Science Seminar: Linking Local Scale Community Processes into Global Scale Ecological Dynamics

I'm excited to welcome Renjan Mua Krishnan. Renjan is an assistant professor at St. Olaf College and he received his PhD from UCLA. He is an ecologist interested in the dynamics of major shifts in ecological communities particularly in the context of anthropogenic influences. He uses a combination of empirical and theoretical methods to understand the network of processes that control ecosystems. uh and he is going to talk to us today about linking local scale community processes into global scale ecological dynamics. Welcome Renjan. >> So yeah firstly let me just say thank you for the invitation. I'm really excited to be here and to have the opportunity to chat with you all and really thank you for taking the time for coming and being open to listening to about the project or about these projects in my general research. Um, so my plan today is actually to just give you a few different vignettes about different projects that I've worked on or really a lot of this is quite recent and ongoing work and so the different vignettes are going to be from different topics but really you know my thought about what's linking these is this idea of moving across scales like starting at sort of species or individual local scale processes um and then using approaches particularly sort of computational approaches to build that up to larger larger scale patterns and dynamics and sort of often making predictions at large scales and in a lot of ways I think that this is really you know it reflects my own trajectory as an ecologist. I started off as very much a in the field like processoriented scientists you know working in grasslands and rivers um and I think that that was an important experience. I think that type of research is really powerful, but for me sometimes it was hard to link that kind of very local work to the bigger questions and the more practical issues that I cared about for conservation and management, you know, thinking about large scale questions. And so for me and then playing out the consequences of these very sort of narrow or local processes at larger scales particularly using computational methods was the key to getting at these bigger kind of why does it matter type questions. And so I'm going to show you a few examples of how I um do that. And but again I want to start off by saying you know I think it is as much as the larger scales are important I think it's really critical to start at this individual or local scale understanding systems in their you know local context um you know whether you do sort of field work or lab work or greenhouse work I think starting there is really important and again like one theme that's going to come up a bunch of times is this idea of integrating across scales particularly the role of computational approaches and I'll talk a little more about that. Um but so right we start at this local scale and then sort of the first level of scaling up is thinking about sort of more complex community dynamics. You know how might species be interacting with each other? How might they be interacting or how might they sort of react to different environmental circumstances and sort of thinking about you know that sort of level of complexity. Um then you know moving even further is sort of putting these processes into real ecosystems or real landscapes. Not just thinking about them in sort of idealized contexts or single contexts but putting them into you know more realistic and spatially heterogeneous contexts and then you know from there moving up to you know regional continental or global scales as appropriate. Um I think this is really powerful. there's really interesting patterns and additional data opportunities when you're getting sort of really heterogeneous areas. But there's also, you know, challenges to working at that scale. You know, generally you can't work at global scales in any practical empirical way. You know, it's hard to run experiments at that scale. And so again, this is where I think computational approaches are really powerful for integration or data synthesis because they can scale a lot more easily. And also I think that it's a really powerful approach for integrating different kinds of information. You know, your sort of ecological theory, but also remote sensing data or environmental data about sort of changes over gradients. Um though I will say I think neon is like one of the very few exceptions to this pattern. Like you really can get field empirical data at really pretty large scales by using neon which is really cool and powerful but still I think for most of us using neon data is still kind of a data management and synthesis sort of approach. um we don't really for the most of us although maybe perhaps some of the people on this call they are thinking about logistics of sampling at continental scales but for me as a user it's more of a data kind of uh task. So anyways as I said I'm going to talk I'm going to give you a few different vignettes. Um firstly we'll talk about sort of testing the robustness of ecological theory across different systems and for that one I really will focus on using neon data but then I'll do two vignettes that are a bit more applied thinking about spread of aquatic invasive species and sort of global predictions of coral bleaching risks. So let's start thinking about ecological theory and how generalizable it is. And I should say, you know, for me, I'm very much a community collegist. So that's my bias. And I think much of what I'm going to say is most relevant at sort of that intermediate community level, but I think it'll be relevant to other sub fields as well. Um, and so really, I think the foundations of community ecology, they come from natural history and, you know, coming up with insights from watching natural history. I think many of us probably know the story of MacArthur's warblers