Science Seminar: Using spectral landscape ecology to infer patterns of intraspecific variation

I'm really excited to welcome uh Dr. Megan Cely Megs who is a post-doal research scholar at Arizona State University and a forest ecologist specializing in remote sensing and imaging spectroscopy. Her research seeks to understand how ecological processes operating across biological and spatial scales shape forest resilience. She integrates field ecology with earth earth observations to study biodiversity adaptation and ecosystem function across landscapes with work spanning riparian cottonwood forests in the southwestern United States and native and invasive forests in Hawaii. Outside of her research, Megs is an avid outrigger paddler and is currently training for a race from Ma Machai to Wahu. She re recently accepted a post-docctoral fellowship at the University of Utah where she will continue her work on cottonwood forests and investigate how hybridization may contribute to forest resilience in the face of rapid environmental change. So direct from Hawaii, let's welcome uh Meg Celely. Welcome Meg. >> Thank you. Okay, welcome everyone. Thank you all for attending today. Today I will be talking to you about spectral landscape ecology as a tool for inferring the spatial organization of inject specific variation and I'll be using neon imaging spectroscopy to do so. And so essentially I'll be looking at and using remote sensing as a tool for scaling up ecological inference. We can measure ecological processes in incredible detail at the individual and site level. But how do we scale those observations across landscapes? And so I wanted to first start by introducing NEON, the National Ecological Observatory Network, which many of you are probably familiar with. They collect long-term open access ecological data and that standardized uh those standardized observations are collected across diverse ecosystems and you can see uh their sites represented here in this figure. And so uh neon is essentially creating a means of understanding ecological processes by integrating observations across spatial and temporal scales. And so the type of data neon collects includes uh leaf level information at which we can directly measure functional traits that tell us about how plants acquire and use resources. We can then scale up from the leaf to the individual tree to characterize canopy structure and function. And then we can continuous to scale up to the site level or to a common garden to characterize variation within and among populations. And so this is a type of multis-cale observation framework that NEON employs to monitor long long-term ecological change. But since field observations are spatially limited, if we want to continue to scale eological inference, we need observations that are spatially continuous. And this is where remote sensing becomes particularly powerful. So NEON's airborne observation platform or AOP provides high resolution imaging spectroscopy and LAR data which allows us to extend our ecological observations from the individual sites to the landscape landscape scale. And so for this talk I'll be focusing on imaging spectroscopy which is a means of image formation from narrow wavelength intervals. And so here we can see a typical vegetation spectrum from 400 nm to 2500 nmters. And we can separate the electromagnetic spectrum into the visible from 400 to about 700 nmters, the near infrared, and then the red edge, which is the transition between the two, and lastly, the shortwave infrared, which goes um to 2500 nmters from about 1400 to 2500 nm. And so by measuring the light reflected off vegetation, we can capture information about biochemical and structural traits. Different molecules interact with light differently across the spectrum which produces distinct spectral signatures. And so because individuals differ in their biochemical and structural composition, each individual can have a unique unique spectral signature. And so a classic example of this is chlorophyll. So referring to the the visible wavelengths, chlorophyll absorbs more light in the red and the blue wavelengths and reflects more light in the green, which is this little hump that we see right here. These are the green wavelengths and this is why we see vegetation as green. And so it's similar for uh numerous other molecules and organels in the plants where different regions of the spectrum provide information about different aspects of vegetation. And so the visible wavelengths are strongly influenced by pigments. The near infrared is particularly sensitive to leaf and canopy structure. And lastly, the shortwave infrared or swur contains information related to leaf water and other biochemical properties. I would like to note though that uh these relationships are not onetoone. Spectral regions contain overlapping information and multiple traits can influence reflectance influence reflectance across the spectrum. So now that we are introduced to these data, what can we do with them and what ecological questions can we ask and conservation tools can we develop with imaging spectroscopy data and the scaling framework more broadly. So first we can um conduct we can create landscape scale species classifications. Um say we want to identify where cottonwoods exist across the landscape. We can use these data to understand their their uh spatial arrangement on the landscape. We can then uh build on this to understand biological variation because imaging spectroscopy is considered an integrated measure of plant phenotypes. We can use these data to characterize variation among vegetation communities or even within populations uh between populations within a species