Interactive comment on “ HESS Opinions “ Hydrologists , bring out shovels and garden hoses and hit the dirt ” ” by M . G . Kleinhans

The paper by Kleinhans et al. is a HESS Opinion piece / commentary. I am uncomfortable picking too many holes in it-as these are their opinions and I respect them as such. In general, I am completely on board with the notion that hydrologists should do more controlled laboratory (and field) experimentation. The comments below offer some constructive criticism for revising the manuscript prior to publication.


Introduction
Viewed from the outsider's perspective of planetary science, or geomorphology, or meteorology, the science of hydrology uses but a subset of the tools for exploring nature as available to all geosciences.Much effort is put in field measurement and in physics-based modelling, wherein hydrological phenomena are reduced to the laws of physics following the optimistic agenda set by Freeze and Harlan (1969).Since their publication, fundamental problems of reductionism were encountered en route to a physics-based generally valid supermodel: underdetermination of model predictions by limited measurements of boundary conditions and initial conditions, underdetermination of model parameters, and underdetermination of predictions by ambiguity about the required level of simplification of physics-based relations in the model.The gravity of these problems, particularly the parameter problem, is attested by the fact that it was given a new name in hydrology: equifinality, but similar problems abound in the other geosciences.Yet there remained a societal need for hydrological predictions, so that Introduction

Conclusions References
Tables Figures

Back Close
Full Screen / Esc Printer-friendly Version Interactive Discussion much effort has been put into building and calibrating models for specific sites.It has been argued that this is a cul-de-sac for hydrology, because it does not lead to progress on big questions but leads to an unchecked growth of models applicable to one unique place only, which is exactly the opposite of what the reductionistic enterprise was about (Klemes, 1986;Beven, 2000Beven, , 2002;;Sivapalan, 2003).
"What then remains for the hydrologist to do if we take away from him the curve fitting, model calibration, the chasing of systems responses, correlations, finite elements, kriging, etc.?", as Klemes (1986) asked.Scientists in the first place want to understand nature.This is not to deprecate the relevance of applications for human interventions and predictions with benefit for human society.But science that only provides facts and useful predictions is impoverished; we want to answer the "why" questions (Mayr, 1985;Kleinhans et al., 2005;McDonnell et al., 2007).Application may or may not follow.
From the perspective of other geo-and extraterrestrial sciences and philosophy of science, one tool for exploring and understanding nature is nearly entirely ignored in hydrology: controlled laboratory experimentation (also see Hopp et al., 2009).This paper explores why the potential for novel insights and hypotheses from experiments is tremendous (also see Hacking, 1984, chapters 9 and 13).From the same perspective, there may be more potential for fresh insights from an experimental approach to physics-based modelling (explanatory modelling rather than calibrated modelling) than currently acknowledged.
The objective of this paper is to discuss why and how controlled experimentation can lead to new insights and is complementary to fieldwork and modelling.First we discuss basic characteristics, benefits and problems of field data and modelling and, in more detail, experimentation.Limitations, such as due to scaling problems, are extensively discussed and we will argue how these problems are not more grave than basic problems of fieldwork and modelling as discussed extensively in hydrological literature.Then we will illustrate briefly how several other disciplines of the geosciences and planetary science employ observations, modelling and experimentation to interrogate the real world.Using a brief exploration of the basic logic of scientific explanation, Introduction

Conclusions References
Tables Figures

Back Close
Full we show why experiments are ideally suited to hypothesis generation.Thus we will argue that more explanatory modelling and experimentation in hydrology will lead to better understanding of its major questions.

Three pillars of the geosciences
There are basically three ways in which geoscientists interrogate reality: 1) observations recorded in data, 2) established laws of nature implemented in models, and 3) intervention and manipulation in experiments (Fig. 1).Understanding is gained with all three, but in different ways and with different limitations which will be discussed below with emphasis on experimentation.That a combination of these approaches is more powerful than each alone needs no elaboration.When the results from all three epistemic approaches converge, we may foster hope that we possess good explanations for natural phenomena.