that use different parts of trees to sort of partition niches and that sort of limits competition and allows coexistence. Um, and you know it's a story of like we saw this thing in this particular case, somebody studied it really intensely, but then they make some sort of generalizable inference, you know, about theory from it, which I think is cool. I think it's powerful. I think for many of us, we kind of got into this game because we love systems. We love, you know, the stories that we tell about these systems. And my first adviser was Mary Power, who I think, you know, to me she epitomizes this idea, this approach of like really understanding particularly river food webs. And from that deep knowledge coming to theoretical insights, you know, particularly work on things like tropic cascades. And so I think this is important. This is powerful. this is where our field comes from. But it also, you know, begs the question like, are these systems special? You know, are these all like unique stories about special places and special systems or are communities structured by more general rules or laws or or really is it all just local contingencies? Um John Lton, you know, posed this question of are there general laws in ecology, you know, almost 30 years ago now? Or he said, you know, perhaps is community ecology a mess just entirely driven by contingency? And so this is, you know, I think been a really motivating issue in community ecology for decades at this point. And what we need is synthesis, not just individual stories. Um and I think again here this is the kind of place where something like neon can be important and valuable for helping us think about questions of contingency. And so to think about this I decided a good set of ecological theory that we could use to think about generality are sort of biodiversity effects. You know, broadly speaking, we tend to think of biodiversity as, you know, quote unquote good, like things that we socially prefer. We generally think happen more when you have more diverse communities. The classic example of sort of biodiversity ecosystem functioning having this positive relationship, but also we think of ecosystems being more stable when they're more diverse or less likely to get invaded. So, their invasion resistance is higher. And there's theory and data to support this all of these ideas in particular contexts. But there's also been a pretty persistent debate about you know whether this is entirely true and how generalizable it is. We have varying results between studies. A lot of the really core theory comes from particularly temperate terrestrial glass grasslands. Um and you know a very limited set of ecosystem functions mostly biomass production. And while there is really good evidence about those things those are you know some set of studies. We also seem to have frequently a dis a misalign or sort of different results when we look at experimental studies which tend to show these patterns much more strongly versus observational studies that have you know often have more complex results um and don't always show those exact patterns um so the question is like you know how can we unify how can we think about what's going on in these different studies why are we getting these different patterns. And I think Neon is a really cool opportunity to try and think about this because you have essentially the same sampling and data collection scheme replicated across one a whole bunch of different sites um but two those sites represent different kinds of ecosystems. are collecting I mean not we you know neon is collecting a lot of different kinds of data that are similar in sort of style and form but in often pretty different types of ecosystems and also they're not just measuring one or two you know variables there's a whole range of community components and ecosystem processes that are being recorded at all of these places so we can think both across many different systems but also like are these patterns persistent or similar across different kinds of processes. So we'll think about a broader range of processes in a moment but just to start us off you know looking at so I picked one sort of diversity relationship to start with because I've done some work on this um and it's the idea of diversity invasibility relationships or the idea that more diverse systems are less likely to be invaded and sort of the classic data pattern that we expect looks like this where in more diverse communities or where at least where there's more diverse native species richness, we tend to have fewer invasive species as a signal of less invasiveness or invasibility. Um, and that's sort of the classic pattern. There's also an additional pattern that we've seen empirically that this is scale dependent. At sort of very small spatial scales, you see that negative relationship, but then as you move up to larger spatial scales, it switches from negative to a positive association. I think this is really interesting and I'm happy to talk about sort of these scale issues um if anybody's interested but for today the point is really just to think about you know can we get sort of these theoretical predictions in the neon data so like I'm not going to worry too much about the details but just we're going to see what look what does it look like when we have in the neon data um and so for a paper a few years ago we pulled a bunch of the neon plant community data. I realize it's not just 1 meter squared. That's a typo on that screen. We are we pulled the data at these different spatial scales from 1 meter to the full plant diversity plots. And you can see both that there is this negative the classic negative relationship at sort of that smallest 1 meter scale, but you also see this um um scale dependent pattern. So looking at a bunch of the neon data, we are seeing this expected pattern and we're seeing it across multiple sites which is really cool. We are getting at