.
Laney Seigner 1:03:56
Cool. Um, okay. Well, so I yeah, I think a lot of you share your email that the last slide or something. So I'll maybe also just follow up and if it's okay to share some of those slides with that, that contact info. I'll post that in our group community channel. And yeah, I look forward to any follow ups and things that stem from this, but it's been a fascinating hour and five minutes, and I really learned a lot. So thank you all for your time.
Anthony Atlas 1:04:23
Thank you. Thanks for having us. Appreciate it. I just stuck my email in there. A couple people replied to me. I think the whole chat maybe can see that Tim so. Oh,
Laney Seigner 1:04:33
yeah. Awesome. Thank you all, everyone. Take care, guys. Bye bye.
Our August 17th cohort is full. Next cohort starts in
* If you work at a non-profit, let us know when you apply