Observation: overwhelming reality in the field
Field data is as close as possible to reality.It contains variation in time and space that is of interest (pattern or signal) or is the result of processes under study in other sciences but left out here (noise).Direct derivation of understanding from data may be seriously hampered by incompleteness, inaccessibility or bias of the observer in the inference of hypotheses from field observation; i.e. due to the limited frame of reference of a mortal observer.
In general, geoscientific theories and hypotheses based on observations and data, ranging from mechanistical theories to explanatory reconstructions of past conditions, usually are underdetermined by the available data (Kleinhans et al., 2005).That is, there is insufficient available evidence to choose one theory over its rivals or select one set of model parameters over another.The term underdetermination is derived from the Duhem-Quine thesis, which states roughly that a theory is never testable in isolation from other theories and data (see Kleinhans et al., 2005, for discussion on weak and strong underdetermination; weak underdetermination is referred to here).The underdetermination problems are so pervasive in all geosciences and other sciences that they determine its daily practice to a large extent.Typical examples of underdetermination problems in all geosciences are (Kleinhans et al., 2005): -Measurement techniques may disturb the observed processes.
-The time scale of human observation is (much) shorter than that of the phenomenon under study.
-Many processes and phenomena cannot (yet) be observed directly or even indirectly.Erosional and sedimentary landforms of the past may have been obliterated by later erosion, and phenomena may not be accessible in practice.
-The laws of physics can seldom be applied directly to the initial conditions to check whether they explain the observations, because they must be applied in models with many different simplifications that provide different answers.
-Many processes are intrinsically chaotic in that they are very sensitive to initial conditions (spatial variation) and boundary conditions (temporal variation) of the system or area under consideration, which are then difficult to specify in enough detail.
An obvious example in hydrology is the practical impossibility to map the considerable heterogeneity of hillslope properties, such as topography (hillslope shapes and sizes), hydrogeology, soils and vegetation, in sufficient detail both within and between watersheds to explain the observed temporal and spatial variability in groundwater flow and surface runoff (Sivapalan, 2003).In fact, so many combinations are possible that every place becomes unique in an arbitrary sense that is nevertheless problematic for satisfactory explanation (Beven, 2000;McDonnell et al., 2007).To include the formative processes for these properties in a landscape model is impractical for hydrology Introduction

Conclusions References
Tables Figures

Back Close
Full Screen / Esc

Printer-friendly Version
Interactive Discussion because the time scale for landscape formation is usually much longer than for typical hydrological phenomena.To revert to the old-fashioned holistic landscape descriptions of physical geography is not fruitful as it does not lead to explanation.Thus data commonly support multiple explanatory hypotheses that are empirically equivalent due to the underdetermination problems, but contradict each other in meaning (Chamberlin, 1890).The clearest approach to reduce the number of hypotheses is by triangulation between many different parameters (e.g.Son and Sivapalan, 2007), such as done in geology (Kleinhans et al., 2005).Triangulation means selecting the explanation that fits observations of more than one types, in contrast to selecting an explanation that just fits observations of just one type.In the case of hydrology, that would mean not merely hydrograph fitting.Rather, it would involve simultaneous fitting to the other terms in the water balance such as groundwater dynamics, changes in soil moisture and evaporation, and fitting to other variables such as deuterium composition, which obviously requires data on these variables (Beven, 2000;Son and Sivapalan, 2007).

Model verification and validation
Models describe reality in terms of mathematical equations, usually at least partly based on laws of natural sciences.Modelling allows full control over specified boundary conditions and laws.Thus, a physics-based model may be used to test whether a hypothesis does not conflict with the laws of physics.For operational purposes such as prediction for catchments, various parameters have to be calibrated.
But a model contains various sets of laws involved, of which some are at best derived from physics, but even then are usually simplified to allow numerical solutions.Also, for many problems it is not obvious which laws apply, and to what extent simplification is possible.Such simplifications include model parameters, for instance parameters in macroscopic laws (e.g.Darcy) that represent more expensive and difficult to model Introduction

