 Today, our first speaker is Dr. Andy Wood, who's a scientist at the NCAR Research Applications Lab here in Boulder, Colorado. He's one of the leading experts at NCAR in the areas of stream flow forecasting, hydrological modeling, and applications and monitoring and prediction for floods, droughts. His work also on climate change, downscaling and impact assessment, water resources management and planning and seasonal climate and hydrologic forecasting. Andy worked in many different sectors and he has worked for a couple of years in the US Army Corps of Engineers Institute for Water Resources focusing on assessment of wetland storage for flood control. Andy also worked at the University of Washington Department of Civil Engineering as a research professor, focusing on research to improve real time hydrologic monitoring and prediction systems, after which Andy was a senior and lead scientist of a private firm in Seattle three tier focusing on forecasting and assessment of hydropower, solar and wind energy. Andy spent three years as a development and operations hydrologist with the NOAA National Weather Service River Forecast Center in Salt Lake City. Andy has served as a chair and co-chair of many national and international committees, for instance he has chaired the hydrology committee of the American Meteorological Society. Andy is currently an editor of the AMS Journal of Hydro Meteorology, and he's also co-leader of the International Hydrologic Ensemble Prediction Experiment, which seeks to advance the application of ensemble forecasting for water management. And he was formerly the co-leader of the first NOAA map drought task force. Thank you Andy for accepting our invite. Looking forward to your lecture. Great, thanks. I will share my screen and try to find my presentation. Okay today, I'm going to talk about the influence really of the land surface on subseasonal to seasonal predictions. I think there's an important aspect of the land surface that and its influence and sort of moderation of seasonal prediction skill and the importance of climate forecasts that maybe not widely appreciated within the climate forecasting and modeling community, particularly global forecasting modeling community. And yet is has been a concern and a challenge for hydrologists and water managers seeking to make use of SOS forecast and hydrology for a really long time. So I showed a few of these slides when we started up the workshop on last Monday, so some will be familiar. And the first couple slides are going to be about this role of the land surface versus subseasonal climate predictions. So in particular, when we think about hydrology, and particularly the variable that we care about for water management, which is often stream flow, probably predominantly stream flow. We have to recognize that that stream flow is generated by processes that are largely in the subsurface or stored on the surface in the form of snow. So in some meteorological events, it's basically raining was such a high intensity that the mineral meteorological drivers of stream flow are dominant and rainfall is running off directly in the streams without really delaying within the subsurface but most of the time, it's kind of a balance stream flow the generation of stream flows kind of balance between hydrologic and meteorological influences so when we look at as to as predictability and hydrology. There are basically two main sources of predictability one being high level called hydrologically hydrological or watershed predictability. Basically the evolution of snow through snow melt and so moisture through drainage into stream channels that are routed to areas we care about. And then the other source of predictability is how well we can estimate the drivers of the hydrologic variability so future climate forecast future weather forecast. If those have skilled those can be filtered through the land surface to provide stream flow forecast skill. I guess that there are always these two sources of predictability that are potentially active. We find that we have to think a little bit differently about how to make how we want to go about making hydrologic sub seasonal predictions. And just to give a bit of background for those who are more on the climate weather side, or perhaps data science side. As I mentioned hydrologic forecast predictability derived from two major sources. For about the last 40 years. The main dynamical based approach for forecasting stream flow or forecasting hydrology at seasonal scales has been called something called ESP, which was originally extended stream flow prediction later became ensemble stream flow prediction. We tracked down the origins of the practice which began at the California Department of Water Resources but with the National Weather Service looking over their shoulders from the California Nevada River forecast center. So in the mid 70s. We started a practice of basically initializing a watershed model, driving it with observed meteorology up to the current date, in which case, the watershed model would capture this watershed predictability in the form of snow. And then driving it forward with historical sequences of climate so they didn't at the time they there certainly was a lot of research into teleconnections and some of that was funded by California as early as the late 50s. We were basically trying to establish connections between sea source temperatures and and land climate. But at the time there were not many developed operational climate forecasts so historical sequences were used to kind of evolve the watershed conditions forward into the future. And so we were recognizing