least on some level this match across systems of the sort of theoretical prediction an indication of some robustness. Um but I will say you know the secret underneath this particular data set is that it was only four sites. that paper was working with a subset of sites for a particular reason and so they were all these sort of southwest aid sort of deserty systems and so we saw that pattern but in a sort of limited subset of the neon data um so then moving from there I thought well let's look at it all like let's just throw all the plant community data and here I will say from here and forward I think it's going to mostly work not just the sort of one meter scale and we're really just going to focus on that negative relationship again just sort of for expediency's sake. Um and so if you throw it all together and you look at the overall pattern again we get that kind of classic negative relationship as we expected at that smallest scale. So awesome, right? That suggests that there is this good robustness of this pattern across many sites and systems, which is what we were, you know, wanted to test that whether or not our ecological theory is robust. But I wasn't satisfied. I wanted to look into this just a little bit further. So I said, well, let's like not just look at the data as a whole. Let's start splitting it up site by site. And so here we've got a bunch of different sites. Again, it's that plant community data and on the, you know, x- axis we have native species richness. Y- axis is exotic species richness in each of those individual plots. And as you can see, the patterns are actually not at all consistently negative. They're all over the place. Certainly, there are a bunch of sites that show those strong negative relationships, but there's also totally flat ones. There's totally positive ones. it is not, you know, that the the theory is not holding at all of our sites. It might sort of hold in aggregate, but what we're seeing locally is actually quite a bit of heterogeneity. So, plotting that in a different way, you can see we've just placed the trend lines for each individual site across our full data set as well as the red line is sort of the overall pattern. And again, you can just see this is we've got a lot of heterogeneity. Um, and so that initial pattern of like this sort of robustness of our theory is actually bearing out to not really be true. And it perhaps argues this idea that if our theory comes from really limited sets of systems or locations, we might not have really robust theory. And rather there's a lot of context or system dependence. And that might be the explanation for why the different studies that we've been looking at for the last 20 30 years are often times showing very different patterns because they come from places that are maybe driven by different processes. And so so right so there's our first question that we want to look like look at this robustness across systems. Um but also we can ask the question of you know the strength of the sort of biodiversity theory is that it suggests patterns across lots of different processes that this idea of biodiversity is always beneficial. And so here I've looked or I've plotted again the full neon data set across three different processes. One is sort of temporal turnover or stability of community composition. Um, one is the ecosystem function of carbon sequestration and then the invasibility is the data that we've just been looking at. And you can see again we have strong overall patterns although notably the carbon sequestration pattern is the opposite of our expected pattern. You would have expected higher sequestration, higher species richness. The stability is increasing by the line going down though. So that is following our expectation. But if we do again the same thing that we did with the invasibility data and we look at sort of patterns of individual sites across those data sets again you can see that strong heterogeneity some sites having positive relationships sites having negative relationships and just a lot of sightbysight variability. So again, it suggests that we have a lot of contingency rather than sort of consistency across sites. And this is true across a range of different processes. But we wanted to sort of delve one step further into this. And so we we did this with a bunch of different data sets, but I'll focus here just on carbon sequestration because we saw some results. Um, so if you took each of these individual sites and you know the trend lines for each of them and the slopes of those trend lines, those might be a parameter that you could think of as sort of describing the relationship for an individual site. And perhaps those site level differences can be predicted by some other environmental or other factor. And so we tested a few, but one of the most interesting and sort of stronger patterns that we saw is, you know, that sort of the sh the coefficient of those slopes, you know, from sight to sight do vary, but they vary in a somewhat coordinated pattern with respect to latitude, you know, with the most negative relationships being at low latitudes and patterns becoming more positive as we move to higher latitudes. Um, we've got some theories about what might be going on here, but this is still pretty initial and this is just one analysis. So, mostly I want to show you this pattern and just sort of show you this as, you know, an approach, but it's really not, you know, it's it's a really interesting pattern, but I think that it's not really explaining everything yet. Um, and so we want to sort of dig deeper into this to try and understand, you know, what might be explaining the rest of the variance and are there other drivers that we should be considering. So that's where we're moving forward in that with that work. Cool. So that's my first vignette that I want to talk about. Um, so now we can switch gears to away a little bit from really theoretical questions