and then we can continue to build on this to predict change and resilience in these communities. And so for this talk I want to talk about a case study where we applied this framework to a foundation tree species Fremont cottonwood. It is a riparian species found in the US southwest. And because these trees drive ecosystem processes and influence biodiversity across multiple trophic levels. Understanding how these trees vary across the landscape is important for understanding ecosystem function and resilience. And while this case study was not conducted at NEON sites, we employed a similar framework by combining common garden and field observations with remote sensing data, specifically NEON's AOP data to scale ecological inference to the landscape level. And so here is the the range of Fremont Continent in the US Southwest. We can see populations in California, Arizona, Utah, Colorado, Mexico. But across its range, Fremont Cottonwood is not a monolith. There is sub substantial genetic variation among populations and in particular we see evidence for geographically structured ecotypes associated with distinct climactic environments. This figure from both at all shows um the three different cottonwood ecotypes distributed across the US southwest. We can see in yellow the central California ecotype which we will not be talking about for this talk. um instead we'll be focusing on the Utah high plateau ecotype represented in green and the Soran desert ecotype represented in blue and these ecotypes exhibit distinct characteristics even when grown under the same environmental conditions and we know this from common garden experiments which help reinforce a central theme of this work which is genotype environment and genotype by environment interactions. And so we can think about uh genotype effects. Um at its simplest it refers to the genetic background of the environment. Um so we can ask questions like do differences among ecotypes persist under the same environmental conditions. The environmental effects describe the conditions in which the individual grows and so changes in traits or performance are associated with the environmental conditions. And lastly, the genotype by environment interactions occurs when different genotypes respond differently to those environmental conditions. And so essentially the effect of environment depends on genotype. So now that we have some of the vocab out of the way, I want to talk about uh first talk about landscape scale species classifications. So we have uh we collected data using uh NEO's AOP. We collected imaging for trust and lighter data across two watersheds in the US southwest. We collected data in central Arizona which represents the snoring desert ecotype. These data were collected in 2021 and we collected some more data in Utah and Colorado and that represents Utah high plateau ecoype and those data were collected in 2022. And so this type of study design where the data were collected um at different times um in different regions is not unique to this case study. Logistical constraints often limit our ability to collect airborne imaging spectroscopy data resulting in temporarily distinct collection times. And I'll touch on how this affects our classifications uh in a bit. I also want to note that these data I separated these data into 17 distinct sites. Each site is 6 km by 6 km. And I did this to um uh make the analyses a little bit easier um and and have uh geographically constrained regions. So when we classify a species using remote sensing data, the rationale behind why it works is that species differ in their leaf and canopy structure as well as their chemical composition. And this results in unique canopy reflectance for each species. Um and so we often refer to this as a unique spectral fingerprint. And so we feed our machine learning algorithms uh examples of canopy reflectance from the different species represented in our data set. So it learns the the unique spectral fingerprints of each species. Once that we trained our machine learning algorithm of choice, we apply it back to our data and we get our species classifications. Now, this tends to work really well in small geographic regions, but when we start to try to scale up our species classifications to larger areas, we often get mclassifications. And these mclassifications um could result in errors in our biological inference. And it can also result in lost time when it comes to conservation effort. So, if we're trying to identify where an invasive species is on the landscape and we identify an area where we think it exists, but it doesn't actually exist, we might waste a lot of time going into the field and and checking that. And so, I wanted to ask why do these large species classifications fail? And there may be a number of reasons. One being differences in observation conditions. There are differences in atmospheric conditions, sun angle, viewing geometry. As we mentioned earlier, we have uh data from two different time points and so as a result the observation conditions um are very likely different between the the two different years. We have biological variation. So coming back to that genotype environment and genotype by environment interactions because we have differences in the environment and in the the genotypes on our landscape and this results in differences in physiology, canopy chemicals, phenology, canopy structure. And lastly, we have the geographers's favorite spatial autocorrelation which follows Tobler's first law of geography where everything is related to everything else but near things are more related than distant things. So when we're thinking about um species classifications, this indicates