This episode we get to talk with Kelsey Josund, a machine learning engineer from Pachama. Pachama is an exciting green startup on a mission to solve climate change. Pachama is making progress towards this lofty goal through a carbon credits program where companies and individuals can purchase carbon credits to offset their carbon usage. Pachama uses machine learning to monitor and measure carbon output from forests around the world to be sold as credits. On the episode we talk with Kelsey about how this process works and what the future of carbon monitoring looks like.
Connect with Kelsey and learn more about her work here:
Learn more about Pachama by following the link below:
Interested in learning more about carbon monitoring? Check out some of our articles below:
Deep: Hi there I'm Deep Dhillon. Welcome to your AI injection, the podcast where we discuss state-of-the-art techniques and artificial intelligence with a focus on how these capabilities are used to transform organizations, making them more efficient, impactful, and successful.
So hi Kelsey. Thanks so much for being here. Please set the context for our discussion. So what's the problem you all at Pachama are solving.
Kelsey: Uh, we have Pachama work on verifying forest carbon credits, which means we work in a space of nature-based solutions for carbon capture. And so the problem we're solving is verifying that these forests are as healthy and robust and real as everybody wants them to be.
Deep: As just a regular person, what does it mean to offset your carbon? What are the credits and how does that link up to what you guys are monitoring? So at a basic level, everything we do all the time is emitting carbon, and we need to be not doing that so much, but forests and all kinds of ecosystems do a really great job of. Taking carbon out of the air through photosynthesis as these plants grow. So the idea behind a carbon credit is if you can quantify how much a forest is growing, you can figure out how much carbon is being sucked out of the air. By that forest, where this gets tricky is in order to be a real offset, it needs to be a forest that wouldn't have grown without that.
Deep: I was just going to say, like, how do you know. But this wasn't just going to be there anyway.
Kelsey: There's several parts to our verification process, but part of it is we need to make sure that it's real, like the forest that people claim is there is actually there and you can make sure that it's additional. That means that the forest is either healthier or. Bigger or trees are growing back in a healthier or faster way, or the deforestation they're claiming to prevent is actually being prevented. And that if what they're claiming is preventing deforestation, that it's not just like going somewhere else, then we need to also make sure that it's permanent, meaning that we can understand like the risk of it being cut down or burning down or being hit by. And take that into account when we're crediting it. So that if that were to happen, the credits could be redirected to a different forest or something like that. And then verifiable meaning we can tell what kinds of trees are there and what kind of an ecosystem it is. And therefore can model how much carbon is in the trees, because generic forest, like the kind of forest you have in California versus Brazil have very different characteristics of how much carbon is stored in them. And then. Finally things about like impact beyond carbon, which is not modeled with AI. That part is just like our experts on the ground. Making sure, basically that when you have a forest project, it's doing good things for the forest, but you don't want it to be coming at the expense of an indigenous community or something. So we want to make sure that the projects are. As a whole, rather than just looking at the carbon.
Deep: Got it. Here's my ideal vision, right? Like I want to be able to offset my carbon. I want to put in, let's say, you know, for every X dollars I put in, I would know exactly what little. Plot of land and trees are what chunk of a tree I actually own. And I'd be able to, you know, again, in a perfect world to have a real-time camera stream or something pointed right at it. And I would be able to like track that back to like, uh, um, like a piece of land. And then in the other direction, you know, I actually might go out and buy a plot of land. Put it into your system somehow, and then go back and feed someone else like me and say, Hey, look, I got this plot of land that had a, I was going to have a Walmart go up or something. And I managed to like buy it and stead, uh, do I like, how do I plug in? So maybe just fill us in on those two sides.
Kelsey: Yeah. So both of these are things that our solutions can help enable, but that the current status quo doesn't support at all. So I guess I can talk a little bit about like how this works without us, which is that. When you have a new plot of land that you want to turn into a carbon project, people have to go there, measure all the trees. And to some extent, that kind of thing still needs to happen because we have to be sure that the kinds of trees are what they say they are, but then they have to do that again every five years to make sure the trees are still there. So that's the monitoring. Which it often is less frequent than every five years. It could be every 30 years or you can just be taking like the landholders word for it. So our justification is that we use remote sensing to monitor that. So after the initial step of being like the trees are this kind of tree. Then we can continue to monitor consistently using satellites. So it's not quite a live stream. Although we have talked about doing that for certain places where it'd be very cool and more accessible to do with satellites, you can continually monitoring all of these places. It's not quite as detailed as a live stream, but it's still like a frequent update, probably about monthly have high quality data. Can't do it too much more frequently than that for a lot of them. Cause if things like cloud cover.
