Resources for Computer Vision Teams:

LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.
Foundation Model Assessment – Foundation models are popping up everywhere – do you need one for your proprietary image dataset? Get a clear perspective on whether you can benefit from a domain-specific foundation model.

Transcript:

[INTRODUCTION]

[0:00:03] HC: Welcome to Impact AI, brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven machine learning-powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research in computer vision for people and planetary health. You can sign up at pixelscientia.com/newsletter.

[INTERVIEW]

[0:00:34] HC: Today, I’m joined by guest, David Marvin. David was the Co-Founder and CEO of Salo Sciences, which was acquired by Planet last year, and is now the product lead for forest ecosystems at Planet. Today, we’re going to talk about monitoring forests. David, welcome to the show.

[0:00:49] DM: Thanks, Heather. Glad to be here.

[0:00:50] HC: David, could you share a bit about your background and how that led you to Salo Sciences and then Planet?

[0:00:55] DM: Sure. Planet’s background is in forest ecology. I received my Ph.D. from the University of Michigan about a decade ago now. Then I went on to work for the Carnegie Institution for Science, which is based at Stanford University as a postdoc. Throughout my Ph.D. and postdoc work, I focused on understanding how tropical forests are changing in response to climate change and uncovering some of what the drivers are in terms of how climate change impacts tropical forests’ carbon and carbon dynamics, so how carbon moves in and out of tropical forests at large scales.

I use a number of different tools to get this question from running experiments in the forest, using elevated carbon dioxide levels to fieldwork. Going out to the field, painstakingly measuring trees with diameter tapes and staking out plots to do those measurements, but then, all the way up to leveraging airborne and satellite-based remote sensing platforms to be able to at scale, understand change over time and identify things like, tree species and the amount of carbon contained within forests at large scales.

About this entire time, in academia, there was always this lingering question and desire to be doing something larger and not just larger in terms of is the question I’m asking the right one? Usually, something that just leads to a publication in a scientific journal, which, of course, it’s critical, right? We have to be asking these smaller, contained questions to be building on the knowledge base that scientists create for other scientists and for applied practitioners and policymakers to be able to answer and do their jobs properly.

To me, there was always a missing piece of, can I be doing something bigger and getting at the direct, if climate change is really making such large impact on tropical forests and not necessarily right now, but in the future, this was 10 years ago, right? We didn’t really have the same level of large-scale changes and evidence we have now. Could we be answering bigger questions? But also, working at bigger scales.

Having that multi-scale experience, working with experiments on seedlings, all the way to working with satellites over whole countries, I really felt drawn towards the bigger picture of trying to map very large areas and answer any of the questions.

It was in my post-doc time at Carnegie that I met and became close friends with my eventual co-founder, Christopher Anderson, who was a staff member at the time in the research lab that we were in, he was responsible for being on the aircraft, the lab operated, which had an incredible sensor package. I’ll get a bit more about what that means and how that’s important for mapping forests more generally, but he operated a number of the sensors on the airplane. Got to fly all over the tropics, Central and South America, Africa, and Hawaii, and just had an incredible experience working with the actual instruments that collected the data that we used at the time, and then eventually, the data we’re using now here at Planet.

He too had some of these same feelings around, what’s the right question to be asking? And am I doing enough? We had a lot of conversations where we were just stunned by the incredible amount of data and science that was being done in the lab we were at and labs around the world. What really bothered us was often, you would get a grant, do the research, publish on it, and then move on to the next grant, do the research, publish on it, move on. It was very rare you had academic labs that could effectively work both at large scales, and be able to produce an applied product that others could use since the goal of the academics is really about that publication. That is the be all end all that’s the currency of academia is those publications and getting out as many as you can in high-profile journals.

It doesn’t really provide everyone else who’s not in academia with effective tools to answer their questions, make better decisions, and be able to think holistically about climate change, the impact on forests, and the way that their behaviors can change to do better work, to be able to reduce the impact of climate and the impact of humans on forests. We decided that there has to be a better way that’s after that academia, because really, we didn’t see any pathway of staying in academia and doing this effectively. We decided, let’s start our own thing. It really was just like, what’s that thing going to be?