that models can learn the location or the local context of a region rather than generalizable species characteristics. So in order to test why these large scale species classific class classifications failed, we used our um neon data or neon imaging spectroscopy data at our different sites. And we first classified we used um two regional classifiers to identify where Fremont cottonwood exists on the landscape at our sites. And the reason why we use this method instead of using the the training data set that I collected in order to create these classifiers is that the training data set was spatially biased to where we could access which regions where we could access um the trees and in order to identify them. And I wanted to create a spatially robust spectral library. And so in order to create this map, I went through an iterative process to get a more conservative estimate of where these trees exist on the landscape. And if you want to read more about that, I'll refer you to um Cely at all in 2026. But for now, I'll just uh skip over that to say that we created the species classifier and we created a to I to gather a spectral library of our cottonwood trees and all the other species represented on these landscapes. And so this figure shows the average cot can cottonwood reflectance at the canopy level for our Arizona data set which is in pink and then our Utah Colorado data set which is in green. And so in order to look at how these classifications vary across the landscape, I separated the data into uh 70% training and 30% testing within each site. And then I trained a classifier on the data from each site on that 70% training data set. And then I tested that classifier across every single site using the 30% testing data set. And I did this for every single one of the sites. And then I looked at how class how the classification varied across environmental gradients. So I used a linear mixed effects regression model and looked at how the classification changed across environmental gradients and I iterated through this a 100 times and I got um these figures which are partial dependence plots and you can see there are a number of environmental variables that I threw into these models including geographic distance which gets out that spatial autocorrelation but most of these were not very important and so I want to just focus on winter precipitation, spring precipitation, mean annual temperature, and ecoype, which are our four most important variables. And so on the on the x- axis is environmental distance. It's a standardized environmental distance. And the y-axis is classification accuracy. And um uh one of the the major results that I found was that it's environmental distance increases, classification accuracy decreases. But what's the biological rationale for these variables in particular? Why were these variables um the most important when um determining classification accuracy across space? So I'll start with our fourth most important variable ecotype because the easiest to explain where we included our Arizona data set versus our Utah Colorado data set. So we had differences genetic differences. There were two different ecotypes but we also had differences in um the time of collection. And so we weren't able to entirely parse apart those u different sources of variation. And so that that variable includes both. And then for the other variables uh we can turn to our field and common garden studies to help explain them. So starting with our seasonal precipitation metrics, uh it's the the the seasonal precipitation metrics that we found being very important for classification accuracy across the landscape was uh where the same uh were the same variables that were predicted to to determine uh genetic connectivity across the landscape for Fremont Cottonwood. So this figure is from Kushment at all in 2014 where they went out and sampled Fremont on the landscape and they developed a resistance to gene flow surface layer which is the figure that you see and they found that overall stream conductivity plus seasonal precipitation were u the the major var the the best variables for for developing this resistance to gene flow layer. And so if our classifications are capturing genetic differences within Fremont Cotmwood across space, it makes sense that our classification accuracies would be related to the same environmental factors that structure gene flow. So then next we looked at the mean or mean annual temperature variable and in order to understand what's going on here, we referred to some of our common garden studies. So there were uh three common gardens that were established across an a thermal gradient. So up here in Utah, in southern Utah, we have our Canyon Lands common garden. It's our coldest common garden. We have Aquafria in central Arizona. And then Yuma, which is our our warmest common garden down in southern Arizona. And populating these common gardens were Fremont Kwood from numerous source populations that spanned a thermal gradient. And so at these common gardens, uh, Cooper at all in 2019 measured survivorship and they looked at survivorship versus budset day of year. And well, we'll talk about uh, phenology in at the end. I want to focus instead on the source population temperature. So the colors of the points and we can see that overall the trees that had the highest survivorship were the trees that um, match the common garden in terms of their mean annual temperature. So the w the trees that survived at Yuma, warmest common garden were the trees that were from regions with warmer meanial temperatures and vice versa for canyon lands. And so uh this work and others has found that Fremont Cottonwood has