the role of Lancers physics and transforming meteorology into future outcomes for stream flow, and this practice became known as ESP. And so we can use all over the world as a dominant form of seasonal or sub seasonal this seasonal hydrologic prediction in the, as we get into the last decade or so, the rise of greater number of climate model base forecast and availability of them and availability of field assessments that show that they have benefit to start to change this practice a little bit, but this is still a dominant form of forecasting. The most common ways that you would include proof hydrologic forecasting focus on basically both the sources of predictability either can improve the watershed models or the watershed observations, or methods that are used to to estimate those conditions including data assimilation, or you can prove improve your knowledge about future weather and climate, which would help shrink the uncertainty that you have going forward from this initial condition. So I'm going to talk a little bit about some work we did to screens frozen. So the work we did to start to understand the interplay between these two sources of predictability. Back in around 2004 I had been making a lot of hind cast ensemble line cast and presenting them at conferences. One morning, I confess I was getting a little bit tired of just presenting the same material so I thought about kind of reversing the concept of the ESP forecast, and thinking about so the ESP forecast basically combines us a well known initial condition with uncertainty in a climate forecast in the future and I thought well what if we look at the role of uncertainty in the initial condition as a contrast with a perfect climate forecast in the future and I call this reverse ESP. I realized that if you contrast the skill that you get from these two ways of making a forecast, and of course this reverse ESP is not a realistic way of making a forecast because you would, you don't have in real world the real world that much uncertainty in the initial conditions, then you can start to understand something about the kind of persistence of forecast signal or uncertainty and either one of these components as you go out in longer forecast lead times. And the reason that this sort of ESP forecast works, or that at times this reverse ESP forecast gives you some insights, is that in the hydrology we have this seasonal cycle of storage components. So in the in these two graphs over here I'm showing basically various major sources of moisture such as sweet which is so equivalent or so moisture, and how they vary throughout the calendar year, and then, and give rise to run off this RO field at a certain time year so in a place like the Rio Grande or Lobatos which is a couple of snowmelt driven system, you have a rise of snow in the winter months that melts in the spring and then it drives a rise and so moisture and the rise of run off, which becomes stream flow is a very seasonal phenomenon so if you can imagine that if you're looking at predictability in March or April when these watershed storages and moisture are very high, these are going to do more to drive the run off and stream flow signal than precipitation which is this P sort of shown as a seasonal cycle here. And that's completely different in other locations such as in California where you have a lot of winter precipitation. There's not as much snow in this particular location. But then that winter precipitation, coupled with the dry summer leads to a lot of predictability going out starting in March and April because you don't have a lot of uncertainty and neurological drivers come forward. So these dynamics are what we tried to pick up in this this error attribution framework. And so, for instance, if you compare the error in the forecast that different lead times and this is for different forecast made a different time to the year October, January, April and July, each line is sort of starting where the forecast starts. The ratio of the error in a from a dip from a from one of the sources so blue is the initial condition source, red is the forcing source, or a climate forecast source, you can see that their influence on the overall uncertainty in the forecast greatly throughout the season so in the beginning of the year for instance, climate forecasts are much more important than initial conditions this is the beginning of the water year. That's because you're going into the rainy season, and the initial conditions are likely a dry small moisture and so the hydrologic variability is going to be determined by the weather you've ever a very rainy winter or slightly raining winter. Whereas down in the Rio Grande in Southern Colorado. You start the year with initial conditions being much more important than climate forecast so the basic point of this is to show this error attribution for framework can tell you a lot about how important seasonal climate forecasts are for a spin flow forecast or runoff forecast compared to the land surface, and this is useful because that can help tell you something about if you look at it over the whole country. The elasticity of runoff forecast skill compared to climate forecast skill or initial condition forecast skills so this is on the left flow forecast. This is related to climate forecast skill in this on the right flow forecast skill related to initial condition skill. And basically, easy elasticity sort of show you for instance that if you're initializing on October one in the east coast. The climate forecast skill is very important and that's actually true most of the time, except as you get into summer. And then if