to something a little bit more applied. Understanding spread patterns of aquatic invasive species. Still we're going to think across different scales but but here we are going to limit to just the state of Minnesota. Again that has a lot to do with data availability more than the approaches. Um, so if you want to think about aquatic invasive species, you might think about things like zebra muscles or aquatic plants or aquatic algae that just grow in sort of really dense mats that potentially are excluding other species are potentially remodeling the ecosystem. As you can see in that middle picture, while zebra muscles are bad in a lot of ways, they clear the water up. So, this diver is perhaps having a really good experience, though there are a bunch of other ecosystem consequences. I also just want to say I'm really excited and proud about this work because we're currently working in collaboration with social scientists to really understand what effective or successful management is in terms of like what people are actually valuing, not just sort of the more quantitative metrics that we can easily quantify. Um, and so I think this is a broader project that I'm really excited about and I'll talk just a little bit about some of that, but admittedly today I'm going to focus on some of the more quantitative and quote unquote objective metrics. So we've got our system, you know, the state of Minnesota and our, you know, 10,000 plus lakes. Um, and so each of those lakes is a habitat or a patch that we are interested in, whether or not it's invaded, how likely it's going to be to inv get invaded, and you know, really thinking about which species are most at risk, not sorry, not species, which locations might be most at risk. And also, we might think about this for different species. But so if we are trying to figure that idea out of you know risk of different locations there's a variety of different components that are acting at the local scale that we want to sort of incorporate into this. The first is we think about habitat suitability and that's very much sort of a local scale condition. You know particularly things like water chemistry or local usage. Um that's really defining at the lake scale what and sometimes even the sublake scale, you know, is this a good habitat for an individual as it literally tries to establish? But we'll add to that things like connectivity. The movement of invasive species between lakes is largely driven by boers moving between lakes. And so how well different lakes are connected to each other is really important to the spread dynamics of of these invasive species. And we actually have had this really cool project going on or mostly some collaborators of mine for the last number of years that there's a pretty robust inspection program across the state of Minnesota where there's inspectors at boat launches and when somebody takes their boat in or takes their boat out of a lake, they they check and look to see if they, you know, have any invasive species hanging off of it or they're doing kind of the simple things to prevent spread. But they also tend to ask like what's the last boat that your lake was in or sorry what's the last lake that your boat was in or where do you plan to go next and you could see on his little tablet they're recording that data and from you know you know having essentially millions of those individual potential connections via single boats. we have been able to build up a essentially a network model of how connected different lakes are to each other across the state of Minnesota to to build in sort of that movement pattern or that um dispersal pattern. Um and then we also want to think about sort of the internal sort of population dynamics of invasive species and particular invasive species in these systems. And this is also we've got empirical data that can sort of feed into this analysis. But again, it's that's like the growth of individuals in specific locations. That's again local scale data that we're trying to integrate into these larger predictive models. Um, and we've started to do or we've done some of this work at working at making predictions at large scales with this little application, not little, this application called AIS Explorer that you can just access on the web that tells you sort of the relative riskiness of essentially every lake in the state of Minnesota for at least a couple different um, uh, invasive species. Um, and I think it's a really cool application, but it's got some limitations. Um, in particular, it's essentially a a static risk prediction and it's entirely deterministic. You know, each lake has a value and that's just the value of its risk. It also doesn't really integrate internal population dynamics and growth and sort of stochasticity. Um, and it has a relatively limited ability to sort of explore different management options. And so I think this is a cool application, but we wanted to sort of delve a little bit deeper into sort of the dynamical processes and the decision-m components. And so we've developed a new tool that is at a somewhat well in some ways at a finer resolution. Um but it's also a dynamical model where we actually build networks of connected lakes and then we allow you know invasions to happen in individual lakes and then they can grow and then they can spread between lakes and we can see how we might expect that that spread to happen over time and space. And while this is just sort of an idealized map, we can map this to real lakes with real quantities or qualities and connectednesses um in the model. We also have the ability to implement different management strategies. And as part of that, we also are structuring it to have a dashboard with information on the costs of these different management strategies. And you can build in budget constraints as you're sort of trying out different management strategies. I'm not even going to talk about the management