We looked at a few different options. The first thing was immediately, should we start a non-profit? That seemed like, it made sense. This was not really meant to be a commercial business and it was very academic-focused. We quickly heard from a lot of mentors that we had at the time, that carries a lot of overhead and a lot of additional work that you two as individuals with no experience doing this and a focus on building something, don’t have time and don’t have money for.

The easiest, simplest thing is to start a for-profit entity to do much more flexibility and control and without the need to have the overhead of a non-profit. We decided, yeah, we would corporate a company and start building. We did this on the side, nights and weekends. We both had full-time jobs at the time. Chris was a PhD student at Stanford. By that point, I was positioned with the Nature Conservancy in the California chapter as a research scientist. We spent about two and a half years working nights and weekends, literally a typical Silicon Valley, but we ran in a garage in San Francisco, working on this in our spare time.

It took us a while. Two and a half years of on the side work. Again, this is the point where we had enough of a product and system working to be able to actually do some consulting work that brought in revenue that let me be my position at TNC to go full time to Salo. When Chris finished his Ph.D. about a year and a half after that, he came in full-time. We began hiring. Eventually built a company to only around eight employees total, including ourselves. We had enough traction in the market. We had enough of a product market fit and revenue coming in that it was a profitable company by the time Planet approached us, we began talking an acquisition.

Ultimately, we were in this pretty incredible position where we had built and grown a bootstraps company to profitability and had full control over it. Our biggest need at that time was how do we go faster and how do we go bigger. We found ourselves back in that same position to me, where four or five years prior to this, where we felt like, we needed to do something bigger and faster, because the issue at hand, there was not enough going on to resolve a dual crisis of climate change and biodiversity loss in tropical forests and forests worldwide for that matter.

We were at our own crossroads with the company of how do we make it go faster and bigger. The options were keep going as we’re going, organically build it, get outside investments, likely through a venture capital firm, or get acquired. Ultimately, we pursued all of those different paths and worked through them and then decided that doing an acquisition through Planet, who, get some more detail in a moment, but build and operate satellite systems. Our data was built on satellite imagery, and we’re like, what better place to go and build this faster, better than the folks producing those data? It just really made a lot of sense to go inside Planet. We knew the founders already do longstanding partnership with the company. Their vision aligned exactly with ours around doing good for the planet and for nature. It was an easy call in the end to come to Planet.

[0:09:38] HC: That must have been an exciting transition for you. Congratulations.

[0:09:41] DM: Thank you. We never set out when we first started the company, and really hadn’t even thought about any of this stuff, until really, the year before the acquisition. We’re two scientists and we had to completely learn how to run and operate a business, how to manage teams, how to hire. These are not skill sets that we had any experience with. I think, pretty sure for both Chris and I, our first for-profit jobs were with this company. Literally had no organizational experience inside of a for-profit company. Very different world for us for sure.

But we found our way through and really enjoyed the time independently operating and building a company and are really excited about the future, where we can go from here. Also, I’d be remised if I don’t mention, a few years in, we brought on a third co-founder, Kyle Gertridge. Kyle had been a longtime friend and colleague, who was an attorney by training. We actually used to live in the same house in San Francisco when I first moved to the city from my postdoc. He was following along from the very, very beginning. He actually helped devise this on the side a little bit right in the beginning, just about entity formation and how to think about intellectual property and how to think about growing a company. But no one had ever said like, “Oh, you should think about what people often hear now term like, what’s your exit strategy?” I’ve never thought about that at all. We didn’t have one. We’re like, I don’t know, build the company, make it profitable and run it for the rest of our lives, right? Isn’t that what people do?

It was a really eye-opening experience to have companies approach us to be acquired and thinking about what that meant for our futures. In the end, like I said, a company like Planet was essentially, probably the only company we would have really ever been acquired by just given their vision and the fact that they have their own satellites and we’re a satellite software company.

[0:11:40] HC: Let’s talk a little bit more about Salo Sciences. What did you do there? What were you building?

[0:11:45] DM: The original vision and the mission of the company was always build solutions that can help other organizations make effective decisions around climate change and travel the forest loss. We never really deviated from that broader mission. That was always what we intended to do and what we still intend to do here at Planet. How we got there was a little circuitous. Our first revenue and our initial projects ended up being about wildfire, because we were living in California. We first began actually being successful with our software pipelines. It was right around the time of the wildfire crisis in California really kicking off at 2017-2018.