Deep: So one of my challenges I have when I look at sites like this and I think, okay, I'm going to buy some offset. They'll point me typically to like some big, huge plot of land. And maybe I see some pictures of it, but I don't actually know what I own, right. What is it I'm buying and how do I wrap my head around it? As opposed to just being pointed to a picture of a forest?
Kelsey: Yeah. So it is a little bit more abstract because you're buying a credit. And part of the reason that we talk about it in those terms is because then you can more easily compare for us of super different types. And also the piece I mentioned about like the risk to the forest, meaning even if like two forests have the same. Biomass per hectare. If one of them is at much greater risk of deforestation, you're going to need to actually. Protect more hectares of that land as a buffer against that risk. And so we can't just say for X number of dollars, you protected Y amount of land because it's going to vary greatly depending on what land it is. Um, you're completely right. That people want to see the degree of specificity. So that's a continual I think challenge on our like marketing side of the storytelling piece of how do you, how do you translate this kind of abstract thing into. It's something that is compelling to people
Deep: Let's dig into both those sides on the risk side, you said something really interesting, like, okay. So for a given plot of land, you have to somehow assess its risk of basically not being a carbon sink. Yeah. Yeah. How do you quantify that?
Kelsey: And there are industry standards of like how much buffer you need to apply. And then the frequency of monitoring, which does allow things to slip through the cracks sometimes. So we increase the frequency of monitoring so that if a forest from which credits have been issued, it gets destroyed and we can notify more quickly. And that will lead to reissuing either redrawing, project boundaries, meaning protecting different land or something or ceasing to issue. Or like shifting where that money has gone, I can go to all kinds of things. And this is an industry that's continuously changing. Meaning like if a forest burns down, maybe the Landstuhl needs to be protected because that land can still regrow.
Deep: Let's say you've got an acre of land inside, one of the California forest that burned this summer. Okay. So for that acre of land, you associate that specifically with a certain number of credits?
Deep: and then once it burns down, do you delete those credits?
Kelsey: It will really depend. On like the specifics of what has happened. So when we're modeling biomass, we're looking at how much is actually in the trees at a given time. But what we're looking at to assess is like trends and differences between the protected land or is unprotected land. Because as long as the tree is alive, The biomass is in that tree. But when the tree dies, that biomass is released into the atmosphere through decomposition or burning. But if it's a protected forest, the total carbon of that stand of trees should remain constant. If an ecosystem can recover, the carbon credits are still valid in the sense that we're not protecting any specific tree or protecting an ecosystem. But if it gets to a point where the ecosystem itself is destroyed, then. That's different. Oh, so like a fire that comes through burning super hot and like obliterates things. It's very different from a fire that comes through and thins, the trees, and they can come back healthily. Like that's some of the complexity and like this monitoring piece, but the biggest aspect that we're looking at often is actually just deforestation.
Deep: Tell me pre-AI, what exactly happened?
Kelsey: Like the very fundamental level of it is like destructive sampling, take a tree and burn it in a special. Sensor to come up with the carbon that's destroyed. Obviously can't do that at scale, because that is exactly the opposite. So from there they can derive what are called allometric equations, which basically are just pretty simple, like matching equations to like fit. That measurement to measurements of trees. And you can go based on like tree density and size. And so those equations will be derived for tree species for tree like locations. There are tons of different ones, and there's a lot in the literature about like which one you should be using. And generally you want to use a specific as you can, to what you have, like what kind of tree, both the species and the location. And they only take into account. Often the above ground biomass, but sometimes they also include the roots. Sometimes you also need to somehow take into account the other biomass in the forest, beyond the trees, but basically you go in, you'll do an inventory measure all the trees, plug that into the elementary equation, pick the correct equation, add the corrective factors for whether the equation is. Into account things other than above ground biomass entries, and then you get a number and that's the biomass for some area. And so it's very manual and very time consuming and there's a lot of room for error. So that's one of the biggest challenges that we face in applying. Machine learning to this is that grandkids data is not that clean. And it's also very expensive to collect. So it's not a place like, you know, you can just go scrape the entire internet and train GP T3 on.