you're initializing on October one, this client this initial condition skill in the in the high mountain western us is very important and so management agency might think about the forecast are important to you and and what kinds of techniques for improving them or new data or new development work is going to be most beneficial. So, in any case, this kind of attribution framework is now being widely used in higher logic predictability studies us your Canada. There have been a number of papers that follow the same ESP reverse ESP way of attributing forecast skill to different drivers. This is just an example of a global analysis showing the same sort of ratio concept being used to separate places where initial conditions versus thriving climate forecast influence skill the most strongly. This is a paper by Shredh Shipman who used to be a teacher in the mind. When I was doing this predictability work initially. So, with those thoughts in mind about where predictability comes from for hydrology. It's worth recognizing, or maybe it gives us a way to understand why there are actually many ways that people around the world are making hydrologic s to s forecast. One of the ways is just to use empirical techniques to basically have a statistical or empirical or could be nowadays machine learning model, where you relate broad scale climate indices to stream flow and actually probably shouldn't have run off here but sometimes run off is also used as part of this. Without any sort of land modeling involved. I could just show a quick example of why that works from a long time ago. So 1999 or so. And then the 90s when people were exploring the impact of teleconnection such as Pacific Decadal Oscillation or El Nino and so on land North American climate and continental, I guess, climate for many comments. And how that influence stream flow and you could find it just knowing the PDO or the, the ENSO state could tell you something about your likelihood of getting above average or below average stream flow. And this is a basin and Pacific Northwest paper by my colleague Alan Hamlet. So it was very easy to get some sort of broad predictability out of just paying attention to these indices and then statistically relating them through composites or re sampling their regression models to the outcome for stream flow. More recently, there have been a move toward dynamical forecast and there's quite a large community that would like to basically be able to take dynamical stream flow forecast directly from global climate models or global NWP models, which would be very convenient. You don't then have to run any kind of separate hydrology model or even have separate hydrology staffers, you just, you know, that are doing their own thing with their own model and have their own needs and budget. You can just run your NWP system and pull out runoff and then usually have to route it through some sort of channel, stream flow channel routing model to produce a stream flow. This will be an example of a fully dynamical approach to generating stream flow, and I'll show an example of this later. Another probably more common dynamical way in fact this is probably the most common dynamical strategy that's being pushed, particularly in Europe where you have ECMWF being quite a dominant and influential source of skillful climate forecast would be to take climate fields from the coupled GCM forecast and then you downscale them and you use weather generation techniques to generate local neurological fields to drive a hydrology model to produce runoff to run through a channel routing model to produce stream flow. So this is, you know, in this case there could be some sort of statistical empirical components within the downscaled but in general it's based on dynamical model, running dynamical models, it's quite computationally expensive. There are certainly more hybrid approaches that combine elements of both of these. You know, sometimes you're using these reanalysis fields or indices or GCM outputs and then have some statistical model relating it to stream flow. You'd also have this kind of dynamical approach and then the two are generating separate stream flow approaches that can be merged through techniques like Bayesian model averaging to produce kind of posterior predictions for stream flow that harness the strengths of both empirical and dynamical. There's also traditional kind of incremental approaches that I would say build off the standard ESP approach I talked about before where you use hydrology and lamma plus routing to produce stream flow. But then you go in and you condition that stream flow ensemble forecast that you get using climate information that could come from a number of sources that could come from GCM forecast climate indices reanalysis fields that are identified as being predictive for your region. And then you have a sort of post process stream flow ensemble. That's something we're looking at right now in this tutorial with the hydro group. In any case, there are many people trying to kind of slice this, this pie in a different way. And by and large, the, these don't have the same levels of success and I'm going to get into that a little bit about why in general, I guess I would say the mob. The approaches that use hydrology and land models tend to be more successful. I should say that one other thing that's been, we've been doing in the hydrology area is trying to make climate forecast more accessible to groups that are focused on watersheds and water management and hydrology and hydrology modeling. So in this kind of a project, we do with a variety of different post processing tech techniques, and