 My name is Dave Hulsland. I'm a staff scientist with AOP, Airborne Observation Platform. You just think of us as the airplane people. We take care of all the imagery pieces. As Courtney kind of did a nice intro for that. And I know you've got a couple of good presentations from Nathan, Tristan, and some of my other colleagues in our department over the last couple days. You're probably gonna, this is gonna be a very quick introduction. You know, each one of these vegetation indices, each one of these data products could be and probably was somebody's thesis at some point. Not gonna get into a lot of detail in 25 minutes. So you had a good overview from Nathan on what the AOP instrument package is and how we fly it in the airplane. And it's all sounds and is really very interesting, very cool. They're flying right now. They've been flying since April. They'll come back in about October. And it is, as Courtney mentioned, quite a logistical nightmare. We've got a beautiful camera, amazing Waveform LiDAR, pretty much the only publicly available Waveform LiDAR you're gonna find. The Discrete Light Arch, Tristan does a wonderful job with, an imaging spectrometer, Averis next-gen. Nobody else makes that freely available the way we do. I wish I'd had that when I was in grad school in the 90s. And it produces about over 200 terabytes a year. So it's this huge, amazing, beautiful package of instruments and data that's so big, it's completely useless. There's no way you could get around 200 terabytes a year. And that's kind of what some of these data products are. That's, that raw data package is kind of the everything. Any question you want, you can probably use that data to answer it. But when you're asking your question, the other 85, 90% of the data, you don't need. So these data products are to help you narrow down, get to the part you want, and not just be overwhelmed with this tsunami of data. So, how did we get to our package of data that we want? I'm gonna focus on the imaging spectrometer products. I know Tristan's taking care of the LiDAR pieces. And we'll talk about how we derive those, the uncertainty within it, how we get to those. So, remote sensing started out with, well, it's, you can definitely get a good answer if you walk around on the ground, what I call my background's geology, boots on geology, walk up to it, look at it. Yeah, there it is. Probably the least efficient form of mapping too, as much as I like to be outside. So folks got thinking, well, from camera, from the air, I can see stuff, maybe a little better, and then I can maybe add a few extra bands, et cetera. But there's problems with doing that too. There's a lot of the spectrum that we can cover. So this is kind of the useful part for passive remote sensing. It's a nice diagram from NASA USGS for running the Landsat program. And you can see we run down to about 350 with our spectrometer up to about 2,500. And you can see if they're going to bother to fly a satellite, and if you ever get a chance, go to Aeros Data Center and actually watch them downlink from the satellite itself. You have 15 minutes once or twice a day to get all of your data that you took with all of those pictures. Now if you've ever downloaded a Landsat scene, you know it's set FTP, go get a coffee, go to several meetings, maybe come back at lunch. It might be done. So if they only have 15 minutes to get several orbits worth of data, they must be doing something much more quickly. But it also means their bandwidth is an extreme premium. They actually use three channels simultaneously. So that means they don't want 400-some bands. They need to pick and choose very carefully. We also have the ability to maintain all our stuff every time it lands there. Once it's in orbit, well, good luck. I hope it works for the next 5, 10, 15 years. Nominally, they're only supposed to last about seven years. We're lucky that they last decades. So that's kind of a different problem, but we don't have to deal with that. So we actually cover this whole range with 428 channels. So while you can answer specific questions with certain bands from, say, a Landsat data set, you can put together anything you want with all the channels we have. Of course, it does mean you have 200 terabytes of data a year to deal with, too. So it's a different problem. So to answer those specific questions, what folks would do is say, all right, well, I want to know about where chlorophyll absorption is or where nitrogen is or certain other features. And they'll sample in on certain sections there. And over time, fortunately we've been doing this since the 70s, spectral indices would get developed. These are the bands that matter. I want to look at their values. I can make a ratio, get down to a single score instead of this huge image cube. And from that, you can get these spectral indices. And that's the data products that I'll be telling you about today. There's a few ways that those get developed. They all start out with mapping the phenomenon the hard way, the boots on geology that I mentioned. Go out, get the real map, then get your imagery and basically look and see what matches up. Main ways to do that, M and F, minimum or maximum noise fraction. It's the same transform depending whose paper you read. And then start to get which bands are most significant, which correlate best with what you mapped on the ground. Then try to standardize that out of the best quality spectral imager you got. I know Nathan did a good job covering how hard we worked to keep that calibrated. And use your exploratory analysis to find those bands, develop your index, and go with