adapted to different thermal regimes. And so at these common gardens, it has resulted in higher survivorship for trees from areas that are more similar thermally to the common garden. And while there's a wide body of literature of ongoing work looking into why this is, I wanted to highlight a study by Moran at all that provides a particularly nice example of genotype by environment interactions showing that cottonwood populations differ in their physiological responses to thermal stress. So here this figure we have on the left to model conductance. So how much water are the are the plant is a plant pumping through its leaves? And then we have leaf temperature on the right. The x-axis is transfer distance. So the difference in um in mean annual temperature between the source population and our common garden. And here the all the data was collected at Yuma, our warmest common garden. And uh the data were collected in May which is a relatively cool month. The trees are are relatively happy. and then in August when it's really warm and the trees are thermally stressed. And so what Moran Adall found was that in May there wasn't a really a relationship between either stormal conductance or leaf temperature and transfer distance. But in August uh when it was nice and hot there there was a relationship where trees that were more similar to from regions that were um more similar thermally to our yumo garden were had higher stomatal conductance which resulted in in lower leaf temperature. And so from this they concluded that trees from warmer populations regulate their leaf temperature via transpiration. And so here we see a an example of these genotype by environment interactions where trees from different areas are responding to the environment differently. And in our classification accuracy, we found that accuracy declines as mean annual temperature becomes increasingly dissimilar from that of the populations we used to train the classifier. And so Matt appears to structure mean annual temperature appears to structure how well we can distinguish those populations spectrally. And the same environmental gradient temperature helps explain differences in the physiological responses among populations which then suggests that the spectral distance differences that we are detecting may reflect adaptations to different thermal regimes. Okay. So why do landscapes large scale uh species classifications fail? We found largely due to biological that it's largely due to biological variation of our target species on the landscape. And we found some evidence for differences in observation conditions affecting our our species classifications. And lastly, we didn't really find much evidence for spatial autocorrelation. So then how can we improve our land our large scale species classifications understanding uh what's limiting our are the the transferability of classifications across space. And so for this we turn to again our common garden experiments where we collected uh leaf level reflectance data of our of our trees uh existing at the common gardens. Um and so here you can see the reflectance of leaf level reflectance and the the lines are colored by common garden. So Yuma is red, Canyon Lands is blue and Aubra Fria is gray. And so one of the things that pops out is that we see that there's a a very obvious difference in our visible wavelengths um at Yuma. And then when we visualize these data in principal component principal component space, we can see that Yuma is separating from aua and and canuly land.and. So we're seeing this this uh environmental effect in our spectra regardless of of the source population and where these trees are from. And so to kind of dig into this a little bit more, we wanted to look at where what uh regions of the spectra are more heritable. And so we use uh variance partitioning to look at the different sources of variation. And I apologize the the bar is covering up some of the bottom of the of this graph in the citation. But um the in this figure we have our mean reflectance uh for reference in in the dotted line. We have heritability which is our solid black line. And then lastly we have the different sources of variation um as different colors. And so here we can see our the variation coming from our common garden in purple. And we see a large effect of common garden in visible in the visible and the reddish wavelengths. And then um we have different sources of variation that rel relate to genotype in the other colors. Blue is residual but everything else other than purple and blue is related to genotype. And so we can see from here um both in the heritability line and our variance partitioning that um heritability is much lower in the visible wavelengths but it's a lot higher if we look in the near infrared and the shortwave infrared. And so we use this information then to essentially cut apart the the electromagnetic spectrum and use different regions of the spectrum to test the transferability of our species classification from Arizona to Utah Colorado and vice versa. And essentially we confirmed our hypothesis and that species classification transferability was highest when we use the shortwave infrared. So that region of the spectra that represents biochemical variables um and it was most uh conserved across our our three different common gardens was also the the region of the spectra that helped improve the transferability of our species classifications. And this is really important when we're considering the types of remote sensing data that we're collecting. There are relatively uh few sensors that collect imaging spectroscopy data. NEON is one of them. But a lot of our remote sensing data comes in the