and sort of decision-m side of this. I'm going to today I'm just going to focus on one other analysis we've been able to do with the model to sort of think about sort of relative risk or relative influence of different lakes across the state. And so the way that we do this is basically we can run simulations and because it's a model we can do essentially as many as we want and we can pick sort of random subsets of lakes you know and thus they'll have different networks of connectivity um and then we could just run the simulation on that subn network. It would be very difficult to run the model on the full lake network because it's just too big. But these sort of sub networks we can run simulations and then we can look at sort of overall outcomes like the number of invaded lakes or the mean abundance of invaders across the whole system and you know try to gauge what's happening you know in these different simulations. But the real power is again because computer time is cheap. We can essentially pick lots and lots and lots of random subsets of lakes and run sub and run simulations. So what I ended up doing was running 10,000 simulations with each simulation having about 2,000 lakes. Um that meant out of our whole subset or our whole set of lakes, each lake got included and about 2500 simulations. But then also there were 2500 or sorry 7,500 simulations that look at the dynamics across the state if you don't include or where that lake is not included in sort of the potential invasion process. And so across each of these different simulations we can calculate in particular the number of lakes that were invaded. Um and then we can compare in simulations where that lake was present in the network. you know, how many lakes got invaded versus when that lake was absent from the network, how many lakes got invaded. And so any single simulation is going to just be kind of arbitrary. But if we look across lots and lots and lots of these simulations and we kind of take the average of, you know, how many lakes are more or or are invaded and we can then say for each individual lake when it is or isn't included, you know, what does that do to sort of the average number or the the difference in the number of lakes that get invaded? So, you know, most of our lakes are essentially not that influential. Including them or not including them has pretty small influence on the total number of lakes that get invaded across the simulation. But we also have a number of lakes where when you include them, the number of lakes that get invaded in a simulation tends to be a little bit higher. um you know suggesting that that was perhaps an important lake that facilitated spread in some capacity. It's a little harder to know exactly what the interpretation of those negative differences is, but um certainly I think the positive ones there's a bit more of a a clear interpretation. So we can see that different lakes are sort of more or less influential. Um, but then the question much like we thought about with sort of some of that neon data earlier is can we try to understand at each location like what's making them different? And so here we did just try out a handful of different sort of lake parameters that we thought might be relevant in particular just this first one as an example. Um, we could measure for each lake across the whole network. We can ask sort of how central is it to the to the network of lakes and particularly in this idea of degree centrality like how many other lakes is it connected to and then we can plot you know the degree centrality of that lake relative to its influence on you know or it's it it does it lead to more or less lakes when included. And we can see again there's a pretty wide, you know, variance, but we also see a pretty clear positive relationship that more connected lakes if they're included in the network tend to lead to more lakes being invaded at the end of the simulation, which you know, I think makes a certain amount of sense. Like a lake that facilitates a lot of con connections would you could imagine influence simulations to have more invasions. And then here are just a handful of other parameters that we did that same analysis with. Betweenness centrality is just another sort of network shape or topology metric that's also about how central or core is a lake to the system. And we also get a nice strong positive relationship. The other sort of really strong pattern that we got was with lake size. So larger lakes tended to also have this positive influence on leading to more lakes getting invaded. That both sort of make sense in the context of larger lakes tend to be more connect or have yeah have more connections. So they tend to correlate with higher degree. They also probably have more people coming to them and they have just sort of in the model basically the potential population size and thus the potential number of propagules that a a lake is sending out into the world is higher as those lakes get bigger as well. Um, interestingly, one parameter that I certainly thought would be important but turned out to be essentially irrelevant was this idea of lake suitability, like more or less suitable lakes. That didn't really correlate very well with, you know, this idea of influentialness in sort of the spread dynamics. I still need to think about what I think might drive that, but certainly it was an interesting result. So overall thinking about what makes you know lakes influential I think we've got some interesting results but there's also a lot left to explore these different parameters that we've been sort of comparing you know are they totally independent are they correlated is there sort of latent variable that's driving all of these patterns we might be interested in looking about whether there are other spatial patterns perhaps other usage patterns or environmental factors that we could be considering And then lastly, you know, if we do this, you