that. And over time, those get to be very well accepted. And the first one we go over is going to be kind of the grandfather of them all. The normalized difference vegetation index. About 40-year heritage just said, look, you've got these beautiful passes of satellite imagery, but I really just want to know what's the health of the vegetation. Give me a map of that. And that got developed as near infrared, so Landsat Band 4, minus red, Landsat Band 3, over 4 plus 3. Very simple, normalized, so it goes from negative 1 to 1. Higher score means it's greener, more lush, et cetera. And if all you have is just a handful of bands, well, that's very easy to use. 4 and 3, and off you go. And it sounded very good. Well, that was the only sensor we had. You know, in the 70s, the satellite actually flew and it didn't blow up, and it did take a picture and it returned it to us. You call that a victory. We had a little higher standards these days. Like, well, maybe I'd want these different bands. Maybe I want a more particular piece. So we need to work with that heritage, stay within the accepted norm of the equation, but put out a product that's as good as what we can produce with today's technology. So people started to look at, well, how well does this actually work? And they started finding things like, let's see, one of these is a laser. When it gets really green, it starts to saturate. It's kind of like, you know, whoever wins gets to write history. The people who build the sensor generally works in the area where they're used to. Well, if you're in North America, that works really great on typical mid-latitude forests and grasslands and that kind of thing. And as we started looking at more equatorial and more heavily vegetated places, tropical forests, it turns out at high values it tends to saturate. Really green and a little bit more than really green look the same to NDVI. So if you're using these indices, you may want to keep that in mind because you'll start to lose a little sensitivity and specificity in different areas that are more heavily greened up. So to that end, when MODIS was developed some decades later, they said, right, well, we want to upgrade this. So when we do, instead of just doing NDVI, we'll incorporate a blue band as well. That'll work on the saturation. We can account for a little bit better variation in soil background and aerosol scattering issues. This kind of starts to hit at the one main, probably one of the biggest sources of air in remote sensing products is the atmosphere. From a remote sensing and instrument perspective, the atmosphere is a complete pain in the butt. We really wish it just wasn't there. Then we wouldn't have anything to study. So instead we deal with it as best we can. And when we get into the uncertainty section, I'll show you how we work with that. Once people said, all right, well, if we're going to mess with NDVI and we're going to try and make this perform better, everybody starts to come up with their own indices. If you search vegetation indices through Google Scholar, you'll find literally dozens, 60, 70, 80, 90 different indices that somebody developed that works very well on their data set in their area. And we've kind of said, well, those are good and they might work in your study area. But as Courtney mentioned, we need it to work in all of our sites. So we've narrowed it down to a handful about 11 spectral indices that do that. Atmospheric resistant vegetation index is another one that was designed to use the blue channel. As you get down into the shorter wavelengths in the blue, you start to see more atmospheric effect. And using that to account for what that might be doing to your data allows you to get a better index. This is particularly good in high aerosol content. So think of hazy stuff. You take a flight down to Florida in the summer and you look out and it looks like soup with all the humidity in it. You go over an area with a lot of aerosol content, pollution, et cetera. That's what you're seeing. And using that blue channel allows us to account for that better. I was actually camping over the weekend up in the high country. We have the same problem here, but instead of pollution or humidity, it's pollen. You can see clouds of it coming off the mountains right now. Then once folks said, well, the greenest index is great, but I want indices for other values. We have the PRI, also known as chemical reflectance index, also known as canopy xanthophil. Again, it's a normalized ratio and the values, so it's negative 1 to 1. Green veg comes out about centrally located, negative to positive 0.2. And really is focusing in on other pigments in there, not just chlorophyll and greenness. Give you a more complete estimate of the vegetative health. And of course, then folks want to get into, we talked about dead canopy stands, et cetera. The living biomass is great, but what about the dead stuff? So we get into NDLI, which is either normalized difference ligand index or canopy lignin. This one is relatively new compared to others. So far, seems to have pretty good support in the literature, but in using it, it will open up an opportunity for NEON and for everyone using our data to do a little bit of validation. How well does this work across different sites? How well does it actually correlate with what's going on with the ground with all the great measurements collected by Courtney and his team? It's supposed to be capturing that lignin and the foliar biomass and work at getting that whole content. So in using that, well, see what you turn up. A lot of our data products are designed. No one's ever collected them