Those really intense wildfire years in California, we saw an immediate application of our technology not to tropical forests right after that, but to be able to help map the fuels, the vegetation fuels that drive wildfire behavior in California. We received some early funding from the USDA for their SBIR program to help build a system to map tree mortality across California. At the time, there’s also a really huge tree mortality crisis occurring because of a long-term ecological drought in California. You had something like, 90 to 100 million trees dying off per year. The best way to map them at the time was to fly airplanes with a trained observer, looking out the window and hand drawing big circles and squares on a map, on a tablet where they saw mortality.

We’re like, “Come on. There’s got to be a better way. We know there.” It’s this approach, leveraging satellite imagery and machine learning properly to be able to create these really highly-resolved and accurate maps that can be updated regularly for mortality. That was our first funding that came in and our first real project was this mortality map of California. As we got into more of the wildfire community that would want to use this data, we saw more areas where our technology would be able to be applied and make a difference. That became this much larger project that became termed the forest observatory, where we would use satellite imagery and deep learning technology to extract high-quality layers of vegetation fuels that are inputs to wildfire behavior models here in the western US.

Because at the time, and really still to this day, one of the main data sets for running these models is a national data set called LANDFIRE, that was never intended to be used in this way. But it is the core data set that folks use to run fire models, and it really doesn’t work very well at higher resolution, so at the tree-to-stand scale. It’s meant for large-scale county and state-level mapping. It’s pretty inaccurate. It’s updated pretty irregularly. I think it’s gotten better these days. But six or so years ago, every two years and take a couple of years of work after the imagery was acquired to actually be released, so it was always lagging, pretty inaccurate, and pretty coarse resolution.

What we did, what really became the foundation of the systems that we still have today was we leveraged high-quality airborne LiDAR measurements of forests all over California. Airborne LiDAR is one of these technologies, these sensors that was on that airplane back in my post-doc lab, shoots out hundreds of thousands of pulses of laser light per second and reflects back to the sensor, and it can basically recreate in three dimensions a forest, or a city, whatever your mapping target is. It’s extremely precise. It’s centimeter-level accuracy, and it’s very high-quality data. We consider that the gold standard of forest measurement.

We can take that 3D LiDAR scan and convert it into these two-dimensional maps of metrics like the height of the forest, or the canopy cover of a forest, or what’s termed canopy base height, which is the height above ground of the first branches of trees in a forest. All these different really high-quality metrics can then be paired with satellite imagery and then fed into a deep learning model, so that the model can learn from that high-quality LiDAR data how to extract those measurements just from the satellite imagery.

Then once that model has been trained, it can be applied to contiguous large-scale maps, let’s say all of California, or the whole Western US, of just satellite imagery without additional airborne LiDAR data make predictions of every pixel across that area. The entirety, the wall-to-wall map of Western US, we can predict the height of every tree at each location across the Western US, even though airborne LiDAR only covers a fraction of that same area. That was really what really broke things open for us was leveraging large-scale airborne LiDAR collections and deep learning to be able to create these really nice maps of vegetation fuels that could be used into power WILDFIRE models.

Even though the product was focused on WILDFIRE and WILDFIRE-related fuels, the same concept can be used to map tree height, tree cover, and above ground carbon in forests all over the world. The application of that same system can be pointed at a number of problems and be used to produce a number of products that would be really valuable. That was what got us our start.

[0:18:08] HC: Then now at Planet, are you focused on something quite similar, or have your goals changed?

[0:18:14] DM: Very similar, but much larger scale. What we’re doing now at Planet is we’re leveraging those same types of systems, but building out tooling infrastructure to do it globally. Back in October of 2023, just around six months ago, only nine months into our post-acquisition period, we released Planet’s first global map of above-ground carbon, tree height, and tree cover at 30-meter resolution, so not as high resolution as the work we’re doing now. But it was meant to be a backward-looking, more coarse resolution version of what we’re building towards currently, so that people could use these data sets to look at over time, what a change in carbon height and cover to evaluate the suitability of new forest carbon projects around the world that are meant to lead more carbon in the forests than there are at the current moment?