some spatial translation techniques to basically process available forecast into watershed scale predictance, which are more familiar to water managers and water management groups. They have, they tend to have climatologies that are more recognizable versus the kind of course scale climate products that are not not more not this securely, not tailored to sectors like this. So this is an example of a system we stood up about four or five years ago to use an enemy forecast and see a press forecast. And just to go quickly through this, I don't want to make this a major part of the presentation, but in this kind of a system, you can look at the skill of forecast from different sources at lead times and for sort of prediction areas that make sense to the water managers and hydrologists. And you can even go in and through post processing, find ways to enhance that skill it's very difficult certainly for precipitation like in the week three forward period. In this case we do some conditioning of the CFS version to climate forecast using other fields from the CFS version to predictions such as SSDs. And then when we combine these in a kind of component based progression, we can definitely improve temperature forecast, but it's more challenging for precipitation forecast. So in terms of approaches we can, again at this watershed scale, actually improve forecast in places and at times but certainly not everywhere. This was kind of a, you know, it was a limited scope study but the point was to show that even at this watershed scale some of this post processing could be helpful in enhancing skill and hopefully some of the other groups that are working in this with machine learning in this ASP colloquium will be finding similar outcomes. So that's so sorry all these areas that are yellow and red and orange are places where we improve skills for this post processing. Getting back to an example, another example of this sort of traditional empirical development of S2S hydrology forecast. I'll just show an example of some work that we did recently with the Bureau of Reclamation, Upper Colorado River Basin, which is here. This is an example of a typical product from our national centers for an NME forecast of precipitation. It's very core scale, you know, boundaries where it's pretty hard to pick out the actual watershed, which looks like this. Again, talking about hydrologists and water managers, we tend to think in terms of these basins and these drainage areas that go into major reservoirs. And as we're, it's especially important right now that we're hitting this really critical drought, I don't know if people have been paying attention to the news, but these big reservoirs for the system, like Powell, like Mead or now, at historical lows. In any case, here we took again this traditional approach where we generated ensemble predictions using an ESP forecast approach. And then once we had those ensemble predictions, we looked at the current forecast and looked at analogs within the meteorological drivers of our ensemble predictions and then gate those analogs that matched our current prediction higher weights so that we could then impose the climate forecast signal onto this unweighted ensemble forecast driven by historical climate. And we did it in different ways. We looked at climate information over this entire river basin, which you can see spans multiple states. We also looked at it in a more granular fashion, breaking out certain areas where you might anticipate you have a gradient of forecast outcomes, you know, you could be forecasting it's going to be drier in the San Juan, river basin to the south and wetter in the green river basin to the north. So there's maybe some argument that you should be trying to harness those gradients when you do this kind of thing. And as I showed before, I largely found that imposing climate information in this way, and this is showing the scores for CRPS. These are forecasts of runoff volumes starting in summer and going out to June, where the predictor is basically the runoff volume, I think it's from April through July. I do find that these methods called KNN these analog methods, they use climate information, do by large reduce the air in the forecast so that's a successful demonstration of forecast, using forecast information from climate models like NMME to benefit hydrologic predictions still. I want to switch now to talking a bit about global systems and in what's possible and what's being done on the global level to do hydrologic predictions. There are a number of centers, ECMWF, Swedish Met and Hydro Institute, that's in Sweden, Deltares and Holland, and other groups around the world they're doing. They're starting to pay attention to hydrologic outputs from their global prediction systems. The most credible of these, as I mentioned before, involve uncoupled land surface simulations are driven by climate outcomes from global hydrology models. So, example of this is the SMHI Worldwide Hype Model, which is forced by ECMWF, renown season forecasts. These fully coupled systems in which the land model is being developed that is basically putting out rough, it's being routed into, to strengthen locations and assess. I think the goal there in this development is that if we start to pay attention to the psychology. We will eventually develop and improve a source of predictability in the land surface that could benefit American weather prediction. So it's kind of an indirect way of eventually getting better weather and climate forecasts as well as hydrology forecasts. So from this batch of efforts, I'd say the least credible outcomes unfortunately