[Speaker: Jacqui Lloyd (DEECA)]
Hi everybody. There we go. Hi everybody starting to file in. Welcome to our webinar. We're just going to wait a minute or two to let a few more people in, but thank you for joining us. All right, I reckon. I think we've got quorum. I reckon we just go for it now.
So welcome everyone. Thank you for joining us today. This is the VicWaCI Webinar series hosted by the Hydrology, Climate and Energy team at the Department of Energy, Environment and Climate Action (DEECA). This webinar series showcases research from the VicWaCI research program. And my name is Jacqui Lloyd - hello!
I'd like to begin by acknowledging the Traditional Owners of the land on which we're meeting today. For me, that's the Wurundjeri People of the Kulin Nation. I'd like to pay my respects to their Elders, past and present, and extend that acknowledgement to those across other parts of Victoria or Australia where you in the audience may be joining us today.
You may have picked up at the start there our team has had a slight name change.
We are the Hydrology, Climate and Energy team now. However, we continue to manage the Victorian Water and Climate Initiative (VicWaCI) and our contact e-mail is the same: HCS.Team@deeca.vic.gov.au. That hasn't changed.
As with previous webinars we'll be recording this one today and by default your camera and mic is off. But we encourage you to use the Q&A function of Teams and we'll have time for questions at the end. I think we'll drop the link now in the Q&A section actually. Because our big news is that we have the recordings from the previous webinars available on our website now.
And just before I introduce our speaker Keir, I just want to mention that we're expecting to soon be releasing the report summarising the VicWaCI work from the past few years and we're pretty excited to share that with you. Because it summarises all the findings from the second phase of VicWaCI and there's heaps of great work in there. And when it's released, we'll share that through our e-mail group.
OK, so on to today's presentation. I'd like to welcome Dr Keir Fowler, who will be speaking on the topic of Drought-induced shifts in the relationship between rainfall and runoff: recent improvements in understanding and modelling. This research was funded through the Australian Research Council Linkage Program with DEECA and Melbourne Water as industry partners, the University of Melbourne and Monash University as research partners and in-kind support from CSIRO. The research is not directly a part of VicWaCI, but it's a part of the research that is supported to better understand water availability in Victoria and potential impacts of a changing climate.
So we're very happy to have Keir here today to talk about it.
Keir is a Senior Lecturer at the University of Melbourne in environmental hydrology and water resources. He has a broad experience across both academia and also the water industry as a consulting engineer. His research aims to improve climate change risk estimation to help water industry partners to plan for the future. His areas of focus include understanding how watersheds physically respond to past changes in climate, and improved rainfall-runoff modelling in the face of a changing climate. And also developing better ways to make decisions about water systems given we're uncertain about future climate change.
So here we are. I'd like to now pass to Keir to begin, and thanks very much for joining us today Keir.
[Speaker: Dr. Keirnan Fowler (University of Melbourne)]
Thanks, Jacqui. And Jacqui, you can see my slides OK?
[Speaker: Jacqui Lloyd (DEECA)]I can. Beautiful.
[Speaker: Dr. Keirnan Fowler (University of Melbourne)]
Excellent. Thank you very much for the introduction.
So I'd like to begin by acknowledging this was a real team effort. And particularly acknowledging two people: Hansini Gardiya Weligamage and Luca Trotter, who did their PhDs as part of this project. And their supervisors included myself, Murray Peel from the University of Melbourne, Tim Peterson from Monash University and Margarita Saft, who is now in the Technical University of Berlin in Germany.
OK, so the way that I've structured this presentation is that there's some initial slides which kind of describe where we were up to in our understanding prior to starting this project. And they'll also serve as a little bit of an introduction to the broad concepts. I'm going to then go through a number of different papers that were written as part of this project and the main lessons learnt.
OK. Sorry, my computer always struggles with the first one, OK. So this is a classic example of the type of hydrological behaviour that we're trying to explain and also model. So this is the Campaspe River which is north of Melbourne. And each of these blue columns represents a year of rainfall and then each of these blue crosses represents a year of streamflow from this catchment up until the start of the Millennium Drought in 1997.