Having a historical look back in the ability to see trends over time is really key to enabling better locations and types of forest carbon projects all over the world. Ultimately, what we want to produce is a near-tree-level map of the world’s forests, and we’re well on our way to doing that and expect to be releasing that later this summer, or in the fall.

[0:19:46] HC: When we were working on this at Salo Sciences, was this already based on Planet imagery, or were you working with other satellite constellations at the time?

[0:19:54] DM: That’s the other key factor that early on we discovered was really important. Much of the academic work on leveraging remote sensing and machine learning was focusing on using a single sensor, like Landsat, or something to be able to train and run these models. What we did was take multiple different data sets from different sensors, stack them altogether, and train the models on what’s now termed multi-sensor fusion feature data sets.

We like to use all available satellite imagery that we know makes sense to apply to this problem, because each one, even if they’re the same type of satellite, so for instance, Landsat and Sentinel-2 are both multispectral satellite sensors. It just means that they measure only a handful of wavelengths of light, and each have a green band that they’re measuring. The exact wavelength of green light that they’re measuring differs a small amount, and just that difference right there gives us different information about the surface of the Earth. This was an insight that the lab that we worked in back at Carnegie had, because they were working with a different sensor that was on the airplane called a hyperspectral sensor.

Instead of measuring 8 to 12 or 16 bands of light, they measure 400 wavelengths of light, and it gave you continuous, essentially, ability to look at the whole spectrum from visible to shortwave infrared light that reflects off the surface of the Earth and back to a sensor. In this case, it was on an airplane. It’s an enormous amount of data, and it’s extremely high-quality information, but it’s hard to work with, right?

Right now, we’re running models on 20, 25 band data sets. 400 bands of data. Difficult problem to work with, but we found out, and this was we, as in the entire lab, through many years of academic study, that leveraging all of those wavelengths of light is actually more powerful than trying to feed your selection. Pull out a subset of the best performing wavelengths, because each one has its own unique interaction, the wavelength of light with a target. In our case, mainly its leaves and branches of trees. It gives you a distinct piece of information about what is happening on the ground. In the case of hyperspectral data, they can tell you about the chloroform content, or the leaf water content of invisible leaves. They can tell you about defense compounds, or the structure of the canopy.

We knew that using all these different sensors would help us to create better outputs. Second key insight, again, the focus in academic literature at the time was really on machine learning methods that were single-pixel only. You look at one pixel of your feature set from satellite imagery, look at one pixel of your airborne LiDAR, or whatever your ground base data was to train on, and it would try to do one-to-one pixel predictions and learning. We’ve leveraged convolutional neural networks instead, which led us to take into account the context around those target pixels. It wasn’t just a single pixel. It was an area-based machine learning, essentially, to make it simplified.

Those two techniques using multiple different sensors, using context-based convolutional neural networks, open the gates for us to do far better-quality mapping that was very different than what was being done at the time. These were things we continued to use here at Planet. The difference now is, instead of as a partner of Planet, having access to just a small amount of data, which we were able to do through a partnership back at Salo, we have access to their entire catalog of data, which goes back almost eight years now with our resolution data set, the tree-image data called Planet Scope, and it lets us work with data from all over the world. We have this amazing space of temporal data set to begin building and training our models on in a way that as a partner, we just couldn’t ever have had.

We learned a lot about Planet imagery and, of course, other sources as well during our time at Salo, so we came in already as experts on the data, but now we’re just – we have a totally different understanding of the data and a much better grasp of it, and it can leverage the power of the internal systems and tools that we would never have had as a partner before being on the inside.

[0:24:57] HC: You mentioned that you train these models by using airborne LiDAR as essentially, your ground truth, the assessment of the canopy height, the different measurements you’re taking, but then you use satellite imagery as your input. Do you only have LiDAR for certain locations? How do you validate that your models work in places where you don’t have LiDAR? Maybe it’s a different part of the country, maybe it’s even a different continent. Are you able to validate that your models generalize to different locations?