Conclusions References
Tables Figures

Back Close
Full microscopic processes (e.g.pore scale flow governed by the Navier-Stokes equations).Thus one cannot be certain whether a mismatch between model results and observations is due to the simplifications and numerical techniques or the underdetermined initial and boundary conditions.As such, extensive physics-based models are not very useful for simulating the details of a concrete existing case (Klemes, 1986;Konikow and Bredehoeft, 1992;Oreskes et al., 1994).
Particularly in hydrology, the incorporation of more physical processes in a model leads to inclusion of more parameters.The values of these parameters are usually poorly known, so that models need to be calibrated for each setting.But limited calibration sets allow a wide range of combinations of parameters that give the same results, well-known as equifinality.Furthermore, insensitive parameters are then very poorly constrained, which is known as the problem of parameter identification (van der Perk and Bierkens, 1997).These problems render model predictions inaccurate for different settings (Beven, 2000(Beven, , 2002)).
Put more precisely, it is fundamentally impossible to verify and validate models (Oreskes et al., 1994), nor to falsify models.Verification here means to establish their truth content, i.e. the choice of laws to represent the natural phenomena of interest.
Validation means establishment of legitimacy, i.e. the model is flawless and internally consistent.Confirmation or falsification refers to the (dis)agreement between observation and prediction, which merely support the probability of veracity, but do not verify a model (also see Klemes, 1986;Konikow and Bredehoeft, 1992).Note that this terminology is closer to the meaning of the words but differs considerably from the way it is commonly used (e.g.Klemes, 1986;van der Perk and Bierkens, 1997).
The world, as geosciences study it, is a mess of abundant intertwined processes, rich history and complexity.There are patterns, but there is also noise and accidents.Scientists have to limit themselves, but by building simplified representations and models of reality necessarily leave out much.The problems of underdetermination partly explain why the construction of a single, generally valid reductionistic model (see Freeze and Harlan, 1969)  is specified in (underdetermined) model parameters, initial and boundary conditions is the result of excluded processes, such as tectonics, landscape evolution, soil formation, climate change, and life.These processes and phenomena have their own disequilibrium dynamics, nonlinearity, thresholds and length scales.These and many other processes and phenomena together formed the Earth and left their imprint and their history, a snap-shot of which is the excruciatingly difficult-to-map of spatial variation.Then there are also accidents, such as landslides and other disasters waiting to happen, the exact course, initiation and timing of which depends on coincidental rainstorms or droughts, earthquakes and so on, which are hard to predict or even hindcast.Small wonder there is uniqueness of place (Beven, 2000).
However, the underlying physics, chemistry and biology of all these phenomena, including the hydrologic, are not unique; merely of varying importance.Models do not fit the data because the choice of relevant physics, temporal and spatial scales throw out much of the rich history, which then has to be brought back into the model as (underdetermined) initial and boundary conditions.Small wonder that our physicsbased models do not fit the data exactly.In geology and biology an explanation for a phenomenon is not complete without reference to both physical factors and history (Mayr, 1985;Kleinhans et al., 2005, and references therein).

The seductive Siren of parsimony
Parsimony is often mentioned as a guide to the question what physics (not) to include.
Yet parsimony, also known as Occam's Razor, is related to one of the most complicated issues in philosophy of science: simplicity.Indeed, scientist's understanding of Occam's razor and simplicity, its usefulness for science and what it actually says about the world or about our possibility to understand it varies greatly (Riesch, 2007).It is quite revealing that scientists pay lip service, or more, to the principle because they think that philosophers of science endorse it, whereas the latter study the principle and its use by scientists because scientists use it.Introduction

Conclusions References
Tables Figures

Back Close
Full Screen / Esc Printer-friendly Version

Interactive Discussion
Loosely put, a parsimonious explanation or model includes just enough elements to explain: not more, but not less either.Overparameterisation leads to underdetermination as is well known (e.g.van der Perk, 1997).Oversimplified models may beg the question, and are sometimes harder to apply because much needs to be specified.For instance, a one-dimensional flow model for meandering rivers needs much more parameterisation than a three-dimensional model because the latter solves the secondary flow pattern whereas the former has to parameterise it.Moreover, using different combinations of laws of physics can result in exactly the same outcome given uncertainty in parameters and input, as frequently occurs in hydrology and morphology.Unfortunately, it is hard to decide then which model is better and more parsimonious without being grounded again on the discussion of statistical measures for goodnessof-fit and uncertainty.In short, Occam's Razor is as helpful in geoscience as a rusty bread-knife for shaving (Kleinhans et al., 2005).