Deep: I'm guessing it's not standardized either like different governments, higher, probably independent consultants or something that do this sort of thing.
Kelsey: Yeah. And so there's a lot of different datasets and a lot of trying to understand the characteristics of those datasets and. From there, figuring out if we need to apply different corrective factors and how do we use it and how much do we trust it? And then another big difficulty is that a lot of the time the sampling is done in healthy forests and governments will want to understand what their old growth forests are like. And so they'll go in and do all these measurements there, which means then when we're trying to model, you know, a second growth forest or new planting, it's very difficult because we don't have that much data that reflects. Since the expensive high quality data tends to be from national parks and things like that.
Deep: So are you mostly buying that data?
Kelsey: So a lot of our data is from government sources, so we don't have to buy it, but we do have to like get access to it, which can take some work. Lots of universities have some like experimental forests. We do buy some data, but that actually tends to be our satellite data that we buy. Usually we can get access to like tree measurement. Data through like partnerships and things. And I know definitely like where this is going, we're going to have to start collecting our own data. Um, but at this point we haven't yet begun doing that.
Deep: And tell me a little bit about walk us through the modeling process. Like, are you directly straight up trying to predict the targets that are coming out of this?
Kelsey: So there's a lot of things that have been tried and like the kind of industry standard approach for this is to you. LIDAR as the input where you have LIDAR, which is not everywhere. So LIDAR is basically point cloud data that describes like in 3d, the shape of something it's often used in self-driving cars, this case it's for forests, but yeah, that can be either space born or airborne. And we have played around with both. Again space when you're going to get it in more places, but it's going to be less fine grained versus from a drone, you can get really high resolution, but in very specific places. So there are trade-offs. So the standard is to use that as your input and then predict biomass from that. This means that you can usually get very high quality. For places where you have good LIDAR data, because the LIDAR data is essentially giving you the inputs to those allometric equations. I mentioned, right? It's telling you the tree measurements, that's what the LIDAR data can do. But the downside is like I mentioned, you may not have it everywhere. And so then people will tend to use. Just optical satellite imagery as an input, and then predict biomass from that. And then you can train corrective models between these. So basically you'll be able to get a biomass model in more places based on optical imagery as your input. But it won't be as good because the imagery can't capture all the things you want it to you. And you're not just using the way the images look to the human eye are often using like ratios between the bands that are known scientifically to capture information about plants or about soil moisture. And so on. We also use various other inputs that are not direct, like captures from like the time we're observing. So this might be like, Uh, weather observations over a long period of time so that we know basic connecting information about where we are trying to predict the biomass, because that can tell us about like physically the potential and a model trained just on those things can give you a pretty good, like. Estimate of biomass. It can't tell you anything specific about the current time. It'll just tell you, like, if there was a super healthy forest here, how much biomass might it have, and that kind of thing can help them model quite a bit.
Deep: You're listening to your AI injection brought to you by that. That's x-y-o-n-i-x.com. Check out our website for more content, or if you need help injecting AI into your organization.
Help me wrap my head around this. So like, if I'm looking at LIDAR data of the rain forest, I'm basically just able to tell that it's a rain forest. Is the model really learning like the girth of particular trees and all that? Or are you ultimately relying on these. Samples that you actually kind of rigorously, manually measured.
Kelsey: So it depends on the how granular the LIDAR imagery is. And they might've been like the pressing applied often does, like in the presence of Rasta rising this point in cloud, you often end up with it being like re sampled to a much lower resolution. So like larger pixel size, which means you're right. You can't tell as much detail, but the raw point cloud can. Tons of information. Like I've seen images where you can see like the shapes of every single layer and like a Cedar tree or something. And if you have both the enough data and the data at all, and enough compute and you know, all the labels you would need to use that data, you could potentially have extremely high performance LIDAR models. The issue was always just having enough of the data of what the total. Which is the biomass of a tree. Like if you can't geo locate your LIDAR imagery and your labels, and that is like the biomass measurement finally enough, then it's not useful. Right? So that's part of why you would, you know, smooth it out a little bit because that way, even if you get the tree offset by a few meters, you still know the trees in that big meet the couple of meter area, and then you can use it. But the biomass is so highly concentrated in the very large trees that. If it works out,
Deep: I mean, is that ultimately what this comes down to is counting the large trees.