[0:25:25] DM: That’s a key question, and it’s one of the major focus of our group in Planet is to, number one, collect as much openly available airborne LiDAR that we can. To date, we’ve collected this today as in from Salo through Planet, so about five years now they’ve been working on this. We’ve collected about 14 million square kilometers of airborne LiDAR from all around the world. We think this is probably the largest public, or private repository of airborne LiDAR in the world. Even at that number 14 million square kilometers, which is a huge number for airborne LiDAR, we have covered a tiny, tiny, tiny fraction of the world, obviously.

What’s important now is trying to fill in a lot of the gaps where we don’t have a lot of Airborne LiDAR. To your exact point here, not having much or any airborne LiDAR from certain regions of the world could introduce a bias, number one, in terms of our model output and its ability to make predictions, but also, leaves us high and dry for evaluating its performance in those regions. To address both of these issues, first to evaluate model performance, where we do have good airborne LiDAR, what we do is we retain around 15%-20% of all of the airborne LiDAR coverage. We remove it, keep it out of our trained data set, out of the validation data set from the modeling, and use it to test the model performance independently.

That only gives us good test data over places where we have airborne LiDAR, obviously. Then what we do is collect, or utilize other existing well-known, well-calibrated data around the world. This mainly comes from two sources. One, field plot data collected from universities, or governments all around the world. We have around half a million field plots that are geo-referenced locations that we can do these comparisons with. Then we also leverage a space-borne system called GEDI. This is the Global Ecosystem Dynamics Investigation. It’s a space-borne LiDAR on the space station. This gives us very broad coverage, between 50 north and 50 south latitude at least. The higher latitudes are missed, unfortunately, on this sensor.

Otherwise, it’s very well distributed. It is LiDAR measurement from space that are covering the whole of the Earth and they’re geo-referenced for the most part. There’s a bit of some quality issues there. But we’re able to use that data set as a completely independent way to evaluate our predictions of our heightened cover models. We approach the validation aspect from a few different angles, trying to source as many different independent data sets as possible to do validation. Then we also like to do comparisons to well-known public data sets, either from academia, or from governments, folks who have done similar global maps of height, cover, or carbon. We will compare again to those since they’ve often been gone through rigorous peer review, or vetting by government scientists to just show how our performance compares to those similar types of data sets.

We’re putting out a really extensive validation report on our global 30-meter data set later this week, early next week, I believe it’ll be coming out, that goes into depth on both validation and comparison of our data to all these different sources. Now, on the bias question, there’s a lot that our team thinks about and works on to be able to reduce the bias that results from only having data in a heterogeneous distribution of LiDAR sites around the world.

Unfortunately, there is so much of this LiDAR we have that we essentially have LiDAR in every forested biome in the world. The model that’s making predictions somewhere likely has seen a similar forest, whether in that exact biome, or in a similar type of biome where the structure of the forest is close enough, where we’re not getting wildly terrible predictions in places we don’t have any data. Does a pretty good job of being able to predict without ever having seen that forest. Because in our point of view, forests just don’t look that different globally. The range of variation of forests compares, let’s say, the range of variation of human faces is just much more constrained, especially when you’re talking about satellite imagery and not something like, drone-based imagery, we’re getting each leaf resolved.

We’re just seeing tree crowns and the structure of tree crowns of a forest. The models we’re leveraging were built to be able to handle things like, facial recognition, which is far more complex and much larger, I think, variability face-to-face than forest-to-forest over the world. That helps us initially off the bat, make good predictions. But we’re always trying to think about how to tune a model and shift our sampling strategies so that we can reduce our bias and predictions.

[0:30:57] HC: To build effective models for forests like you keep down here, it sounds like you need a fair amount of knowledge, not just in machine learning, but also about remote sensing, with the LiDAR and satellite imagery that you use, and about forests themselves and trees. How is all that knowledge integrated in order to help your machine learning developers create these powerful models?

[0:31:19] DM: That’s a great question. It’s a particular insight that we have come to understand over the last decade or so of working on this, not just at Salo, but beyond our academic work, is you really need to have domain expertise in all three to make this work well. It’s really difficult to come from the machine learning community with zero background in biology, or ecology of forests and be able to produce high-quality and scalable models like this.