And so you can see that up until that point in time, this catchment had been delivering a fair portion of rainfall as streamflow around about 20% on average. And then we hit the Millennium Drought in 1997 and we see around about a 10-15% reduction in rainfall on average. But the reduction in streamflow seems to be well out of proportion with that reduction in rainfall. So much so that it's well below what we would have expected even given the reduction in rainfall. And you can see that clearly if we plot each of these years on a scatter plot. We have rainfall on the X-axis and streamflow on the Y-axis, and so if we kind of focus in on say, all the years that had about say 800 millimetres worth of rainfall you can see that the years that with that rainfall during the drought delivered a lot less streamflow for the same rainfall compared to years that happened before the drought.
And so I guess before this happened, what we naively would have thought was that there was kind of a single relationship between rainfall and runoff and that during wetter years, we go up that relationship and during dry years, we go down that relationship but it, but it's the same relationship throughout. Whereas what these kind of results are suggesting is that the relationship itself is shifting to give us less streamflow for a given rainfall.
And this was not an isolated case. Here's three more cases, most of them for a bit further north. And here's a map that's based on Margarita Saft's PhD work, which was published 10 years ago now. So this is based on information nearly up until the end of the Millennium Drought in 2010. And so Margarita developed a statistical test to be able to determine whether there was a statistically significant shift in the relationship. And all of these red dots are those where there is a statistically significant shift. So you can see that it's general across this whole area of south-east Australia, but also it's not universal, so there's plenty of catchments where it's not detecting that shift.
The other thing that we had just sort of started to understand prior to the start of this project was the idea of the persistence of this behaviour. So this is a plot from a study led by Tim Peterson, and Tim used data up to 2017 so it included
seven years after the end of the drought. And I guess given that the drought ended with record floods in 2010 in many catchments, we would have expected, optimistically, that that would have kind of knocked the behaviour of catchments back to their old rate. It kind of would have reset back to the old rainfall-runoff
relationship. But the thing to focus on in this plot is the red dots and the red line, which described the years after the drought. And you can see that that those years in that line is a lot closer to the behaviour during the drought than it is to the relationship between rainfall and runoff from before the drought started.
And so the meaning is that in catchments that have shifted and not recovered, a year of average rainfall gives less streamflow today than it did prior to the drought.
Just give me a moment. OK, so I have a slide here introducing the project, but oh sorry, I forgot that this slide was here. So this slide shows the sequence of recovery across the state of Victoria. So you can see that in the east of Victoria, the catchments never shifted. They maintained a stable rainfall-runoff relationship throughout the drought, so that's in the Victorian Highlands, whereas the majority of the rest of Victoria did show shifting during the drought and then by 7 years after the drought many of those catchments had recovered, but still around about 1/3 of catchments had not.
OK, so now on to that slide. But I think, Jacqui, you covered this pretty well. You covered the funding partners and the research partners and the project went for 3.5 years and it was completed last year.
OK, so one of the challenges in trying to synthesise across the whole project was just that there were really a lot of papers and a lot of different insights that came from the research. I'm going to do my very best to sort of summarise. So what we're going to do is essentially we're going to fill out this table together. So there's some big questions that we're getting asked on the left hand side here and then for each of those, there's at least one paper that was written in response to that question. And I'll kind of try to summarise the main point of that paper in that third column.
OK, so the first big question is: how were other aspects of the water balance behaving? So what I mean by other aspects is we've looked at rainfall, we've looked at streamflow, but there's other aspects of the water balance such as actual evapotranspiration. And also, if you think about the water balance, it's usually water in minus water out is equal to change in storage, right? So there's this change in storage component there as well.
And so Hansini in her first PhD paper looked at these other aspects of the water balance. And that's the paper there. And a key figure from that paper is this one, I've kind of pulled it out. It was for several different catchments, but this is just the results for one catchment. It's in the same basin as before: the Campaspe basin a bit further downstream, and this diagram is showing essentially a water balance. So this is for the years prior to the drought. You've got your precipitation that's giving you the total water available and then how is that split between actual evapotranspiration in orange and streamflow in green. And obviously you can see that most of the water, even prior to the drought, was getting used for actual evapotranspiration (AET).