Kelsey: Um, it's not quite that simple sense that wouldn't allow you to capture things like regrowth and things like that. But I think that's, that's definitely a big part of it. It can be really astonishing when we're looking at, are they different? A couple of times when I first started working on this kind of project where I would look at our labels and be like, But, but this couple square meter section, that's an impossible number. And then our forest science extra be like, no, that's just where the old growth tree is. Like that number is so huge compared to all the things around it, because there's one giant tree, you know? So they're, the outliers can contain a lot of the biomass. So, I mean, at the end of the day, like how accurate is it? I mean, that is the million dollar question. The level of uncertainty capture in all of these steps is so high. Basically the thing we look for in accuracy is matching up with the existing state of the art, which as I mentioned earlier, the measurements through the elementary equations and the destructive testing, even that has a lot of error. So we want to match that as well as we can. And then what we really want to do. We'll look at trends. So we want to make sure that if our model that detects deforestation and detecting deforestation, that our biomass model is correctly saying that biomass has been lost. If we see if you run this thing over. You know, a forest that we know is second growth versus old growth. We want to be sure that it's getting the correct trends. So basically you want to be sure that all of the trends are matching up and then that the. The influence of those environmental potential variables are correctly being reflected. Like we want to be sure that we're not just predicting based on potential. And as long as all of those things are happening, then the model is pretty trustworthy. But all of those things. Are very hard to match up with.
Deep: but like you must've been able to measure your stuff on a very particular scenario.
Kelsey: Yeah. We have accuracy bars that we require ourselves to meet before we can. Say that the model is good enough to use for evaluating a project. Basically we have all these, these field plots, all these measurements, all this LIDAR data, all of this optical data, all of this radar data, which we didn't even mention, but sometimes we use that to, and. All of these things. And then ultimately we need to be able to build a model and then run it somewhere else because the places where we have these measurements are not the places that we're concerned about. Right. Cause we're looking for the potential new reforestation places where new protected, existing forest, not the places that have already been investigated really heavily. So we want to be able to determine is the model good enough to run into specific place? So that's the other factor. You can't just say the model is this good? Because running on. Like a test split that comes out of your training data, which is pretty standard practice that doesn't work very well for this kind of model, because everything is very spatially auto correlated, which means if you just take a chunk of your training set and use it, you can get really amazing numbers, but they don't really necessarily mean anything. If the forest that you need to evaluate, isn't specially auto correlated with your training data. There's a lot of. Basically just a whole lot of plots that we need to do. We need to inspect, to look at the trends and all these different ways to assess whether we can trust any of the quantitative numbers that we're getting, because that qualitative assessment is still a huge part of determining whether the models are viable.
Deep: So from a customer standpoint, it might be depending on the customer, but like, you know, you've got end consumers, then you have corporations who might have quite exacting standards. But, um, bottom line is, you know, if some cattle farmer and Botswana agreed to reforest their land, you're going to know when they started sneaking their cows. But. Do you feel like as of today you're able to satisfy those stakeholders? Or do you feel like you're at the very beginning of a very long road or, or kind of both, or
Kelsey: I definitely think it's both right. There are places in the world for which we have sufficient data and experience to be able to do this well. And there are places that we don't yet. We're definitely. In a really exciting place because of that fact, like the fact that we know we can do it, but we haven't yet is really exciting. And most of what we do right now is building on top of existing carbon registries, which means someone else losing a more traditional approach has already verified. The project and we are doing an additional step of verification. We don't accept all projects from all registries. I don't know our rejection level, but it's pretty high. There's a lot of projects that we take into consideration. And then we determine the amount of verification has already happened is not sufficient. So our standards, because we're able to apply this more frequent monitoring and more detailed monitoring are seeing problems with things. So we try to give that feedback back to product developers so people can improve. And then. Our next step, what we're really working towards next year and starting now is our own projects, which we will like co evaluate rather than being a secondary step.