are still the ones from these coupled forecast systems where the land models have never really been specifically developed for hydrologic prediction. And here I'm just going to show some examples of these types of efforts, which often have quite nice websites where you can go and look at predictions. This one is from SMHI, showing a sort of monthly mean river flow and SMHI's land models are implemented on a catchment basis, so they're not gridded. You can often see different catchments in the spatial patterns. The easy end of the couple system approach is basically to run their IFS, their integrated forecasting scheme. And then take the runoff and route it through a model called ListFlood to produce global forecasts, global strength forecasts and do some post processing to turn it into regular alerts and warnings and other kinds of usable information products. We have a nice talk from Rebecca Emerton last week who helped develop the seasonal side of this global flood awareness system go fast. Before it had been used for medium range prediction. Here's just an example screenshot from this product which is on the web that you can go and access it and query it. I think you have to make a free registration but aside from that the products are free. But definitely there is this attempt to do a kind of global sub seasonal to seasonal hydrologic prediction that's being picked up by agencies like ECMWF. Here's another example. If you drill into this global fast system and you click on a reporting point where they're highlighting a forecast perhaps because something is happening there and seeing the kind of plumes for river flow going out. They're again being generated by this couple land scheme and then various alert oriented products where they show risk in different time periods across the sequence of forecasts. And then nice thing about what's being done, I would say this work is which started back in around 2012 to put this together is reached the point where skill assessments are being published and you can see that there's both good and bad in these skill assessments certainly there are a lot of locations where there is no skill if you look at a skill score like the rock. And then there are places where there's much higher skill. This is kind of to be expected. Another nice thing about these efforts is that they're starting to go beyond just forecasting to generating all the other kinds of products that are needed to understand the forecast so pre forecast and re analysis that you can go back and used to do skill assessments and the systems are becoming quite, I guess, elaborate. And one notable thing to say about this glow fast system is that, although it started with a couple landscape, they have now moved to spinning up an uncoupled global landscape because they recognize the need to calibrate and improve the land component of the forecast there needs to be more control over how the land surfaces are responding so that they can improve the quality of the forecast so that's in a way it's kind of a recognition that we're not quite there on the use of runoff directly from couple forecasting models such as the ones in our s2s database. And there's a lot of hope that we will get there through model development I think probably the next speaker Paul Germer will talk a bit about what's being to improve land surface performance and understand the predictability there so Sorry to interrupt a couple of moments. Okay. I'll just, you know, I'll just say that we're trying to use these in in a new system that's being set up by the world neurological organization. This is a combination of coupled and uncoupled approaches. So, I will just say one thing here. It's, this is a slightly different angle taken from the long range projection hydrologic projections there's a increasing trend to start to use hydrology from not just s2s but Decadal and multi Decadal or scale projections. I just want to point out, perhaps why we're in this situation where our land models don't automatically give us good results when we take the runoff from them. Certainly there are a lot of reasons related to scale and the ability to resolve physical processes and these models are running very large scale so don't necessarily resolve the features of, of watersheds including their But in addition, the way that we evaluate land models in couple or system models or GCMs is as just a small fraction for hydrology amidst all the different other factors that we're trying to evaluate such as atmosphere radiation energy cycle carbon cycle and other things so that in the end we may only be looking at rough as just one small piece of the big picture and as a result is going to be very hard to tailor it and improve it in a way that makes makes us sure that we'll get reliable results so With that, I won't spend too much time but just give some final thoughts. The main takeaways I like people to have is that there's multiple paths for creating s2s hydrologic predictions. It's really critically important to recognize the role of the land surface on forecast skill that just because a global model or a couple ESM has runoff in hydrologic fields. They may not be high quality. And until we improve this land surface hydrology are most usable forecast for coming from these kind of uncoupled calibrated land models driven by client information. So with that, I'll stop. Hopefully this can help us bring our communities together so that the hydrologists and water manager type people like myself are more in communication with climate developers like many of those in this symposium. Thanks. Thank you Andy. Thanks for very comprehensive talk.