form of uh drones and satellites that collect data primarily in the visible in the red edge which is the most responsive to environment um as our data showed. And that's great if you want to track physiological responses or stress over time. But if you're trying to do for example species classifications it's much more challenging. And that's not to say that uh multisspectral satellites don't uh collect data in the shortwave infrared, but there it's it's relatively understampled. Okay. So, putting aside our uh species classifications now that we have them, we're then going to use these data to understand the biological variation of cottonwoods across the landscape and then try to predict change and resilience. So coming back to that idea of of reflectance spectra as an integrated measure of the phenotype, there are a number of different hypotheses that suggest this. Uh there's a spectral variation hypothesis which assumes that vegetation or reflectance spectra serve as a proxy for taxonomic information because spectra can capture variation in leaf chemistry, structure, and physiology. And so we we build on those on those hypotheses to estimate um functional variation and functional and diversity across the landscape. And so I use methods developed by Ozner at all in 2014 where they took the the canopy reflectance data and they applied a cayman's clustering and then they used that uh those clusters as a uh they termed it spectral species. But essentially what they did is they applied the Shannon diversity calculation and the Bray Curtis dissimilarity index to estimate alpha and beta diversity. And when they did this across uh Peru, they found that they were able to estimate um alpha and beta diversity rather well. And so I adapted these methods and instead of using them across um uh multiecies communities, I applied them just to our cottonwood species to get at or inspecific variation or within species variation. And so starting with our our results from our alpha diver or within species uh functional variation as calculated using the Shannon index, we looked at the inpecific spectral variation across our environmental gradients um using a linear mix effects regression model. And these are our partial dependence plots. And so again we tested in numerous environmental variables but here the most important variable was maximum temperature. And what was interesting was that with maximum temperature the relationship was not linear. We had um lower intpecific spectral variation in areas that were more in more thermally stressful environments. And so our intpecific spectral variation peaked at um intermediate maximum temperatures. So now when we think about uh what the future has in store and what's already happening in the southwest where the southwest is uh warming rapidly, what does this mean for cottonwood trees? Um if we're if we have lower intpacific spectral variation in areas that are warmer and these w warmer areas are becoming more um prevalent. And so I wanted to to come back to this figure where we looked at survivorship and especially where we looked at survivorship in Yuma, which is our our hottest common garden. And survivorship is is highest in populations that are that are adapted to warmer areas. And those and according to to our results, those trees that are more likely to survive in these future heat waves have lower intpecific spectral variation suggesting that future climate change may select for a narrower narrower range of phenotypes. So then shifting topics just a little bit, we then looked at our between sight spectral variation and so we use the break Curtis dissimilarity index to to look at the spectral dissimilarity between our sites. And so this is the correlation matrix um where the yellow represents regions that are more dissimilar and teal represents regions that are more similar. Unsurprisingly our Arizona sites are fairly dissimilar from our Utah Colorado sites. And then we used this this uh similarity matrix and compared it against um uh environmental distance matrices across our landscape. And for this one, we uh added back we added in that resistance to gene flow surface layer that we referenced earlier. And we found that um this resistance to gene flow was the actually the best predictor of of our our of our beta spectral diversity uh dissimilarity. So this first of all underscores our earlier hypothesis that um our classification accurate our classifications were capturing underlying genetic variation because spectral beta diversity between sites was higher where resistance to gene flow was higher and this suggests that reduced gene flow is associated with greater differentiation and spectral composition among populations. Interestingly, it also emphasizes the results found by Kushman at all in that human modified stream flow and future drought may have reduced genetic connectivity between our populations between our cottonwood populations. Okay. And then I lastly wanted to uh go from looking at our within or intrapecific spectral variation and turn to a slightly different um ecological focus which is phenology. And for this we asked uh does a phen phenological sensitivity of Fremont cotwood differ among ecotypes? And so phenological sensitivity refers to how responsive a tree is to interanual variation. And the the inspiration for this work came again from those common garden experiments where Cooper at all in 2022 looked at budet and bud flush across a common garden. So first they found that uh buds set and bud flush differed between the different populations even with a common garden within even within a common garden. So we have these different genotype effects within a common garden. But