OK, so then when we moved to the drought period, we've got that about 15% reduction in rainfall. And immediately you can see that in real terms the orange, the AET, has not changed very much. So the average millimetres per year equivalent amount of water that was being used by AET is around about the same during the drought as it was prior to the drought, despite the fact that the rainfall has reduced.
OK, so you're probably asking, OK, what about change in storage? This is a little bit more difficult to quantify. We know that there's been long-term changes in water stored in the landscape because we have groundwater bores and we can see declines in those groundwater bores. But it's really difficult to get volumetric information out of that. Luckily, there is a satellite called GRACE, where the G in GRACE stands for gravity and believe it or not, by following minute changes in gravity over time, we can relate that to the amount of water that's stored in the landscape. When we do that, when we look at this study area regionally across say western Victoria, we get that the amount of storage change that was occurring over central and western Victoria over the drought was around about 25 millimetres per year equivalent depth.
So if we add that in, it doesn't look like much, watch closely and there's a tiny little black bar there, OK? But if we express these as changes between these two periods, it comes out a little bit more clearly. So the rainfall reduced by 150 millimetres. The actual evapotranspiration really didn't change very much. The storage was changing by 25 millimetres, which is essentially lending extra water into this equation essentially fuelling the AET, and then really with those two, orange and black, accounted for really the only place that this rainfall change can be apportioned to is the streamflow, OK.
So the take home message there was that the AET, actual evapotranspiration, remained constant-ish during the drought compared to prior to the drought
and that the storage change, while small, accounted for around about 10%
the precipitation change.
OK, so the mental picture that we're building up here then is that
the actual evapotranspiration, so plant water use is a big component of that, tends to get the first cut of whatever rainfall is available, and then the streamflow kind of gets whatever is leftover. And this is actually consistent with work that had been done prior to this project that was talking about the role of vegetation in driving water scarcity during droughts, particularly in the context of greening. So greening, meaning that we've got more carbon dioxide in the atmosphere and carbon dioxide can act to fertilise plants and make them grow more vigorously. So the hypothesis was that that was exacerbating water shortages in terms of streamflow.
What hadn't been done until this project is that no one had looked at shifts in vegetation behaviour concurrently and in the same locations as shifts in streamflow behaviour. So that's what Hansini did for the second paper in her PhD, and this was the main plot from that paper. So on the Y-axis, we've got how much did the catchment shift their rainfall-runoff relationship? So the lower down the catchment appears, the more it was shifted - so the more water was missing essentially hence the negative values. And on the X-axis, I realise you can't see the full label, but I'm going to replace it here. So what we're looking at is has the vegetation behaviour shifted? So relative to what we would expect given the weather conditions during the drought, OK, and the RS stands for remotely sensed, we're using satellite data to get a regional or at least a catchment scale view of what the vegetation is doing.
And so if it's shifting more, it's going to be on the right hand side. And if the hypothesis is true that the vegetation is driving the hydrological shifts, then we should be seeing a correlation between shifts in, sorry, so a catchment which is showing more shift in remotely sensed vegetation behaviour should also be appearing lower down, so in other words, it should be the same catchments that are shifting hydrologically.
And by large that is what we saw. So, we did see a correlation between the two, but there is a fair bit of scattering in this relationship as well, so it's clearly not the only driver in the equation.
So the take home message then was that the ET was higher than expected given the weather conditions during the drought and that the places where that occurred seem to be correlated with the places where the streamflow was also shifting.
OK, so in terms of broadening our view of this problem, if we think about the introductory slides, they were all dealing with annual totals of streamflow, they were not, so we lost the information, for example about hydrograph shape, and so we're particularly interested in baseflow, because baseflow is traditionally associated with changes in groundwater and we think that in addition to vegetation, groundwater may be playing a role here.