What we have focused on primarily is in building our team at Salo and now here at Planet is trying to bring in folks who have at least expertise in two out of those three. Whether it’s either background on remote sensing and machine learning, but no ecology background, or you have a ecology background with really deep experience in machine learning, but maybe no remote sensing experience. We try to get two of those three, because we think given the composition of our team here at Planet, who collectively have deep expertise in all three of those areas, we can get some up to speed pretty quick in some of their missing background.

Not being strong in one of them is definitely a large detriment, and we’ve seen this again and again with oftentimes, it’d be customers we’d be talking to back at Salo, they’d be like, “Yeah, we’re thinking about using your data, but we got really good machine learning team here. We’re just going to give it a shot and try to produce it ourselves.” We’d be like, “Cool. See you in a year.” Almost always, six to nine to 12 months later, they’d be back, being like, “Boy, was that difficult. We did not do a great job. Let’s talk about using your data.” You really do have to have the three legs of the stool to be able to build a quality operational product that is meant for forest monitoring.

[0:33:19] HC: You mentioned publishing earlier and publishing papers is the primary goal in academia, but an industry is definitely not. Yet, on Salo Science’s website and on Planet as well, I see a number of research articles. What business benefits have you seen from publishing?

[0:33:35] DM: We continue to publish and make publishing a priority with our group here at Planet. I think a number of the other product focus groups here at Planet do as well. I think this really comes from leadership here at Planet that values and understands the necessity of continuing to work on primary and applied science as part of the business strategy. It’s very hard in this vertical of forest monitoring and I’ll extend it over to ag monitoring as well. When you’re producing brand-new product like this that hasn’t really been seen before, keep learning and be skeptical.

A lot of your end users are going to be government agency scientists. Their first question would be, “Let me see the validation. Do you have it peer reviewed yet?” Even though the ultimate users of the data, they don’t control the budget and they don’t have the decision to buy that data, but they’re going to be the ones that give the green light, yes or no, to the person who’s made that decision. Making sure you have scientists on your team, making sure you’re still active in the scientific publishing community, that you’re up on the latest papers that are coming out and basically, acting like a scientist in an industry position is crucial to make any product work, especially in branding markets, like forest monitoring and carbon markets.

Now, we hear all the time and it’s a crucial piece of it. Granted, we haven’t yet published on this most recent product, but we have one in the works. We plan on publishing at least two or three papers over the next year on these new products coming out, because one, it’s critical to get that scientific validation through peer review, but it also lets the community know, hey, this is something that we’ve done that’s new. Come try it out. Come try our data out. Give us feedback on where you see it not working well. Come critique it. We want to leverage that feedback to improve the product over time. That’s hard to do that if you don’t have publication that you can pass around to other scientists and a data set that also is in a public repository.

That’s one of the other key alignments that we have at the Planet was their focus on impact and a public good data set. What we’re doing now is working with the impact team here at our Planet led by Andrew Zolli to figure out what is it be our digital public good strategy, where we can make a lot of these data on forest carbon and high-end cover available for free to the research community to evaluate and build new discoveries on top of. That’s a key strategy that Will the CEO and Robbie the co-founder have always had is getting this data out there and into the hands of researchers who can make discoveries and it really push the envelope.

Because we all know as scientists, there’s non-budget in their tight grant cycles to buy data sets. It’s spent on other things. There’s a default to using free data. We prefer if they’re using higher quality free data than the status quo that’s been out there. We really see publication, open data and engagement, especially with the scientific community as key factors and uptake and acceptance far outside of that community.

[0:37:08] HC: That’s good to hear that it helps advance not just what you’re building, but enables others to build upon it and extend what you’re doing and use your data, advancing science overall.

[0:37:20] DM: Yeah, the ultimate goal, right, still goes back to why we started Salo Sciences. It’s around, how do we build better tools that includes data to come back climate change and the biodiversity crisis? If that means losing a small amount of revenue from what you might be going to get from some grants here and there to let scientists build and understand the systems and the role in a different way, that’s part of the mission. That’s part of why we started Salo and why we continue working towards at Planet.

[0:37:53] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:37:58] DM: Going back to the earlier question on those three domain pieces, ecology, machine learning and remote sensing, we still see it today and I think we’re seeing it less and less, but a new company starts and then a business model around some really incredible end market platform, or tool and they just assume that, well, we have machine learning scientists working on this part of the platform. We can just have them also work on the data production. It can’t be that hard. Don’t do it. You’re going to waste a lot of your investors’ money and a lot of your time.