Deep: Do you have like a region where you're going to track carbon and therefore you have to kind of keep going with your data and your sampling and everything you need to do until you reach a point where you're confident in the analysis. And at that point, the next unit of land becomes. Sort of accessible to your customers and transparent and everything, and now it's being monitored properly. And now your team moves on to the next, you know, plot of land. Is that in essence though? What do you work and doing it...
Kelsey: So we don't do it in as orderly a fashion is that we tend to kind of go where the data is and where the projects are rather than us being. Here's an area we want to roll out. I have heard talk that's maybe we want to start doing, and as we scale, hopefully we'll be able to, but for now, it's like we have this great dataset that some university is willing to let us use.
Or some other researcher met somebody at a conference and is going to like, let us use their data, you know? And then we can build a cool model for somewhere that we previously couldn't or we'll have. Company be like, Hey, we want to support forest projects, close to our headquarters and do you have any? And we don't. And so we find one and we have to verify it. We can just go to find data like that kind of thing. And I don't know, like the level of specificity that goes into like all of those decisions, but definitely for now, it's like finding. Data and projects both pretty much described almost every machine learning project. Like we all, we all do that all the time. Right. It's just different in your case. Cause it actually, you know, it actually translates into something interesting, like for us on the ground, I guess it makes it feel different to me is that it's like physical place. I literally mean where is the data like Chandra?
Deep: Yeah, no, I think that's actually, it's actually funny because you know, as data scientists we're so. Comfortable in the non real world. I can't in this abstract intellectual land. Here's a weird question. Have you thought about tapping into private landowners, you've got apps or something and you, you walk your property and maybe you actually tap into them somehow as users to actually generate stats for you to know that you don't have to, you know, go off and send, you know, highly probably what are expensive teams to go to, to take measurements and stuff.
Kelsey: So, if you, if you want to ask, have you thought about doing this cool thing with forests? We definitely have most of them, but you know, some of the new iPhones have LIDAR. How cool would it be to have, you know, citizen science teams go out and do that? And then of course our forest science expert will be like citizen science can be dangerous. People can do things wrong, you know, like there's so many cool things and there's so much satellite data that we don't yet. use. And we could do things with it. And I can't, I can't even explain how many cool projects and problems that are still in this space to solve. And for now we're just doing forests, but like grassland also captures carbon and algae does like, maybe we could do that kind of thing. We have no plans for any of these yet, but the possibilities are endless here.
Deep: I totally appreciate, you know, your forest experts saying that sending out. You know, individuals has got his own set of problems, but isn't that kind of what always happens with machine learning like that. There's different levels of it's rarely the case where you can just go out and get easy ground truth. You know, those problems, if they're intersected with value, tend to have already been taken up, you know, you can just measure efficacy of your humans as well as you can measure the efficacy of your models. So you could. You know, send your, you really do have so many people that really want to do something about the state of the planet. Right now. If you organize teams, they go out, they take measurements. Then you could measure them against your actual ground truth and use.
Kelsey: If we could develop an uncertainty model of the humans collecting the data theoretical thing that could be done, there's truly like unlimited number of cool projects that could be done with the fact that we can collect more data.
Deep: And does the demand for your credits?
Kelsey: Right now, I think we're more bound by supply than demand actually, because we're kind of in this new space of this high frequency and high fidelity monitoring of this, a lot of projects don't yet pass our bar. It's definitely a matter of they're more big companies, mostly who want to be spending money on. Real carbon offsets. And it's hard to find enough. We need to tap into this market before they get frustrated and go away. I, and also we need to, you know, for. The planet's sake, you know, most people at Pachama, aren't just like the people who want to protect forests for any reason. We, even if the money wasn't involved, but like that's not the world we live in, we need to find ways to direct money to these projects so that they will happen.
Deep: The scale of this problem is. Yeah, daunting. Right. It's just daunting. And then the timescale with, within which we have to solve it is comparably daunting. One question I have for you is these corporate buyers. Are they the ones that are like pushing you for like higher and higher efficacy in your modeling practice and your quantification of the carbon that's actually being offset for a particular area?