at this scale across these different sites at this kind of frequency over this time span. And we don't know how well they're going to measure up. And that's kind of the nice thing in science. The best answer is, I don't know. Let's go find out. That's a great opportunity for you with the data, which, as Leah said, you're getting your hard drives today. So you'll get to see some of that 200 terabytes. Nitrogen index. Speaking of new indices, this one is very, very new. And while NDLI seems people are kind of centering in and it's getting pretty good support so far, normalized difference nitrogen index is a little more, I wouldn't say controversial, but not a lot of people have used it yet. It hasn't been very widespread. It's been tested in a bunch of different biomes. I mean, that's made for neon data. Let's fly everything and see what it looks like. So huge opportunities here to see how well this comes out, how well it correlates with Courtney's team's work, getting the information off the ground. I think we're going to see some excellent work there. We mentioned Greg's doing a lot in Amazon and other canopies. And this is one of those focuses. I'm very excited to see what comes out with this. It's really going to vet the index and we'll see where it goes. Maybe we'll get to be part of fine-tuning that in the literature. That is very new. So new it's not in the ATBD yet. It's the algorithm theoretical basis documents. Basically, why did you make this data product? Details out the formula, has whole sections on the uncertainty in it, et cetera. For the handful we've got here, it's 31 pages. I think it's awesome reading because I wrote it. But if you're having troubles going to sleep, it's probably a good fix for that too. But it does have a lot of good detail. Again, I don't want to go through all that in 25 minutes because maybe you care about NDVI, but you could care less about it. You want nitrogen. So kind of give you the top level of that. Very, very new one as in last month that was just added is the soil adjusted vegetation index. You start getting an idea. It's the index isn't the problem. It's all the other junk that gets in there and all that stuff. Everybody tries to compensate for that as best they can. The reason for that is that is an input to the algorithm that produces leaf area index. That's like the nitrogen one. There's a lot of research back and forth about how well does that actually correlate to what's going on in the ground. When I did use to teach in the defense area a lot, we'd work with the Intel community and the biggest thing we had to drive home for them was just because it's on the ground doesn't mean it's in your data. They'd always say, well, I know the target was there, but it's not in my image. You didn't measure something that would show it up. So it's the same idea. We try to get to these indices. We want to measure leaf area from the air. It sounds so very simple, but you get into the details and that's really where the hard work comes in. But it's a good start. It does have a little momentum in the literature, a lot like nitrogen. And I think it was a tremendous opportunity in research to show how well does that algorithm work? What are the peculiarities of the sensor and your processing and different areas in which you're going to operate to show does that really measure? Does it scale? So we can go from Courtney's measurements to the AOP measurements to MODIS and Landsat and the Continental Scale. And it's a great opportunity to see how well that works. So all the plant matter is an interesting piece to it. That's funny. It cuts off on the bottom. But the water indices are kind of the next piece. It's nice that it's green, but if it's dry, it's going to be dead soon. I mentioned about the pollen coming off and my wife and I were talking about that. It's like, well, fine, I have allergies right now, but it's not a big deal because it's 100 degrees in a week. Everything will be dead anyway. So this kind of allows us to get to that. It's green now. Will it stay green? So when we talk about canopy water, there's really a family of indices in there. There's the moisture stress index. It says it's green now, but it's stressed out and it's going to wilt off soon. It's a simple ratio, one band over the other. Tends to range from zero to three. Might go a little high. The main reason I want to highlight that, though, is it's backwards. So high equals dry on this one. Most vegetation indices, the higher it is, the more green, lush, healthy, et cetera. But this one is basically the higher it is, the more it's stressed out. This is very good in forested areas for estimating fire susceptibility and risk in it, as well. Similar family, the normalized difference in forehead index. But the nice thing is that this one is normalized. It uses the same wavelengths on it, but it ranges from negative one to one, so it's a little easier to use. You'll find green veg in about .02 to about .6 there. And it's back to the usual. Higher is greener. Normalized difference, water index. You kind of start to see the theme here. Everybody wants it normalized. You want two, maybe three bands. This minus that over this plus that. Call it a day. Keeps them very comparable and also makes it much easier to use. This goes in for the actual spectral water absorption. So that part of the spectrum where you actually see if there's water, there's a feature there that absorbs the photons. It also helps in terms of working with the scattering that you'll see in canopy studies. So, again, in a forested area, very good. It's often used in conjunction with LiI, and all