And so Luca in the second paper of his PhD, looked at the changes specifically in the recession of the hydrograph. So usually, we have a flood peak which is driven by the rainfall and then the hydrograph slowly recedes, which is typically driven by the slower sort of groundwater parts of the of the catchment.
And so what Luca is showing here in this plot is he's pulling out, so it says days two to four up there so we've had our rainfall event, it's not the day after that, it's the day after that again, and then it's for three days. And what he's looking at is he's quantifying the recession behaviour by what's called a recession constant. And he's looking at differences in values in that recession constant between the pre-drought period, so prior to '97 and the 2005-2015 period, OK. And within that, there's going to be lots of different flow events. Some of them will be really big and so more streamflow during the recession, and some of them will be smaller, but still a recession occurring. So he split that off in the different categories of the X-axis and then finally you'll see different colours of boxplot here, representing whether or not these catchments showed a shift in the relationship between rainfall and runoff on an annual basis.
And so what we're seeing here is that there is indeed a big difference. So there's a big difference between the pre-drought period and the drought period in terms of recession behaviour. So the catchments are receding more quickly and those that are doing so are the same ones that were identified in Tim and Margarita's work earlier.
So the take home message is that there were steeper recessions during and after the drought, and that this then suggests disconnection between the groundwater and surface water components of catchments.
OK, so if we try to bring this all together, concurrently with all this research happening in Hansini and Luca's PhDs, we also decided to get together a really broad range of researchers, and also people from different government, such as DEECA, such as Bureau of Meteorology. And you can see the big list here and there were a number of universities involved, CSIRO as well. And even more importantly, interdisciplinary, so there are a lot of vegetation people in the room, surface water hydrologists, groundwater hydrologists, remote-sensing experts, and the purpose of the workshop was essentially to present them with this evidence. So what has been happening hydrologically, which is essentially the first part of this presentation and then ask them to sort of discuss and explain their sort of perceptual model, their mental model, for what could explain this behaviour, which is obviously informed by each person's kind of background and discipline.
So we then sort of went through a brainstorming, no-idea-is-too-crazy kind of phase. And then we shortlisted all of them according to whether they are consistent with available evidence. And this slide shows the ones that passed that test. So I'm just going to go through them 1 by 1.
So firstly, the first thing we noted was that the Millennium Drought was the first long drought for which we had widespread streamflow monitoring. So we know of previous long droughts, such as the World War Two drought, but the thing is we weren't measuring streamflow in many locations then. So it can't, the World War Two drought, can't really inform our expectation for what is normal hydrological behaviour during a multi-year drought. That was the first thing.
The second thing was that the evaporative demand, in spring, was higher than in previous droughts. So springtime is a really crucial time for our catchments because it's when they're at their wettest. And so there's essentially the most water to lose if the evaporative demand is higher.
Thirdly is well, I've already talked about this, so changes in vegetation as a hypothesis. There's the greening aspect of that there's the fact that our catchments were recovering from salinisation that had occurred in the '80s and '90s, and there's a bunch of other biophysical adaptations that I'm not going to go into in detail today.
OK, so then we move from there into the groundwater part of the system. Obviously, whenever you have lower rainfall during drought, you're going to get less recharge.
The idea here though, is to explain what's different about this drought compared to previous droughts. And so we talked a lot about the role of different groundwater systems, whether that's local groundwater systems, through kind of fractured bedrock, or whether that's regional brown water systems, both of which can be essentially removing groundwater from the landscape. Once then you've got that kind of multi-annual decline, you've got a thicker vadose zone which then changes the recharge dynamics, and so you sort of get this snowball effect with then it feeding back into the behaviour.
OK, so now you've got groundwater that's too deep to then be interacting with the surface on hill slopes, and in the valleys, you've got less interaction, so you have a contraction in the length of the stream network. So essentially there's a smaller proportion of the landscape that's connecting and interacting with groundwater.
Moving downstream then. In the alluvial aquifers, whereas before the drought, the groundwater levels were quite high, so the stream would be gaining water from the alluvial aquifer, it's now losing that water.
