 So our next speaker is Andy Wood. Andy Wood is a scientist here at EMPA and he also kindly co-led one of the tutorials, namely the hydro tutorial. Andy is a hydrologist by training and he will be talking about S2S full casting and hydrology. We're looking forward to your talk, Andy. Thanks, Judith. Yeah, I'll talk about some methods and concepts and trying to do S2S prediction for hydrology and screen flow. First part of my talk will mainly be about seasonal screen flow predictability and these concepts of where we see sources of predictability and then I'll give some examples of ways in which the community is trying to harness climate and watershed predictability to create real-world forecasts. So one of the main things to think about when you're trying to do S2S prediction hydrology is that there are several sources of scale involved. It may be a little bit different from climate forecasts in that regard and climate forecasts, you think about the land surface and a lot of the driver of that predictability comes from the oceans and hydrological predictability or screen flow predictability. You think about the watershed, the amount of moisture and energy sort of in watersheds that needs to evolve over the forecast period and so that forms one major source of predictability that we call initial hydrologic conditions. And then the second form, major form of predictability is meteorological predictability. Basically how well can we estimate future climate and weather that will drive the response of the basin after starting from those initial conditions? And so in hydrology for many, many years since we didn't really have great knowledge about future climate conditions when hydrologic prediction really got going to try to drive stakeholder communities like water management and energy management. We mainly tried to harness the initial condition, the initial hydrologic predictability even though there were these two sources. So back in the 70s when climate forecasting was not well-developed, you at least had the ability to take a watershed model, spin it up to try to get a good estimate of the current moisture in the catchment. And then the main technique that evolved was called ESP which is basically driving that initial condition forward using historical sequences of weather from the starting date that you were making before the initialization date. And then the uncertainty of just having to select a sample of past weather sequences would emerge as the forecast that you time became longer. And so this is kind of a depiction of that framework. One of the main weaknesses of this framework is that it assumes that you know these initial conditions perfectly. And I should say that one of the main ways that people have tried to incorporate climate forecast information into this ESP approach is to take those historical sequences and then to condition them based on information about climate as I'll show that can come from multiple sources. And that's called a trace weighting, ensemble trace weighting approach. There are a lot of papers on that to go back quite a bit. In any case, I've been doing this kind of research for a long time and back in the early 2000s after making quite a few presentations on using this technique and kind of a research mode also to do forecasting and hind casting and skill assessment for different parts of the world. I started to think about what else we could learn from hind casts. So a hind cast is basically a whole series of past forecasts that you initialize and direct forward. And typically the way you look at those hind casts is the ESP forecasting way where you just look at the skill of forecasts that's been driven by uncertain climate with this assumed perfect initial condition to compare that to observations. One of the things I decided to do was to reverse the approach and then look at what kind of skill or uncertainty you would get, how that would propagate through the forecast. If you instead assumed you had a perfect meteorological forecast and then you combine that with uncertain initial conditions and then being able to look at how uncertainty or conversely forecast signal propagates into the forecast period, you can tell throughout the year what the relative importance of these two sources of predictability are. And then you can sort of combine them in a ratio of uncertainty that gives you a quantitative attribution. So this kind of started to develop a more formal framework for attribution and quantification of hydrologic prediction uncertainty. And one thing to be aware of is these contributions very quite a bit seasonally. If you look at the sort of precipitation inputs in this gray going throughout the season in two different locations, one in Northern California, another one in Northern New Mexico, you can see that not only do the contributions from precipitation inputs to watershed very seasonally, but also kind of the balance of moisture stored in the watershed, where here, SWE is a snowpack, soil moisture is SM. And those sources are the drivers of that initial watershed's predictability. So just to highlight really quickly that this type of analysis can really draw some contrast in how hydrologic predictability works. This is showing again for those two watersheds, this ratio of errors that you would get from these two formulations, ESP and reverse ESP. And one of the things that immediately shows you as a contrast where if you initialize a forecast in October in California, the climate forecasts are a much more important driver than the initial conditions initially, this flips by the time you get to spring, sort of April, May initializing forecasts. And here, the initial condition uncertainty is a much more important driver of future uncertainty. And then the reverse is true in Northern New Mexico. So it just kind of illustrates the complexity of this interplay between these two sources of predictability. People have since gone further, including myself in characterizing these contributions both nationally, this has been demonstrating a concept of a forecast skill elasticity. So the sensitivity of runoff forecast skill to skill in estimating initial conditions or estimating future climate forecasts. And this shows for about 400 or 500 reasons around the US that there are these seasonal variations and how these different sources of predictability impact the runoff prediction. So I don't wanna spend too much time talking about this particular result of one notable feature you can see from this kind of analysis is that say in the West Coast, the initial condition uncertainty is very important as you get from January into spring and summer, not so much the climate forecast predictability. So they're just nice to be aware that this varies quite