of these go into productivity and fire estimates. So these are going to be kind of your main family of products that you'll use as you're looking at health and productivity across the whole stand. Again, negative one to one. Green veg, negative .1 to about .4. And all the detail on this and all the formulas, et cetera, in the document as well. Beyond that, we get the normalized multi-band drought index. So this is kind of like we had NDVI, and then you saw EVI where we added the blue band. This is the water version. We had our earlier ones, but we add in a third band for better drought sensitivity on it. .7 to 1 for dry soil, .6 to .7 for intermediate, less than .6 for wet soil. So this is a nice condition of your plant screen, but the soil's dry. Again, tells you it's not going to last long and it gives you a little more prediction on what's happening in the stand. Water band index, a much older one. It's another, just a simple ratio, but it has been used quite a bit. Again, we don't have, we just saw what was at three or four, new elements got officially named by IUPAC a couple of weeks ago. We don't have that. I suppose we could probably have a naming convention, but if you go to AGU Fall Meeting, you're probably more interested in a pub crawl than a naming convention. As a result, water band index has the same acronym, the same name in the literature. There are several different versions of it. So when we say WBI at NEON, we're going off of the earlier Pinoeles 1995 paper, 970 nanometers over 990. So as long as you know which one you're using, that's really the key part. Whoever named it the second one hadn't done that, but we're stuck with it. The green veg in that comes out to 0.8 to 1.2. And band aggregated products. So, pardon the rapid flip, but if I go back here, that's what Landsat has. You're going to fly the satellite, you have all the attendant problems that go with going into space there, so you only get a half dozen bands, and you want to produce NDVI, and everybody's used to it being four minus three over four plus three. Well, what do you do when you've got 40-some channels in there? Well, they said they wanted the near infrared. Okay, well, that's what the filter you could get that was small enough and strong enough to sit on a rocket and not just be obliterated when you launch it. We have a choice. Which bands should we include? Right now, we're just using the center, other ones available out of Landsat and others. But we have the opportunity to use more. So, I have a version coming up that we'll actually use in the case of the near infrared there, about 25 bands or so. Now, by doing that, that allows us to actually push down some of the noise as well as we get into the uncertainty section here. And the same for band three. So, anything that's referred to as a broadband index means we probably are going to have a few channels in there rather than just one. So, that'll be a nice upgrade as well. That'll be coming this fall. There are some, though, that are not broadband, like nitrogen. They're very specific about which band they want. So, we'll just have the one in there. But where possible, I will aggregate together to allow us to produce a little bit better product in terms of pushing noise down as much as we can. As you heard with Nathan, you know, that's our whole job is to get that reflectance data, get that LiDAR data in the camera as good as we possibly can so that the derived products and the science that you do with it are as good as they can possibly be. Which talked to two, we may be able to incorporate an uncertainty image or band in there. So, you'll have a pixel by pixel estimate of uncertainty as well. So, we talked about the uncertainty. Where is it coming from? So, we start out level zero. Yay, you turn on the spectrometer, that's it, right? Perfect instrument. I'm sure Nathan mentioned all the little details that seep in there about ghosting across the panels and stray light and all the other issues. So, that's a source in there. That's our systematic air, band separation, how well is that grading actually separating the wavelengths when you say you have 890 nanometers, do you really? Or is it more like 891? And that does matter. So, we obsess over that and do the best we can. We calibrate that to the radiance. We measure calibration before and after the flight season in the lab. Before and after every line we use onboard calibration, both dark and laser. We spend a lot of time just making sure that instrument is really doing what it's supposed to do. Once you have that, that gets to your top of atmosphere radiance. And that's good value, but what we want is that reflectance on the ground. Account for the atmosphere. That's the piece I mentioned. How do you model what the atmosphere is doing to get that out of the signal so you really see what's on the ground and not just a picture of how hazy it is? We use the industry standard on that. There's several kind of well-accepted pieces. ATCOR is the one we use. They're all ModTran 5 based. That's in constant refinement for exactly how we apply that. But that works with your cloud cover, your shadows, thin, serious clouds, water, vapor, aerosols, zone pressure. Geolocation of it, if you're correcting this model, did you really have the path correct? We do a lot of work with that. Is your surface model correct? It's nice that we co-collect LiDAR so it allows us to tell about the exact orientation. That all goes in there. But every one of those, how well you estimate that, how well you measure it, is another source of error. That should get you to your surface reflectance. To help