And lastly, we acknowledge that there were more farm dams in the landscape than there were in previous droughts.
OK, so there's a really wide range of hypotheses that were proposed here. This was before Hansini published her kind of correlation between streamflow and shifts in vegetation work so that the conclusion at the time was that groundwater was seen as the most likely cause. I would say, like my personal idea would be that groundwater and vegetation together are seen as the most likely cause.
OK, so that's the end of talking about processes. I'm now going to talk about models.
And specifically, rainfall-runoff models, which are very important in much of the planning in the water sector that we do. So before this project started we already, as part of Margarita Saft's PhD, she had asked the question: OK, if we look at the catchments where we see shifts in rainfall-runoff relationship, are they the same catchments where our models tend to behave really poorly? Where they simulate really poorly, particularly during this Millennium Drought. And the answer was unequivocally yes.
What we didn't know, though, is that, we hadn't yet delved into which aspect of the flow regime was getting simulated poorly, and there's so many different aspects that you can look at.
So that's what Luca looked at for the first paper in his PhD. And this is one of the key figures from that. So there's a little bit to unpack here. This is for one model and what Luca's done is he's tested this model across 150 different catchments in Victoria. And what he's done is he's compared the performance prior to the drought with MD is Millennium Drought and then post-MD is after the drought. And if you're seeing more red it means that the performance of the models have degraded in that later period compared to the pre-drought period, OK. Now the KGE, if you know what the Nash-Sutcliffe Efficiency is, NSE, it's very similar to that. So it's kind of overall model performance. We sort of already knew that that was poor, as I just said, but the subsequent 3 rows here unpack 3 different aspects of model performance and there were many others that Luca looked at as well.
The first one, Q*, relates to overall bias. So is it overestimating or underestimating the water? The second one is the CV, so the coefficient of variation, so are the simulations over variable or not variable enough. And the third one is the BFI, which is the baseflow index and relates to the shape of the hydrograph.
And so the key takeaway from this plot is the fact that the BFI and the CV are not showing much degradation during the drought relative to prior to the drought. Whereas where it's all concentrated is that Q* - so that overall model bias.
So then the take home message here. Oh, sorry, just noting that this was done for
five different rainfall-runoff models, which is repeated here, and the story was more or less the same. So then the take home message then, is that the model struggled with the bias, specifically with too much water during the drought, rather than errors in hydrograph shape.
OK, but in assessing models, we don't only have to focus on streamflow, they also simulate other variables. So actual evapotranspiration, you can also look at the total amount of water that's stored in the model because we've already been talking about the importance of that. And so the next two studies looked at each of those in turn.
These kind of aspects of model simulations are really important, particularly if we're using models for future scenarios, particularly under climate change, because if we're doing well in these non-streamflow variables, we're more likely to have a model that is capturing the dominant and important processes, and therefore we can have more confidence in the projections that it's giving under future climate change scenarios.
OK, so for the later part in Hansini's PhD, she looked at the actual evapotranspiration question. Now, there aren't that many places in Australia where we're actually directly measuring AET, and each of these colours corresponds to a location in Australia where we are. So from that you get like a time series of measured AET and then Hansini has set up a bunch of different models to simulate in those locations. And she's asking the question of or for a particular aspect of model performance - is it doing it well? So in this case, it's the coefficient of variation. So are the model simulations too variable in the AET simulations? Or maybe not variable enough?
And the result that she got was this one. Far too variable. Now, you might ask, OK, we've got measured AET in these locations, what happens if we calibrate directly to that in addition to streamflow? Which is the traditional way. If we do that joint calibration, we get a much, much better result which was pleasing to see. But unfortunately, it depended very strongly on which aspect of the dynamics of AET you focused on. So if we move that over to the side a bit and add another aspect, this is the strength of periodicity, so it's asking the question how seasonal is the signal? And you can see the results are pretty bad depending on the location, the models are either over or underestimating the strength of the seasonality of the signal. And when we calibrate, when we do that joint calibration, we include AET in the calibration, it unfortunately doesn't improve the situation.
