A –Streams of Thought– contribution by Harsh Beria.

Dmitri Kavetski is a Professor of Civil and Environmental Engineering at the University of Adelaide in Australia. Prof. Kavetski is a renowned expert in the field of uncertainty quantification in hydrologic modeling, having developed Bayesian Total Error Analysis (BATEA) framework (Kavetski et al., 2006a; 2006b) which has been widely used in environmental modeling. He kindly accepted to answer our questions about his early career, his current research interests and how he sees the field of hydrology evolve over the coming decades.

Figure 1: Dr. Kavetski at work!

HB: Can you tell us a little about your background, your formal education and what got you excited in the field of uncertainty modeling in general, and Bayesian in particular?

DK: I always enjoyed mathematics, chemistry and other sciences. My first university semester, at the University of Newcastle in Australia, was in a Science degree. But I wanted to ensure that my knowledge was practical, and switched to Engineering. To be honest, I did not care exactly what field I was in – I was just enjoying learning new things. Environmental Engineering had a particularly diverse range of subjects, I was able take classes ranging from Fluid Mechanics to Programming to Differential Equations to Microbiology. I quickly developed an interest in numerical methods – I was fascinated how numerical methods can trivially solve problems often intractable to “traditional” mathematics. Plus I enjoyed the creative aspects of programming and being able to tell computers what to do.

For my PhD, I decided on hydrological modelling and uncertainty analysis. My lowest marks were in those subjects – but I went ahead because I wanted to learn both about system modelling techniques and about prediction under uncertainty, which I felt were extremely useful general skills that would serve me well in whatever I wanted to do in the future.

After a postdoc at Princeton University in the US, where I worked on a carbon sequestration project, I returned home to Newcastle, where I got a 5-year research fellowship that really helped me consolidate my research career. In 2012, I moved to the University of Adelaide to take a professor position in the School of Civil, Environmental and Mining Engineering.

HB: Can you talk a little bit about your current research projects?

DK: I try to maintain a mix of research interests. For example, with Martyn Clark and Fabrizio Fenicia, I am involved in several projects on improving and applying hydrological models (mainly at the catchment scale). We have this idea of a hydrological model needing to be sufficiently flexible to accommodate the range of application conditions, and that in order to properly test a model, one need to be able to test individual, often subtle, variations of this model. I am also very much interested in probabilistic prediction methods, where I collaborate with Ben Renard, Mark Thyer and George Kuczera on projects ranging from classic residual error analysis approaches to Bayesian hierarchical methods where we try to disentangle the contribution of different sources of error (for example, data errors in rainfall and runoff vs structural error in the hydrological model itself).

But I am very mindful that without practical application, research can become sterile, and so we actively engage with the water and environment industry. For example, over the last 8 years, we have been working with the Australian Bureau of Meteorology on seasonal streamflow forecasting, where we can test and apply many of our theoretical developments.

HB: Was there any non-traditional source of inspiration which got you interested in your research endeavour?

DK: Growing up, I was really into science fiction, for example, into Japanese manga cartoons about giant robots, space flights and all that. Although I did not get into electronics or robotics as a career, when doing work in mathematical modelling, prediction, and programming, I often visualise myself in that sci-fi universe. In other words, you gotta keep yourself amused (laughs)!

HB: What inspired the development of BATEA framework?

DK: During my PhD, I was interested in modelling systems while explicitly recognizing specific sources of uncertainty affecting prediction. As such, BATEA is the most comprehensive approach for dealing with the problem of prediction under uncertainty. But of course it is also the most demanding approach, computationally and data wise. So within the broad umbrella of the BATEA, we try to accommodate a range of approaches of varying degree of scope and complexity.

HB: The current era of global hydrologic modeling is one which uses global scale land surface models. How do we integrate BATEA in such a modeling framework and ensure computational feasibility?

DK: Well, scientists and engineers always try to push the envelope, to solve ever larger-scale and/or finer-resolution problems. Initially, when applying a method to a more demanding problem, one can make certain simplifications, especially in the computational sense. Later, the effect of these simplifications can be appraised and improvements made as required. This “from simple to complex” approach can be used when applying a framework such as BATEA to larger scale problems, and is something we are currently looking into.

HB: Can we theoretically come up with a framework which accounts for different modeling uncertainty (initial condition, input, structural, etc.) separately and hence improve our forecasts? How?

DK: Sure, theoretically. I would say the challenge is less in the mathematical “development” of such a framework, but in being able to collect the information needed to support such a framework. Our work with Ben Renard a few years ago (Renard et al., 2011) made some inroads into this challenge, but of course, much more can be done, and again, this is one of my research interests.

HB: Two of the most famous uncertainty frameworks in hydrologic modeling are GLUE (Generalized likelihood uncertainty estimation) (Beven and Binley, 1992) and BATEA. Both have distinct advantages. Do you see an integration of the two with computational feasibility?

DK: The motivations behind BATEA and GLUE are the same – to improve environmental model predictions not just in the sense of “absolute” accuracy, but just as importantly, in terms of furnishing meaningful uncertainty estimates. As an engineer, I am not inherently attached to any particular technique – what is important is the quality of results, which must be stringently scrutinized to ensure they are “fit for purpose”. In developing BATEA, we have closely followed the framework of probability theory and Bayesian inference, whereas in GLUE more latitude is taken in terms of heuristic variations. No doubt many of these variations deserve further attention, especially in terms of improving the representation of structural uncertainty in our models.

HB: What’s your take on ensemble modeling to quantify uncertainty?

DK: Ensemble modelling in itself is an uncontroversial technique. Indeed many if not most modern applied statistical techniques represent uncertainty using “ensembles” (“samples”) rather than analytical probability distributions. Like any other numerical technique, it comes with strings attached – for example, the ensemble must capture the full range of possible system behaviour. This in practice is the challenge to be met, especially when using model ensembles, because many models might be “wrong” in the same way (and hence may not add genuine diversity to the ensemble).