know, I think of it as sort of can we identify sort of hidden super spreader lakes and if we can like, you know, how might knowing those things influence or inform management? would we, you know, want to be over expending res or should we like bias our resources towards these um, you know, super spreader lakes that might not seem particularly important when we just look at them from other metrics. So again, you know, there's a lot to be done, but we're seeing some really cool initial results. Okay, so lastly I want to go and give you guys one more quick vignette. Again, just another sort of way that I have been using um local process local process data to inform predictions at larger scales. And here we really will be able to look at sort of the global scale thinking about coral bleaching predictions at really the global scale. And so I assume many of you know about coral reefs that they're this beautiful and valuable ecosystem and resource across particularly the tropics. Um though there are corals in nontropical areas too and but they also are highly susceptible to climate change particularly when we have high temperature anomalies. When water's warm, corals tend to expel their symbiotic photosynthetic algae. And when they expel it, it they sort of get this really stark whitish color. That's why it's referred to as coral bleaching. But those symbiotic algae are really important. They're providing resources to the corals. And so if they don't essentially recolonize with essentially better algae, those corals will die. you'll have, you know, mass mortality events. And so it's a really big, I think, important question trying to understand and most importantly like predict in kind of real time whether what reefs around the world are likely to be bleaching um because that's actually really uh used in a lot of management sort of contexts. So the Noah, the National Oceanic and Atmospheric Administration, they have a really cool office and program called Coral Reef Watch that do exactly this. They are taking particularly remote sensing data at the global scale and they're trying to make these predictions of across the globe um at essentially every day who is more most or more likely to be at risk of bleaching or when would they say we expect that there's probably a bleaching event going on here right now. And so this is their bleaching alert area product which is again a a global product at the I think 5 kmter resolution predicting you know where they do or don't think that there's going to be bleaching and it's really built on uh thermal anomaly data. So they they're watching constantly sea surface temperatures and they they basically are adding up sort of a cumulative thermal stress over sort of historical averages or historical maxima and they calculate what they call degree heating weeks. It's kind of in the vein of you might have heard of degree days from your introbiology classes. Basically when the anomaly is above a certain amount and they're counting over I believe 12 weeks. And so if you are one degree over that historical maximum for one week that's one degree heating week. If you are one degree over for four weeks that's four or four degrees over for one week that's also a four. And then based on that they they have a threshold for identifying essentially when they think bleaching will happen. Um so that's kind of the the standard sort of approach to making these predictions. But we also know that coral reefs are really complex and diverse ecosystems. They've got lots of different species and those species we know vary in their bleaching resilience. Um um sometimes uh you can actually see like really clear patterns of two different species right next to each other and some are completely bleached and others are not. And so if we know that there is this species level variation at you know local systems, how might we be able to integrate that information into these large scale predictions? Because right now where we stand is it's clear that there is local scale variation but at the global scale when we're making predictions this variation is essentially ignored. And so what I was really interested in doing is can we is figuring out if we can incorporate this local community data to refine our global bleaching predictions. So you know at the heart of sort of figuring out how to do this was firstly just getting data getting large scale global scale data on species distributions. Um, and there's lots of individual places where we have good species lists, but it's not necessarily comprehensive. So, we actually worked with data from the Aquamaps platform that actually basically create species distribution models for, you know, thousands of species. And then we pulled those distributions for essentially all the coral species we could find. Um, and we were able to get about 960 distributions that we thought were good for coral species. And then we were also able to pull that Noah thermal thermal anomaly data, the degree heating week data. And they actually have that data historically for the last 30ome years. So, we had all of that data and then what was also key to this approach is we had we were able to get a hold of a really big um corpus of times that actual human people went out to a real reef and assessed whether or not that reef was under a major bleaching event or not. And so over again 30-ish years we have 35,000 assessments spanning you know the tropics and about a third of those are events where they went out and assessed and said there was a major bleaching event about twothirds there was no bleaching event. So, we took all of those events or assessments and then we took all of that sort of environmental data about distributions and temperature. Um, and then we used machine learning approaches, particularly random forests and boosted regression trees to build predictive models of when you would or wouldn't expect bleaching that incorporated both sort of degree