then they also looked at plasticity of bud set and bud flush. And they found that overall the trees from warmer source populations were more plastic, more responsive to differences in in common garden temperature and they they changed their their bud set and bud flesh the most. And so I wanted to see how this played out on the landscape scale and see if we these patterns were detectable in situ. And not only that, but what does this mean? Uh when we're looking across the landscape and we're looking at these populations, how are they changing in response to to climate climate change? So in order to do so, uh we turn to our landscape scale species classifications of Fremont cottonwood that we created using our NEON data across our sites. We downscaled these uh data to 30 by 30 meters to match LANCSAT and we collected 25 years of LANCSAT data. This is the normalized difference vegetation index which is essentially a measure of greenness and we looked at greenness uh over time for each of our our cottonwood pixels um at all of our sites. And for every year and every pixel, we fit a double logistic regression curve to to look at the change in greenness over time. So you can see winter and then spring green up right here or summertime the NDVI stays relatively high and then we can see the fall scinessence at the end. And so we use this curve to identify the point um at which uh the 30% of this this amplitude which is how we defined spring green up. And so using this, we then looked at um change uh simp change over time of our Utah high plateau populations versus our sonor and desert populations. And we found that overall the sonor and desert Fremont cottonwoods accelerated in the spring greenup in the past 25 years whereas the Utah high plateau didn't really change over time. And now the question is is this uh is this change related to the fact that um the sonor and desert ecoype has formed faster than the Utah and Colorado or is it a result of um more higher phenological sensitivity or is it a combination of the two and so we found that it's likely a combination of the two because and I apologize again I don't know how to make this bottom part go away but we looked at u minimum temperature growing degree days total precipitation This is uh maximum vapor pressure deficit. And lastly uh chilling degree days for the winter and spring uh before that year. And we found that overall the snow and desert populations had a greater phenological sensitivity. They were more responsive to our inner annual variation as you can see from the slopes of the red lines u which represent the snor desert versus the blue lines. And then we threw uh this into a model to just confirm these results. And we found that the most uh the variable that was the most important for predicting spring greenup timing was our our thermal variable, our brewing degree day. And not only was it important for predicting spring greenup timing, but it differed between the two populations, the slope was much stronger for a sonor and desert population. And so next we wanted to look at what this means for the future. And so we applied these this model to two different climate scenarios from 2030 to 2100. Um we have uh less aggressive warning warming on the left and more aggressive warming on the right. And we found that regardless of the the climate scenario, the spring greenup timing of Fremont populations, Fremont cottonwood populations will continue to diverge over time. Okay, that was a lot of information and uh thank you all for for sitting with me through that and I just wanted to quickly go through the the overall conclusions uh to to summarize everything that that has been discussed thus far. So landscape scale species classifications are limited by the spatial structuring of of tree phenotypes. And if we use only spectral phenotype or spectral features conserved across environmental conditions um this will improve our our classification transferability. Fremont cottonwood intpecific spectral diversity as a proxy for functional diversity was lowest in our thermally stressed stressful environments and our functional beta spectral diversity was best explained by resistance to gene flow surface. Again that integrated uh seasonal precipitation and uh stream connectivity and spring greenup timing of the warm adapted sonor and desert ecotype was more sensitive to interanual climate variation. So then what does this possibly mean for the future? It the implications are that anthropogenic activities that disrupt riparian connectivity alongside climate change, higher temperature, more drought uh may constrain the adaptive adaptive capacity of these populations. We also found that continued warming is likely to amplify phenological diver divergence among Fremont cottonwood ecotypes with potential cascading consequences across multiple trophic levels. And with that, I'd like to thank everyone here um on this slide. They without them, this work would not have been possible. They were all very important uh contributors to to all of the work shown here. I'd also like to thank Neon for for hosting me and allowing me to share this research with you. And then lastly, thank all of you for for coming and showing up and listening to this talk. And with that, I think we have some time for questions. >> Awesome, Megs. That was great, man. There's a ton of uh ton of really interesting research backed in there. Just a reminder to everybody, please uh type your questions in the Q&A box. Looks like there is time for many questions and so actually I have uh one question to kick us all off from Shashi Kunduri. Uh hey Megs great talk. I have a question regarding the cottonwood classification map you