Kelsey: Yeah, I won't say no, because I think they definitely do have teams that are pushing on that. I know someone specific who, um, They have like very specific ideas of how they want us to do this, but more what I feel on the machine learning side is the, you know, university or government entities putting out papers in the field, who we need to look at what they're doing and compare ourselves to them and see if our stuff matches. Sometimes those will be a more tractable problems because they will. You know, they're not trying to do any just anywhere. They usually have a specific data that they're trying to model, but we need to be able to match their performance. At least if we have that and we need to be able to explain why we can't, if we can't, you know, that kind of thing. So that's where I personally really feel it in my work, but I think it's coming from all sides. A lot of governments are working on very similar problems and also. There are both voluntary and government run credit markets. And I don't know what kinds of standards are in like being enforced within the government ones. But a lot of the papers that we're reading are coming out of people who are in government positions.
Deep: For the sake of our listeners and me friendly. Cause I don't know all, but like tell me how do the carbon markets work? What's the equivalent of the NASDAQ and the Dow in my, what are you buying? And like, are there ETF equivalents? And like how, how does that whole thing work?
Kelsey: It's simultaneously like a wild west and also extremely monitored. Everyone playing in the field. Super invested in it being real and wanting it to work, but there isn't like a single entity, like the equivalent. Yeah, it's monitoring it all. And so there's lots of big players. Like the one that I followed the longest, I guess, is Vera. They don't do things that are technically similar to us, but we piggyback off of a lot of their more manual work. And they've been working in this area for a really long time. And then of course not all carbon. Credits are nature-based like we're talking about that are green energy projects, all kinds of things. And so that's a whole other can of worms of how much carbon emissions are being avoided through that kind of project. And there are other organizations and probably companies trying to make that more. Legit, but by the whole thing is just a need that is not being filled. But to your point about like, is there a criminal I'm like ETFs, they're just, there can be bundles where, I mean, this is something that we sell is like, rather than buying, we were envisioning initially like a very specific thing. It's more. Here's a chunk of money, get me some carbon credits with it. So some companies will be really specific about who they want to support, whether that's location or type of forest, whether some people are very particular about whether they want reforestation or avoided deforestation. And I honestly, I guess it's like an aesthetic choice. I don't know. You know, the more balanced portfolio wouldn't be to be doing some of each in different parts of the world, because they all have different kinds of risks and different kinds of uncertainties. And if you diversify it probably doing better.
Deep: I mean, I get the idea of companies offsetting and buying these credits, but are they selling them to. You know, are they, are their investors trying to buy a certain type of credit?
Kelsey: I don't know
Deep: if the demand is low and the price starts spiking. At the end of the day, that's going to translate into more dollars available to, you know, some or somewhere to
Kelsey: As long as it does. Actually like it doesn't just involve a bunch of money, you know, like ticket, scalper style going into the wrong person's pocket while there, rather than the creator. It's definitely possible. That would be good. I think the real fear is that prices inflate. And then the things that were being speculated on reveal, like ended up being not very high quality credits and then their big, scary headlines around that. And industry is unregulated as this, when there are. Things that are not being properly vetted, or even if we're being people that are attempting, but don't know what they're doing, you know, that kind of thing.
Deep: And you want transparency around the credit. And you're basically saying, look, this credit connects with this forest, which has this, you know, audit, capable, uh, trail behind it and this commitment to a future trail behind it. So you basically. Validating the worth of that credit in essence,
Kelsey: That's what exactly we are doing. And we just hope that we can scale that up to meet the level of need. And that's, I think why supply and demand is really weird to talk about in here, because ultimately we keep coming back to like, Obviously the supply and demand matter for us as a company, but they don't matter for the world. The world needs this, whether there is supply or demand. And as the big biodiversity, carbon is only one of our big problems that this kind of solution can solve. Like switching. To using a carbon credit that's involves switching to green energy. Doesn't address the biodiversity crisis, whereas nature-based carbon credits do.
Deep: Yeah. I mean, you could imagine a world where let's say you're wildly successful. These credits are worth a lot and they increase in value and it so much so that people, instead of cutting down a region, Um, or leaving it alone and allowing it to maintain biodiversity, start planning, bed dances, trees imaginable in the most dense fashion imaginable, in which case, like no living species care about it, but that would still be effective.