Focus on what your actual product is. If it is not producing data from satellite imagery, if it’s just utilizing it as part of a workflow, or a platform, or some insight for a customer, don’t build it yourselves. Go to the range of different folks out there who are doing really high-quality work to extract data from satellite imagery. That would mean just enforcing all types of industries. It’s a very hard, complex problem. It has taken us over a decade to get there. It has taken the competitors to plant it, almost all of them are founded and led by former academics that we used to know and work with back at Carnegie.

The world is very small in forest carbon mapping and it continues to be very small. Now the business rolled around with forest mapping and data is still that same very small group. It is not grown substantially, because it is so hard, because it requires such deep domain knowledge to do correctly. Also, I think to the other point, from a business level, you need credibility and you need to be able to talk to other scientists and understand lingo and understand the historical development of these systems and the academic history of it to be credible. That was my main piece of advice is don’t try to do the data piece if your ultimate product is not the data itself.

I think the other piece of advice is stick with it. You probably have heard this already. I want to reiterate this, that there will be times that you in your heart think maybe even know you’re going to fail and you’re not going to get through. You’ve got to get it out. As long as you are disciplined about your runway, disciplined about being realistic about revenue coming in and your ability to close deals on realistic timelines, push through the hard parts, because eventually, if you really are building a amazing product that can change things, you will eventually get uptake. You’ve got to just stay in business long enough to get there and boy, did Chris and I hit some hard times about five, six years at Salo, before finally, and it was very much like, year, year, year, year, like very little revenue. Then finally, boom. It just took off in a couple of years. You’ve got to just be able to wait it out, until you can get to that point where you get product market fit and you get enough customers to believe in what you’re producing.

[0:41:16] HC: Finally, where do you see the impact of the forest monitoring efforts at Planet in three to five years?

[0:41:22] DM: My hope is that we continue to build these global products that bring really high value to not just the forest carbon market, which is the initial real focus of this data to bring transparency and accuracy to a marketplace right now that’s suffering from some pretty major credibility and transparency issues, but far beyond that’s it. Really, anyone that is working to restore nature, protect nature and to fight climate change, we want to build the data that lets you focus on that mission and not on all the very long, deep, upstream issues to extract and find high quality data to go into your system, or go into your study, go into your product.

We want to be able to power a number of these solutions all over the world in all different industries, where if you were trying to make a difference in leading more trees in place, leading more carbon in the ground, alternatively removing more trees and more fuels in wildfire prone areas, we want to help make that happen faster at larger scale and cheaper, so that the work can just get done. Someone told me last week, they’re like, “We have 70 months till 2030.” I like that they were thinking in that granularity of months, not just years. But 2030 is this key global deadline for climate change, for biodiversity, for a number of global problems that we’re facing and we’re so far off track right now, and we don’t have much time to turn around, and we’re already seeing the impacts month over month, year over year. It’s way worse than anyone thought 10 years ago.

We want to help people go faster, go bigger, not make costs, that’s a big issue. I think in three to five years, success to me would be if Planet is known as one of the global standards for high-quality data on all types of hard problems, whether it’s wildfire, carbon, sustainable agriculture, restoration of nature, Planet should be known as the go-to data source for trusted, accurate, and usable data to produce better outcomes.

[0:43:52] HC: This has been great, David. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:44:02] DM: You can go to planet.com. Find out about a lot of the different data solutions that we’re currently producing, whether it’s in forest, or agriculture, simple government, and beyond. But the next vision, the future of Planet is going to be taking that data and producing insights. I think, just next month, we’re launching a new platform to help folks do this. Keep your eye out for new and upcoming insight-based products from Planet via website, or I think on Twitter as well would be a great resource to follow along.

[0:44:36] HC: Perfect. Thanks for joining me today.

[0:44:38] DM: Thank you.

[0:44:39] HC: All right, everyone. Thanks for listening. I’m Heather Couture and I hope you join me again next time for Impact AI.

[END OF INTERVIEW]

[0:44:49] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share with a friend. If you’d like to learn more about computer vision applications for people and planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]