out with that, we do go out in the field, use field spectrometers, the tarps, and we also have all the observational system measurements that we can work in there as well to make sure we're getting this as good as we can. Once you're down to that, your calculation of indices, the whole family we just went over, is really dependent on exactly the quality of that. These are only as good as those. Now that's good because you're not introducing another source, but it also means you have no control over the error there. Using the Taylor series expansion for every one of the formulas for the vegetation indices based on all the uncertainty pieces that were presented from Nathan's talk and in the ATBD, the Algorithm Theoretical Basis Document for the reflectance product, we're able to come up with error estimates for the indices. Basically, how much error or uncertainty you're going to see depends on the values of the reflectance input. Here's one example. There are three plots for every index in that document based on what we kind of see for reflectance error. Your absolute best case reflectance error is down around the 1% ish, so that's your ideal conditions, perfectly illuminated, nice well-behaved atmosphere, not any kind of crazy geometry or hills or shadows. More typical, seems to be about the 5%. In your worst case, with a whole lot of shadows, really bad atmosphere, this is an amazing thing, but it looks like we're going to have two years in a row. We will have collected data over great, smoky mountains. We spend a lot of time sitting there looking at clouds, waiting for good conditions to fly. Somebody asked, why is it so hard to collect, and I mentioned, well, it's the great smoky mountains, not the great clear mountains. In the situation with the topography, as Courtney mentioned, so steep you walk on and it causes erosion, you might be up into more about the 10% uncertainty for your reflectance product there. So what we've done is based on 1, 5, and 10%, this is a 5% for NDVI, use that expansion formula and actually run through every possible combination of the reflectance, in this case, the near infrared and the red, and what would it do to your uncertainty? You can see that as long as you're up in higher reflectance values, 0 to 1, you're keeping your uncertainty down about 0.1, 0.2 for NDVI. So remember, NDVI range is negative 1 to 1, so that keeps us, you know, try to keep down around that 5% of your range, that's pretty good. You can see what happens is the lower values you get, your uncertainty blows up. Well, if it's 4 minus 3 over 4 plus 3, but they were both 0, then NDVI would become meaningless then too. You'd have 0 over 0. So that explains why that does that. So we did a surface for every one of those formulas for 1, 5, and 10% error in reflectance. Those are all in the document there. But I wanted to show you what we have there. Now, the reason we'll be able to convert this into an uncertainty band is we can say for this pixel, you know, 0.6 went in and 0.5 went in, and so that should put us at this value of uncertainty here. So that gives us an image. So that's kind of where we're going with that. But again, as I mentioned, I think it's a great document, but it's worth it just to see the surfaces there. It'll give you an idea of the uncertainty. So as you're using that data, using those products, using those indices, it'll allow you to better incorporate the uncertainty in that comparison with the ground measurements and hopefully give you a more robust publication, a better characterized result in the end. It works out well. You don't often see this. Remote sensing is kind of notorious for not publishing a lot of its uncertainty. As I mentioned, it used to be if you got an image at all, it was kind of a miracle. And we have a little history of kind of overselling what remote sensing can do. The classic case for that was in the 80s. We're going to have precision ag. I think everybody heard that term and it's come back now. You'll never have to worry about how to farm your crops again. Everything will be perfect and of course it didn't play out because the uncertainty was in there. And this creeps in on all sorts of interesting things. Again, just because it's on the ground doesn't mean it's in your data. So it's important to remember those pieces that are in there for error. You can have, again, you narrow down to just a couple bands. Different materials might start to look at like the same thing. I've seen that in a couple cases. It turns out drying green crops look an awful lot like somebody monkeying around with nitrogen chemistry to make explosives. So you would have an intelligent sector would look and say, there's a signal for a bunch of nitrogen here. And it would turn out it was just a farmer drying a crop. Well, if it's in Afghanistan, that's an awful big area of concern for people that are looking at imagery. And so you'd have to go through and look at the greater context. So when you're using these indices, don't just take it as a picture of it says it's green and healthy here. Use our camera data. Use our LiDAR data. See what's really going on. Use that extra piece. The context is everything. Remote sensing has really spectacular-looking results in the imagery. And it kind of has this effect of dazzling. But remember, it's a model. And the great piece I heard in that is the danger in modeling is you start to believe what you're saying. It's this beautiful picture. That's the accepted algorithm. This should be perfect. So use the other data. We have it all there for you. And it should result in some really interesting results for you.