Kelsey: There are definitely, yeah, there are definitely ways things can be done poorly. And they're also, that's why it's important that we have that final step of like, is this, it passes all of our carbon. Checks, but does it also have other good side effects? You don't want just like a giant tree plantation to be the only thing that you ever are growing? Cause I mean, yeah, trees, a tree plantation is better than a strip mall, but it's not as good as it could be. So if that lands could otherwise be developed in a better way. Yeah. I think that's ecosystems. Aren't. Optimized for carbon extraction. It's maybe a factor that happens to be, I mean, I do think there are many places and I'm not an expert on this part, but it seems from what I've seen, that a lot of cases, if you let the forest manage itself or let indigenous communities manage the forest, it does better than, than like what we would guess would be the best. Because if you grow at super dense, it often will have higher rates of disease. Be at greater risk of forest fires. And so the thing that short-term optimizing for just carbon, isn't actually going to optimize for, you know, a 50 year scale carbon. So tell me, like, what are the big AI problems, kicking it back to the machine learning side?
Deep: What are the big AI problems in this space that you think represent the next five to 10 years of effort?
Kelsey: I think we are going to continue to get more different kinds of satellites. And that's inevitable. There are so many satellites being launched at the time, finding how to integrate that information and use it productively for this project. Like in my dream world, we would have both enough satellite information. And, you know, like labeled data to, to be able to model biomass at like a total scale, like everywhere. I don't think that's impossible. I think it's definitely possible. I don't know how likely it is because they don't think it's a priority. I mean, when you launch a satellite, they're often like telecommunications, they're not, they don't have this kind of camera on board. Even the ones that do are not always, you know, fully available, but I definitely think that's a thing that could happen and it'd be super cool if it did. I also think expanding the same general technology that is currently being used. So less pie in the sky is taking. The current approaches and just finding ways to correctly and forever improving how we monitor the risk. Because like I mentioned, there's an uncertainty, there's all the things that can hurt a forest and all the things that we don't know and all of the messiness in our labeled data. So if we can come up with clever ways of. Understanding all of the places where that breaks down, like we mentioned before, some theoretical model to tell how bad the people were at getting the data, then we could get more data. Right. If we knew that. And so I'm sure there are ways we can plug in. More things.
Deep: So here's a question that kind of, this one popped in my head. I was, um, driving through rural Botswana and I just can't, I can't help, but wonder for that one farmer. Who's going to burn down that one acre of land. Like, what is the dollar cost of that acre of land and, you know, to that farmer cause these aren't a lot of these folks are not super wealthy people. It just doesn't seem like it's that expensive for us to, you know, take the model, take your carbon credit valuation. Figure out what that stuff's worth, maybe pricing some, some biodiversity value in there and have it. So the rest of us in the, you know, are buying these things on the wealthy side of the world. And those that farmer is literally just getting paid to not cut down their phones.
Kelsey: Yeah. So I think that's like, you know, democratization of carbon credits is a big part of it, like right now. And all of the projects that are running on the major registries are pretty big because that's what makes sense economically when you have to go inspect them. But if we can get to a point where we can do this using a fully remote sensing model, we could be assessing the value of the land right next to one farmer. Property and offering him whatever amount and giving him the skills and the incentive to protect it. And I think it's made easier by the fact that a lot of these people already do love that land just as much as you know, a tourist there. And they don't want to have to cut it down. But when given a choice between. Eating and not like, it's not a fair position that we put people in.
Deep: So how far away do you think that time is where, you know, your team can like go into a map of the world, draw boundaries around areas that you're dying to protect and then price that land and route cash to those folks. And just literally by the forestry rights or whatever.
Kelsey: Yeah. I suspect it will vary a lot by location in the world, both based on like having government support for that sort of thing and community support. Like if it's in a place where we have been able to establish trust, that kind of thing, we'll go in more quickly. And this was something that was talked about by some of the protestors at cop 26 last week was about. This whole issue of it's like colonialism, the rich countries coming in and telling indigenous communities, don't cut your forest down. We're going to give you money. And no one is actually doing that yet, really, because it's not technically feasible, but it is a thing that's being proposed and it's going that needs to happen. But it seems like to them on the ground, a bunch of Western people who know nothing about their land showing up and telling them don't cut this forest down. I mean, why should they trust that we actually know anything? Why should,
Deep: I'm saying like, literally they download an app, they walk the perimeter of their property. They take their iPhone, they shoot some light or imagery, and then they put it in their bank account and they press a button that says, uh, here you get a monthly payment.