heating week as a predictor. That's our temperature anomaly data. But then also incorporated for, you know, all of those locations, we could say which species do we expect to be present at this location. And so if you know the species that are here and you know the temperature anomaly, what would the prediction be about essentially whether or not this system would be under a significant bleaching event? Essentially, we used a cutoff of greater than 10% reefwide bleaching as a significant bleaching event. And this aligns with essentially the level two alert from Noah's bleaching alert areas which they refer to as risk of reefwide bleaching and also they use this sort of 10% standard. So that's sort of the machine learning approach. So if we look at the way that this data or the model makes you know we can look at essentially the scale of model predictions and just to sort of orient you to this figure this is the you know the predictions that the Noah bleaching alert area makes. So it predicts again in that historical data we could go back and say at a location where we have an assessment would Noah have predicted a bleaching event and this upper number is essentially the overall accuracy of all assessments both bleaching or not bleaching and you can see they get about 2/3 right but if we look the red are essentially events where there was bleaching and blue is events where there was no bleaching. Um, when we look at the actual bleaching events, they're only getting about 31% of those correct. So, while, you know, sort of twothirds accuracy doesn't sound so bad in terms of the events that we really want to be predicting, they're not doing a particularly great job. But then if we look across our different models either so RF is random forest, BRT is boosted regression trees, DHW is a model that only incorporates that degree heating week data and all is the temperature and species distribution datas. And you can see that we increase our overall accuracy to you know 75 85 86% which is I think a pretty good increase over the the Noah predictions. But what's really I think most important and most powerful is if you look down at the accuracy of predicting bleaching events in with with the species distribution data we get to 74 75% which is more than doubling almost you know two and a half time is that right almost two and a half times as accurate as the the Noah data which is I think just a huge increase in accuracy which I think is a is a real step forward. in in sort of these predictions. But more than just increasing our accuracy, this also provides a couple other you know there's a couple other things that come out of this. So if you So these are essentially the Noah predictions. You know the original temperature data on the left and then you know their predictions of where there is or isn't bleaching on a given day. And so red is essentially they predict that there is bleaching. blue they predict low risk and then the white is sort of the warning but like sub threshold risk. Um and you can see there's some heterogene or some spatial variation that's reflecting the actual temperature variation but it's still like at a pretty gross sort of spatial pattern. And then if we look at our own data or our predictions, it's you can see quite a bit of additional sort of spatial heterogeneity. And this is driven both by the variation in thermal conditions, the DHW values, but also because we So this last map that I pulled up is essentially if you said the whole world was at a twoderee heating week anomaly, which Noah would actually predict to be under their bleaching threshold. And so they would just say across the board this should be no bleaching. we see that there's actually quite a bit of spatial heterogeneity in sort of more or less susceptibility. Um, and so our increased increased sort of like resolution of our predictions I think comes from recognizing that there are different species in our different locations. And then here we can also show that across different sort of sort of constant predictions. And you can see that even at relatively high anomalies, there's still some places where we would predict relatively low risk of bleaching. Um, and just in the interest of time, I won't delve too much into that. And I'll just offer one last piece of information, which we also were able to take our this data and sort of do a cross validation analysis where we really built the model with data only up to a specific year. So all the data up to 2000 year 2000 or 2001 or 2003 and then we can make predictions for you know essentially all locations at a anomaly of one degree heating week two degree heating week what would our prediction of risk be and then we can compare like 2000 is sort of our first prediction and then we said well what if we included the data from 2001 does that make the risks go up or down and we can keep doing that for every year up to 2000 or sorry 2020. And what we see is that when we incorporate the more and more recent data for a given thermal anomaly, at least the lower ones, our predicted risk is actually declining. Um, and I'll just give you the, you know, the quick sort of version of that that it s suggests that we are actually getting a potential acclamation. Like these reefs have endured warm or high temperature anomalies in the past and so now when they see that same high temperature anomaly, it's less likely to drive a bleaching event, which I think is a pretty cool pattern to see. So just sort of some overall, you know, thoughts from this section. Incorporating these sort of local species information certainly leads to better predictions, but I think it's also giving us additional insights to spatial and temporal patterns of resilience or susceptibility. Um, and the other thing that's really cool and I think where we want to move forward is applying this to forwardlooking forecasts which Noah already makes. And because our data is again very connected