showed from Celi at all 2026 paper. Did you use any topographic variables for the classification besides the spectral signatures? >> That's a good question. So, I didn't use topography for these classifications, but I did use the LAR data collected u by the NEON AOP to look at uh canopy height. And so, I removed all vegetation that was below 2 m so that we didn't have to include grasses or low shrubs, but I didn't include elevation or aspect or anything like that in the classifications themselves. >> Okay, great. [clears throat] I've got one additional question to kick us off. The you you were mentioning something about the the sphere bands tend to show a lot more information and discrimination among the ecotyp ecotype characteristics of of the cottonwoods. How well do you think that approach will apply to other species, other plant species that perhaps are less studied kind of across large regional scales? >> Yeah. Um, so that so I can answer that by talking a little bit about u my work in Hawaii. So I'm also doing classifications in Hawaii to look at invasive versus native species. And since uh doing this work on cottonwoods, I have started testing out um just using the the swear bands to for our classifications and I found that overall it works a lot better. Um for my dissertation, I mapped out OIa lea across Big Island and I had there were a lot of challenges associated with that since it spans numerous environmental gradients etc. Um and it's a very plastic species. But since switching to using kind of this swore only method, I found that there have been numerous advantages. One one is that I see fewer flight line artifacts. Um so when you collect uh your imaging spectroscopy data, right, the plane uh flies in a line and it and then it can circle back and then fly in the next line. And so when you do so uh you have differences uh sometimes in between uh the different lines that's a result of or observ differences in observation conditions. Um and so sometimes we can we can see that perpetuated uh through our species classifications. And I found that when I use just the swur I see fewer of those but then I also see less of you know random species that are considered that aren't even related to OIa. So, I've been finding that it works better is kind of the short story in in my study system in Hawaii, but I haven't tested on other species. >> Cool. Thank you. Another question from John Mickelson. How about looking at any soil or geological parent material gradients? >> So, I didn't do that for this project. Um because the Fremont cottonwoods they exist along riparian corridors and usually those corridors are really narrow and so the the data sets that we often have are don't quite match in terms of um spatial resolution in order to capture to capture that um information. Um, also the in the riparian corridors, it tends to be just rather silty and and it's there's less variation than if we were to like look more at the highlands. And yeah, so I I didn't include it um because for this uh specific study area just because we were looking at the those relatively narrow corridors where they're usually rather silty and um quite changeable. So >> great from San Sharma. Do you find >> um John John, can we give Claire a chance? Claire has her hand raised to ask a question over audio. >> Oh, absolutely. I'm sorry, Claire. Go ahead. >> Yeah, thanks. Hi, Megan. This is a really interesting work. Thank you. Um I'm curious, you had in the common garden experiment uh an estimate of heritability of reflectance at different wavelengths and I wondered if you could just talk in a little bit more detail about uh how that heritability estimate was developed. Um yeah so this is taking me back a couple years so uh forgive my uh my me for not remembering everything but we use variance partitioning to look at the relative contributions of common garden versus a number of of of variables that related to the genotype because at these common gardens we had different ecotypes we had different populations represented um and then we also had some some clones as well. And so we put all of that um those those information and put that th and so when we v partitioned the variance for those we we kind of lumped all of those uh variables and then we separated it from the environmental the common garden and the residual and then I don't remember the exact calculation that we used actually one of my my collaborators was kind of the one who who primarily uh came up with with those formulas. Um if you want we can refer I can also refer you to the the paper where we uh actually have all the methods for that. But um it was Celi at all. It was in Oohia was uh one of the papers but um yeah so essentially we partitioned the variance and then we looked at the uh heritability was a calculation of the relative influence of our genotypes versus our common gardens and residuals was kind of the the bottom line of it but I I'm sorry I can't tell you the exact the exact math off the top of my head. >> That's great. I'll I'll check out the paper. Thank you. from Sancha Sharma. Do you find the geographic distance acts primarily as a proxy for unmeasured microclimatic variables or is local adaptation playing the dominant role? >> Can you repeat that again? Sorry. >> Do you find that geographic distance acts primarily as a proxy for unmeasured microclimatic variables or is local adaptation playing the dominant role? So yes, likely um our geographic distance is capturing variables that we that we haven't measured um and didn't or didn't include. There are a number of those including stream flow for example since we are working with pairing systems. Um or is it a measure of local