a bit. There have been subsequent studies of this kind of predictability balance in the US and Europe and Canada and global analyses. Operational forecast centers have gone and done analysis so that they can try to understand where best to put investment in effort to improve a runoff forecast, whether in climate forecasts or better initial conditions. And so with this knowledge that these two sources are important, there are many, many ways that people are currently working to try to harness this predictability that are purely empirical to school in now more recently machine learning ways where we try to assess sources of skill in climate agencies or re-analysis fields for dynamical model combinations or sequences or chains of analysis that use instead of empirical methods for client modeling or watershed modeling to come out with stream flow. Or of course, hybrids of these approaches include elements of empirical techniques as well as model-based information and sort of traditional, the traditional ways are still being used where as I mentioned, you might take a hydrology model, try to forecast and then condition it in climate information. I actually spent a number, I believe one study with one postdoc really delving deeply into how different ways of combining the source of predictability would play out if you're doing watershed-scale forecasting. So using those initial benchmark methods like the ESPR mentioned, which capture mainly initial condition predictability, we may kind of cast a spring runoff forecast which are told worse by forecast and then just start to layer in different ways of also trying to capture meteorological predictability in climate and then eventually to combine it with the initial condition predictability. So one example is just, for harnessing meteorological or climate predictability is just to do kind of empirical regressions, component-based regression on climate states and climate variables and relating those to stream flow but that doesn't necessarily know anything about the initial condition uncertainty. There are various ways of merging the information, one of which is that trace waiting scheme I talked about before. And then there's also just model averaging type techniques like Bayesian model averaging, quantum model averaging, skill-weighted ensemble blending, all kinds of things that you can think of to do. When we went back and looked at this, this is for predicting watershed flows in Northwest of the US, in various watersheds, we could systematically compare how well these kinds of approaches played out like the ones that use only the watershed initial conditions versus the ones that use only climate information versus ones that merge both. And in general, as we go through this, we tended to find that as somewhat as expected, methods that merge the two sources of information like the BMA or QMA or skill-weighted ensembles, these green boxes in this purple box tended to do well throughout a whole forecast period. When you start in October and you make spring run-out forecasts going all the way up to the spring, the only climate methods tended to do well early in the season, but as again, as we saw in those earlier plots, when you start to pick up watershed moisture in the winter and spring, then you really have to also have, that is a main source of information. And by the end of the season, if you're only accounting for watershed information, that can be quite skillful. So one of the things I wanted to mention is that this week in the tutorial, the students took a set of hind casts that I had made for Buckle of Bill Resilor, a kind of small case study to the north of us in Wyoming and they applied a trace weighting or climate conditioning scheme on the ensembles. And it was given the short amount of time that we had, it was fairly ad hoc exploration and experimentation with different weighting schemes. But in general, in the time that you could get marginal improvements in these watershed only ESPs if you conditioned them on climate information, which came from ECMWF, CSM2, NSF2S forecasts as well as climate indices. So not in every year did things work that way, but in this particular year that I'm showing they did. If the green line is the observed street flow, the blue line, which is the weighted ensemble mean is closer to it than the red, unweighted ensemble mean. So this kind of a, I was checking the time, but this kind of approach is being actively used in real world, larger regional case studies. I'm just showing an example on this one slide where given a strong user need for improved forecasts for big reservoir systems, such as Colorado River with Lake Powell need, there's a lot of effort going into trying to translate information being produced on a kind of large synoptic scale by forecast producers down through post-processing and other techniques down to the watershed level and then connected to specific water management info forecast and management decisions. Just to give a quick example, this is work with Sarah Baker, who's the Bureau of Reclamation. We're able to show that you could reduce the error of info forecast for Lake Powell by conditioning them with a sort of analog technique using National Multimolar Ensemble. Climate information, just this bears out when you look at the errors where the yellow error is again, a volume forecast starting in August going through the next year or for the spring runoff. And we can see that by including this climate information is purple and blue errors tend to end up being lower. So that's very helpful. And finally, I just wanted to say that this work in hydrologic prediction is not just local, regional, national, but also global. There are an expanding number of activities to try to bring together the kinds of forecasting that are done in major centers around the world and then local experts to produce global actionable forecast. Again, by combining this knowledge of predictability but also taking information from coupled forecasting approaches and uncoupled forecasting approaches and trying to meld them all together into reusable large scale product or global scale product. So with that, I'll just go to my final thoughts that I think are important when thinking about hydrologic prediction. One, it's really critical to recognize that there are seasonally varying combinations of land surface and climate predictability that plus play that get combined when you need to try to make a hydrologic prediction. And it's really important to understand how they vary and the best ways to combine them. A lot of different strategies for combining these two sources of scale. For now, we don't tend to be able to take land surface hydrologic outputs directly from global climate forecasting models because