Kelsey: I think that it's a thing that can happen and that we will be able to reach a point where people are comfortable with that. But to a lot of the people who currently are the ones closest to this land, Involved in this whole world of global finance. They're like, why should I give you my bank account information? You got to take all my money
Deep: It'll be like that for a little while. But as soon as their buddy gets a hundred bucks a month, they're going to go do it.
Kelsey: Getting to that point. It's going to be harder than we want it to be. And I also think there is room for exploitation here that we need to make sure that we steer clear of, because anytime you're talking about wealthy people trying to influence the behavior of less wealthy people, you're in a dangerous place. And even though this case, we're all well-intentioned, there are going to be bad actors and we need to make sure that we catch them quickly before they hurt anyone. I think it's very important. And I think that. Have room to make the world and people's lives a lot better. We'd have to make sure we actually do.
Deep: So. I think this has been a totally like enlightening conversation for me. I feel like I understand this whole carbon credit arena, you know, better than I did before. So I, I want to thank you a ton for that. Um, I want to end with a question that I suppose is on everybody's mind, who is following cop 26. Like what's going to happen. Are we really going to get our shit together and say, or not?
Kelsey: Man, I wish I knew. I feel like, so when I took this job, I was like, I'm either going to feel way better because I'm going to be spending all my time trying to solve this problem, or I'm going to feel way worse because I'm going to never be able to avoid hearing news about it. And I think it's really been both. The news is often bad, but not always. There is progress. There are major companies who at this point are doing this because they want to, because they feel the consumer pressure to, uh, to invest in offsetting their emissions, putting a lot of companies who are offsetting their historical emissions, which is yeah. Critiques that I've heard that are like, well, isn't this just a license to admit more, not if they're going back 50 or a hundred years and offsetting all of their mission. Oh yeah. Can you imagine if, if shell and Chevron and Exxon did that? I mean, that's amazing. So it's, it's very cool. And there is progress happening. It remains to be seen whether it's fast enough. I think we're at the point now where we have to like, be honest with ourselves and be like, Anything we do is better than doing nothing, but I don't think we're going to be able to do enough. I mean, we're already seeing effects of climate change. We're not going to be able to undo the heat wave that happens last summer and killed people, right? Like it's already here. We can do more than we're currently doing. And hopefully we all.
Deep: a lot more, oh my God. I know. And it's so real. I mean, like, literally two days ago, three days ago, we had these massive floods in Bellingham and Vancouver and Abbotsford and my best friend in the world's got a pump and he's just going into friends' houses and pumping out, basically people marketing themselves as safe from the flooding. Yeah. We both sadly know that, you know, in another two months there's going to be another something somewhere. I don't want to go down the dark side because I really want to focus on how much awesome work you guys are doing. What can regular people do? Like maybe this is the final, final note. Like what can our audience do today? Can they go somewhere and offset some credits and like immediately do something.
Kelsey: I mean, it's a constant debate I'm having with my son though, too. Individuals can offset their carbon emissions. I think a big thing is if you're at a corporation, push for them to do it to our website, does support individuals purchasing. But our main focus is on. I'm just so much bigger. Individual transactions are pretty small, but they're super important. I see them coming in all the time. So yeah, go offset your carbon. But just more than that, like be aware that it's a societal level shifts that needs to happen and to be the person in your social circle, encouraging everybody else to. Make whatever changes they can like Google now will tell you which route on Google maps is going to emit more CO2 or which flight you purchase is likely to be more efficient, like pay attention to those things. Look at them the same way you would like spending money because you are spending the carbon budget and we are already over budget.
Deep: That's all for this episode, I'm Deep Dhillon, your host saying check back soon for your next AI injection. In the meantime, if you need help injecting AI into your business, reach out to us at xyonix.com. That's x-y-o-n-i-x.com. Text audio, video, or other business data, we help all kinds of organizations like yours automatically find an operationalized transformative insights.