adaptation? I'm sorry. I don't quite understand how those are entirely different and kind of what the question is trying to get at because the we know from our common garden experiments that there are different local adaptations of these species to different regions uh usually across a thermal gradient. And we are able to capture that in our data and we've shown it kind of numerous times in numerous different ways that our our remote sensing data is capturing those different local adaptations. Yeah, I'm not sure quite sure I answered that question, but I I don't know if I quite understand the the question itself. >> Okay. Um I'm just going to read out a comment from Courtney Meyer. Not so [clears throat] much a question, but it is fascinating to me how the spectral data reveal the limitations of the concepts of species when it comes to predicting how ecosystems may respond to change. >> Oh, thank you. Um yeah, I I think that I think that like using these remote sensing data and trying to under and think of of spectroscopy as a means of understanding these trees beyond just you know what we can go on the ground and measure because we can predict a lot of those different measurements um uh using spectroscopy data. There are a number of papers that have come out in recent years showing that and I think that uh as we as we start to accept it as as a form of understanding biological variation, it opens up a lot of different avenues for exploration and understanding our our landscapes because it's a much easier to to go and and capture imaging spectroscopy data than to measure like the actual, you know, physiology of of these trees or the canopy chemicals or all these different um components and so I'm really excited about what it kind of holds for the future and what we can do with it. So, thank you for the comment. >> Yeah. And a comment from uh Michelle Ptorius or a question. Did you run into saturation issues? I know tassled cap parameters for example brightness etc can saturate at high values. If not an issue in your area, should this be a consideration for applying these methods to other regions? like saturation for the reflectance data. No, we didn't. Um yeah, that wasn't an issue. I've never actually come across issues with saturation for data. I will say uh that for all the data that I present uh we applied a brightness normalization to all the data to kind of standardize brightness uh differences across different sites or as a result of of different observation angles and conditions. Um but yeah, no brightness issues, although that's probably more on the the technical neon side as they they had issues with that. they probably um you guys uh and Ian probably uh corrected for it and and fixed it and I know you guys have rigorous um um calibration u procedures and standards that that you guys do. But that's a yeah I I've not personally had to to encounter that. But I I think that that's more of the technical side of the the neiop that I don't always get into. >> Right. Okay. Uh question from Sophia Gman. Have you thought about how image spectroscopy perhaps only captures the reflectance at the top of canopy levels? And if imagery spectroscopy could be used to somehow look beyond just the top of canopy of ecosystems. Can it be used to look beyond the canopy itself? >> Like into the understory. Yeah. So I get this one quite a bit because oftentimes what we're interested in is the future of the canopy, right? What's growing underneath? Is it an invasive species layer or is it is it native regeneration? Unfortunately, based on the nature of the data, we can only look at that that top canopy layer. We can't really look underneath. Sometimes we get u glimpses into the understory if the the canopy isn't fully closed. Um and but it's just essentially what the the sun bounces off of is what we get. This is where the LAR data can um uh can be useful because the wavelengths of light that uh is used in LAR data can penetrate through the canopy. And so then we can start to get the structure of the overstory and the understory. But then you're just looking at structure and that has its own complications because you don't get the full vour spectra that gives you um chemistry composition and whatnot. So unfortunately no we cannot go past the the canopy layer as much as I would like to but that is a limitation of the data set. >> Okay. Great. Just as a kind of a addendum to this great presentation and question answer uh session. This project was a a neon research support services project I believe funded by national science foundation. This is people probably noticed it was not conducted at NEON sites. This is an example of a an external research team applying for funding and then contracting NEON to perform data collection for their specific project. So shout out to the team that uh put this whole thing together and it was great working with you guys. I'm going to uh close out questions and turn it over to Samantha. Thank you mix. >> Thank you. What a wonderful talk. Thank you so much for all that thoughtprovoking content and those interesting environment genetic interactions. Really excellent. If folks are interested, our next science seminar is going to come up um one month from now in October. I put the link in the chat, but the talk will be about from field the lab a full AI solution for biodiversity for uh a scientist will be presenting from Microsoft's AI for good lab. So it should be interesting. We hope to see you. And if you'd like to refer back to this talk, there should be a recording posted within the week. So, thanks everyone so much for joining. Thanks, Maggs again for a great talk and have a great rest of your day. Bye all.