they haven't been validated or calibrated well enough. And so we have a lot of these kind of uncoupled sequences of analysis to create our forecast. And finally, throw a lot of different means working on this problem from the science angle to the engineering angle and more than we bring them together the better we do. And with that, thanks for your attention and I have time for questions. I'll be happy to answer it. Thank you very much, Amy. Yeah, thank you for this nice offer field that's sort of quite different in its challenges from at least what we have in the atmosphere. Are there any questions? So I have a question. I know this wasn't the topic of the talk, but my understanding is that these hydrological models have a lot of parameters that are being tuned and they are tuned differently for different catchment areas. I was wondering if you could comment on the unification of those to have a parameterized consistent hydrological model? Yeah, that's one of the longstanding challenges in hydrology. And the reality is that to get the best forecast or simulations from a hydrology model locally now, we still need to do that kind of tune parameter estimation in local places, which can be somewhat scientifically unsatisfied. There's a huge effort to try to think about how we can do better regional parameter estimation that is based on less on the data in one particular watershed, but they are across many, many watersheds, large samples. So to be honest, what I've seen is that for the most part that can get you maybe 70% of the way to what your optimal performance would be. And so we really still need to think about how best we bring together the local techniques with these regional techniques to get something that will give us what you might have more in the weather and climate models. On the flip side, I would just say that for the longest time, I think the climate forecast community and weather forecast communities tend to think about this kind of parameter tuning as being kind of engineering-minded, maybe a little bit non-scientific. And maybe have ignored some of the needs to estimate parameters in their own climate models and some of the science that's needed to improve climate model simulation through parameter estimation, similar to what hydrologists do. So I do think maybe there are some learning that can be done from both communities. I think the climate science can learn from the hydrologists and engineers about how they've done this kind of thing. And we can learn from the climate scientist about how to use parameterization changes instead of parameter changes to make improved model simulations. Yeah, thank you. So the parameters are the known unknowns in hydrology where the unknown unknowns in the atmosphere. Anish, go ahead. Thanks, Stuart. Thanks, Andy. A really nice overview of both challenges, but also the global collaborations and efforts on this important societal topic. My question was regarding data assimilation and observations of, especially on the coupled data assimilation side, like other efforts where observations from the hydrological or hydrology side of river runoff or related variables can then be assimilated into impacting atmospheric precipitation, atmospheric variables, and are there groups that are working on coupling the data and data assimilation between the hydrology and the atmospheric observations? Yeah, I would say, data assimilation for the land surface has always been a really difficult problem. And for the most part, just for hydrological prediction, it's not used anywhere near as widely as it could be. It's quite difficult to estimate deep source of predictability like soil moisture, other things like measurements like stream flow or snowpack or just kind of a mix between imperfect satellite information and sometimes very sparse institute point measurements. And so there's a lot of work that can be done in this area. I do think like there's been potential for from satellite products to do global scale assimilation and they do get used in weather and climate models. How much the land surface state influences the scale of atmospheric predictions is also a very active, very research. And I think as we learn more about which states when, where impacts, what's in the atmosphere or circulation in the atmosphere, we'll be able to target our data assimilation efforts more carefully. These are all great questions and great areas of research to pursue. Thanks. So just sort of to understand better, it seems that if we would want to go to a next generation hydrological modeling approach that actually uses the atmospheric information, we would need to have better calibrated input for hydrological relevant variables. Would you see this as the biggest hurdle as opposed to missing observations for the initialization and what would be the next step to bring those two fields closer together and propagate the uncertainty through the whole system? Yeah, I think there, I mean, I think that the land data assimilation and I should mention groups like the GMMAO at NASA are very active in pursuing this, particularly using satellite information. But I do think that the combination of the land surface data assimilation to try to bring models closer to good initial states for forecasting that plus improved parameter estimation and process representation in a couple models and then maybe bringing more hydrologists and people who think about the land surface in a different way into global couple modeling groups and development efforts is going to be really critical. I sit in a group in CBD now that has a lot of people thinking about hydrology, but I've noticed and learned that while being there that a lot of the thinking is on fluxes that I often don't think too much about. It's more about evaporative flux into the atmosphere than about runoff and base flow from the land model. So it's just there are these different perspectives. And once we get really serious about improving this land surface hydrology and its ability to capture variability at different scales, some watershed scale, regional scale, we will struggle to really fully leverage the kind of predictability we can get from the land surface. I do think there's part of the community that and I would say most are coming around to this idea that in the future, if we can take a couple of systems model and get good quality hydrology from it, that would be a preferable solution to having all these sort of uncoupled chains of analysis. And but I think that goal or that drill is still out ahead of us. And I don't think it's unachievable, but I think it takes careful thinking about how we get there. Thank you very much. There are no other questions. I would like to thank Andy and introduce our next speaker. Our next...