OK, so much for AET. Now if we look - if we think about the groundwater trends, so the total water that's stored in the landscape. This was a paper that actually predated the project, but very relevant to the topic, so I'm including it here. So this is an example catchment, that so happens to be, well-instrumented from a groundwater perspective. So there's four different bores here in this top plot, and they're all showing a similar story. So a multi-annual decline in groundwater during the Millennium Drought.
So now what we would like to see, is we would also like to see some sort of multi-annual decline in the water that's stored in the model. And in the case of GR4J, and we tested a lot of other models and it seemed to be pretty consistent, we're not seeing that decline at all.
And the reason appears to be that essentially GR4J and other models are just - they're composed of buckets, right. And when these buckets are just running out of water every year, so they get to the end of summer and they're out of water. And once a bucket's empty, you can't remove any more water, right? So it's not possible to start to have these multi-annual deficits start to build up in these very simple conceptual bucket models.
OK, so then summarising across these, firstly, the models don't simulate AET very well even when calibrated to it, or at least that there were definite aspects of the AET behaviour that weren't being captured. And secondly, that models struggle with gradual trends like groundwater decline.
OK, so I've got one more thing to say about modelling and then I'm going to kind of skip to the end just in the interests of time.
So I'm just going to go over this paper that Luca presented at a ModSim. All right, so what we're talking about here is we're talking about calibration strategies, OK. So if we, if our aim is say to get good model performance during the Millennium Drought, we could of course just calibrate directly to the Millennium Drought, and that will probably give us good performance. And the problem, as noted earlier, is that when we calibrate only to the period prior to the drought, and then we test the model over this Millennium Drought, that it hadn't seen yet, it wasn't part of the calibration data,
we typically get a poor performance, OK.
So before I make the point that I'm trying to make, let me show you the results that correspond to those two results. So in the right-hand box plots here, if we calibrate directly to the Millennium Drought, we have low bias over the Millennium Drought, we have higher performance over the Millennium drought. But if we calibrate to the pre-drought, we have higher bias during the Millennium Drought and lower performance overall during the Millennium Drought. OK.
So then what this paper was testing was that it had been suggested that if you just give the model all of the information, both the pre-drought and the drought, and in this case, after the drought as well, we could do that too, then surely that will ensure
that the simulations during the drought and other periods too are pretty good. And unfortunately that is not the case at all.
In fact, using sort of standard calibration techniques, it seems as if that information that was given during the Millennium Drought essentially just got ignored in favour of the wetter years, and that's part of this sort of model bias to that, that often models are much more focused on the floods than they are during the dry years. So that was quite an insightful little paper that Luca did.
So the take home message is that naively including more calibration information is no solution.
OK, so I would love to talk about the question of, can you understand and give us more reliable models and I think we're going to circulate the presentation, I'll ensure that these the links to these papers are included. The take home messages are that we have found some simple changes that can make models more realistic, but it's really not easy and the most difficult part of it seems to be that extrapolation question. So our goal is calibrate a model on wetter years - it can perform in drier years, that's what we're aiming for, and that appears to be the most difficult aspect of this. So even if we create models that we know, if we give them all of that information in a special way, then they can actually produce simulations that are good throughout those contrasting periods. The extrapolation question, where you only calibrate them on one and expect them to perform on the other is a very difficult aspect of this problem.
OK, so this is the second last slide. Some other research challenges. So for those catchments that have not recovered, or that recovered recently, there's a research challenge there to understand what are the specific climatic or other circumstances that trigger a catchment to recover to its earlier rainfall-runoff relationship.
A big one is building process understandings into models, which was that last question in the table. We have a couple of PhDs looking specifically at the AET and groundwater aspects of this at the moment. So kind of watch this space.
How to calibrate models so that they can pass the crash test. That's what I was just talking about in terms of the extrapolation problem.
And then lastly, I think there's this broader challenge of better ways of incorporating uncertainty. I guess all different types of uncertainty, but particularly from models,
to make more robust decisions under climate change.
Page last updated: 27/03/25