HB: A recent interview with Martin Clark came up on HEPEX where he was asked, if you had to choose between great parameters, great model structure, great parameterizations, or great forcings, which would you select. What’s your take?

DK: I am not sure how this choice came about. I think if you have good models and good data, the rest is (almost) trivial to obtain. But lacking one of these two, you are always hobbled. Maybe as a modeller you would prefer to have the perfect model structure whereas an experimentalist would favour the perfect data. Another reason why the two need to work together instead of perpetuating these rather artificial divides.

Figure 2: During a snow-depth measurement campaign near Mt Cook in New Zealand (October 2009)

HB: How much do you think does a modeller benefit from field work? A recent commentary from (Burt and McDonnell, 2015) advocated that young hydrologists should invest more in fieldwork. What’s your standpoint on this?

DK: Absolutely. To be honest, I am surprised at how little fieldwork (and lab work) is taught these days in environmental engineering degrees (I cannot speak for other disciplines). I think it exacerbates this disconnect between “modeller” and “experimentalist”, which in my opinion is completely unwarranted in the first place. If environmental variables become just columns of numbers, the modeller is bound, sooner or later, to make embarrassing – and potentially costly mistakes!

HB: What topics do you see yourself working in the next 10 years?

DK: I will be investing some time to deepen my collaborations with the water industry in Australia and elsewhere. I find these collaborations very insightful in terms of identifying research and application opportunities, and for expanding my research interests. Water availability is a major issue in Australia, as many of our catchments are arid or semi-arid. These are tough catchments to model, and are on my research agenda for the coming years.

But I also enjoy teaching, both at the undergraduate level and at the postgraduate summer schools, such as those we organize in Switzerland and Luxembourg. So teaching modelling concepts is another endeavour that will keep me busy.

HB: What are the biggest challenges for hydrologists in the next 10 years? What about the next 50 years?

DK: Well, I think in the next decade or so hydrologists will experiment with combining multiple modelling techniques and more extensive data sources. There is a lot of new developments to digest and properly test. I am also expecting gradual migration of operational hydrology to use current theoretical developments, adjusting them where necessary for practical realities.

Looking at hydrology in 50 years … now that’s much harder. So much will depend on technical resources available for computer modelling and environmental measurements. Who knows, maybe in 50 years hydrologists will be able to control instead of model catchment conditions? I am pretty sure hydrology in 50 years will bear little resemblance to today.

HB: In this era where thousands of research articles come out each week, how can the community stay up-to-date with the latest literature?

DK: Yes, there is a lot coming out, and it’s not always easy to keep up with. Still, the combination of the internet, conferences and chatting with colleagues usually does the job. If not, reviewers are quick to point to any omissions!

But that said, I think hydrologists – and scientists as a community – will be facing some decisions as the number of journals and papers proliferates. As we know, quantity and quality are often inversely related, and it is not clear if the current trends are sustainable.

HB: What advice do you have to offer to the young individuals having just started their research path?

DK: You’ve gotta find research topics you are genuinely interested in and enjoy working on, and develop an understanding and interest in the background, methods and outcomes. It is sometimes tempting to forego true deep understanding of scientific and engineering disciplines and swing towards more superficial application and presentation aspects. Treating methods and especially computer code as “black boxes” can really contribute to this. So make sure to invest in learning the fundamentals. It may disadvantage you a bit in the short term, but in the long term you will be on much stronger ground.

The other piece of advice I have is to strike the right balance between learning from older “grey beard” colleagues and collaborating with colleagues of your own generation.

HB: What is the most important decision you made to help you get there and why?

DK: You never get anywhere through a single decision, no matter how important it may seem. Many decisions do not even seem like “decisions” at the time you make them, for example, when you get interested in something, or pick up a new hobby. For me, one of the critical decisions I gradually converged to in the course of my life is to try to adopt “professional grade” standards in whatever I was trying to do, and to get things right from first principles. This in combination with hard work will take you a long way towards your goals.

You can learn more about Dmitri Kavetski and his research on his faculty home page.

References
Beven, K., and Binley, A.: The future of distributed models: model calibration and uncertainty prediction. Hydrol. Proc., 6 (3), 279-298. DOI:10.1002/hyp.3360060305, (2011).

Burt, T. P., and McDonnell, J. J.: Whither field hydrology? The need for discovery science and outrageous hydrological hypotheses. Water Resour. Res., 51 (8), 5919–5928. DOI:10.1002/2014WR016839, (2015).

Kavetski, D., Kuczera, G., and Franks, S.W.: Bayesian analysis of input uncertainty in hydrological modelling. 1. Theory. Water Resour. Res., 42 (3), W03407. DOI:10.1029/2005WR004368, (2006a).

Kavetski, D., Kuczera, G., and Franks, S.W.: Bayesian analysis of input uncertainty in hydrological modelling. 2. Application. Water Resour. Res., 42 (3), W03408. DOI:10.1029/2005WR004376, (2006b).

Renard, B., Kavetski, D., Leblois, E., Thyer, M., Kuczera, G., and Franks, S. W.: Toward a reliable decomposition of predictive uncertainty in hydrological modeling: Characterizing rainfall errors using conditional simulation. Water Resour. Res., 47 (11), W11516. DOI:10.1029/2011WR010643, (2011).

About the author
Harsh Beria (@harshberia93) has Bachelors and Masters degree in Engineering from Indian Institute of Technology (IIT) Kharagpur, and will start his PhD at University of Lausanne, Switzerland. Harsh is the social media manager and an active member of AGU Hydrology Section Student Subcommittee.