to and aligned with those Noah products, I think that's going to facilitate our forward-looking forecasts to also use that same structure. And because managers are already using that data, hopefully our insights can also get plugged into that. Cool. Okay, I'll just say thank you to my collaborators and funders. Um, I realize I may have gone just a little bit long, but hopefully we have at least a few more minutes for questions and thank you for your time and letting me talk to you about this all. Thanks. Awesome. Thank you so much. All right, so folks, if you have questions, you can put them in the little Q&A box. Um, you can also raise your hand and we can unmute you. But we have a couple of questions in the Q&A already. The first couple came in during the time that you were talking about the lake invisibility study. So um referring to the lakes that are the most influential in the number of lakes invaded um are they the most visited or most connected? So I think after the question came in you answered with respect to the the connection but was it related to visitation frequency as well? I I will say I don't think that we know at the moment partly because there's a little bit how would I say there's a little bit of confound between sort of the connectedness and between connectedness and sort of frequency of use. So it's a little hard for me to tease those two out but also I haven't I haven't tried to like explicitly look at that question. So I I certainly want to and I think it'll be interesting. I just don't know exactly what the I don't know exactly how it will play out. So I just don't know the answer. >> Sure. Thanks. Okay. Next question is also about the lakes. To what extent do the analyses or simulations include the management that is already being done? For example, the lake that has the highest connectivity has a negative impact. Is this because there's stricter management at that lake? Can a lake with management that enforces boat cleaning help reduce the spread if boats go to that lake? And therefore that would explain the the negative effect. So I would say that at the at the moment the simulations are not including any current or active management. That's just that's just not in that particular sort of what I'm calling the super spreader analysis. That's just not it's not incorporated. Uh so yeah. So maybe that's the answer. I'm I'm intrigued. I was trying to think back like, oh, is the are the like highest ones or like the individually highest connected lakes are is it showing a negative pattern? And that'd be interesting. I still don't really have a great insight yet or intuition about negative values and so I I just don't I don't have a great answer to that question. >> Yeah, cool. Okay, so now looking at the >> stop sharing my screen. Sorry about that. >> Great. For the first topic that you discussed, the diversity and visasability work asking whether the work that was looking across the larger number of sites has that been published and is that available yet? Okay, >> it's not available yet. We're we're trying we tried to get some money to support the work and that has not been successful yet. So, I might just try to publish it ahead of like in advance to try and support like the next grand proposal, but yeah, it's not published yet. >> Okay. I think I think we are all excited to see it one way or another. >> I mean, getting to sort of go through some of this data in preparation for this talk got me really reexited about working with some of this data. So, >> awesome. We have one more and this might be what we have time for. So this is coming from Dave Barnett who oversees the the data product for the plant presence and percent cover data which is the data you use you used in that first vignette. Um so he asks I came to Neon from diversity stability work and saw how much value data comparable across time and space could add. Uh have you thought about how accounting for abundance might shape these scale dependent patterns? In a 2023 paper, colleagues found that the positive native non-native richness correlations are mostly driven by both groups piling up in productive places. Totally. Yeah. I think that I mean do I I mean I guess this is abundance like that's just not the that's not the word that I had in my head but I think it is the point I think of like everybody likes to be in the good habitat and that's totally what I think drives the the positive relationship. Um the question is and I but I think that we see that again that larger scale. The question is at small spatial scales, do you still get something like competitive exclusion or displacement? Um, and they're and this makes perfect sense that like even if you are getting at the like smallest spatial scales, yeah, displacement or exclusion or whatever, I could imagine in sort of the good habitat the potential for other coexistence supporting mechanisms being higher essentially. And I think from a whole other thing I think we've seen some other data across latitudinal patterns where when you have in that case it's sort of at low latitudes which you know might be more productive but I think they're sort of less harsh in some ways. I think there's this potential for like more things happening that's kind of letting just like more species sort of be there. And so abundance wasn't the word that I had in mind, but totally I think that's exactly what's going on in some way or another. >> Yeah. Awesome. Well, we we are a minute over time. Sorry. I think Samantha has one more thing, but thank you so much. It was a great talk. >> Thank you. >> Fantastic seminar. Thank you, Ron. Yeah, just to thank everyone for coming. We appreciate you being here. And we're off for the summer, so we'll come back in the fall for more neon science seminars. Please consider nominating yourself or colleague today and share the opportunity with your networks and have a great day everyone. Thanks again Renjan. >> Thank you.