"Experimental" approach to explanatory modelling
Models are very useful to present results of complicated sets of equations that the unaided human mind cannot comprehend.They serve the same purpose for laws as data reduction does for data.In contrast to predictive (site-specific) modelling, explanatory modelling merely attempts to explain the general phenomenon under study rather than predict a case as accurately as possible.Explanatory modelling also covers the downward approach described in Sivapalan et al. (2003), where model complexity is increased step by step until the simplest possible model parsimoniously explains the general trends in the data of a specific case.
Furthermore, models are extremely useful to study sensitivity of results to certain parameters and to explore the viability of hypotheses given certain laws of nature (Oreskes et al., 1994;Kleinhans et al., 2005) and boundary conditions are required to yield a certain result.For specific cases, the downward approach can be turned around to do a diagnostic analysis: by adding and calibrating a single process description, the hypothesis that this process explains a certain aspect of the data can be tested (Samuel et al., 2008).Moreover, models can be used to twist a lion's tail as if they are an experimental facility in which certain effects are included, excluded or modified at will.This second way of explanatory modelling is experimental, without point-by-point comparison to data of a specific case but directed at explaining general trends and creating hypotheses.
For instance, an unexpected phenomenon was identified by combining a reductionistic biophysical model of plant growth and a saturated-unsaturated hillslope hydrology model.The plant growth model demonstrated that carbon assimilation rate was larger under slight water stress than under unstressed conditions.This was caused by reduced evaporation, which led to higher leaf temperature that in turn caused higher carboxylation rates (Brolsma et al., 2009a).When coupled to the hillslope model, a reduced precipitation scenario resulted in a groundwater level rise.This surprising effect was caused by reduced biomass upslope due to the water stress, which reduces interception evaporation, which in turn increased groundwater recharge (Brolsma et al., 2009b).Uncertainties in model parameters, initial and boundary conditions were hardly relevant in these cases because the model was applied to a hypothetical case.Rather, model-derived hypotheses such as these can be used in the analysis of field data and for the setup of new measurements and even for experiments dedicated to testing such a hypothesis.
To summarise: models can be used as tools to mediate between nature and theory based on physics, chemistry and biology (Morgan and Morrison, 1999); i.e., to gain understanding.Particularly explanatory modelling, including diagnostic analysis, hypothesis testing and experimental modelling to generate hypotheses, is useful for such understanding.Introduction

Conclusions References
Tables Figures

Back Close
Full

Materiality and serendipity
Experimentation by definition allows good control over initial and/or boundary conditions (Hacking, 1984), and involves to some extent the same materials as nature (Morgan, 2003) but with much better accessibility.Experiments also produce serendipitic results.Note that field "experiments" -a terminology often found in hydrology, biochemistry, coastal science and other geosciences -is then a contradiction in terms, because barely anything is changed and controlled in nature.The word experiment means nothing in that context as it merely refers to data collection.This is not to deprecate field analyses, which are very valuable.In particular, comparison of similar catchments in similar conditions but one significant difference such as climate or human interference may give a strong signal on the effect of such a difference (e.g.Likens, 2004;Samuel et al., 2008).Yet these field studies are hardly controlled if just one of many variable is controlled; even if the meaning of the term "experiment" is watered down then these "experiments" are only weakly controlled.
Materiality is maintained in experiments, contrary to modelling (Morgan, 2003).This is quite important, because the behaviour of the material (water, soil, plants) is unlimited by a simplifying description in terms of laws as in models.One could say that it is the ultimate reductionistic approach, for the material must obey all relevant laws of physics even if we do not yet understand which and how.For instance, all terms in the three-dimensional Navier-Stokes equations are "retained" in experimental flows; in fact, experimental setups have to be devised specifically to exclude certain effects or dimensions.Due to materiality, experiments may confound us like field observations, whereas models can only surprise us because we can go back to the underlying equations for understanding (Morgan, 2003).
Materiality also allows us to obtain a different sort of knowledge: "a feeling of what happens".This feeling for the behaviour of water and sediment is acquired by fiddling Introduction

Conclusions References
Tables Figures

Back Close
Full and tinkering with the experimental setup, materials and instruments, and much more so than in the field where unknown variability may overwhelm the pattern.When this feeling is added to data reduction and to description in terms of mathematically posed laws, it conveys a deeper understanding of what the latter describe and mean.This combination is also very powerful in teaching to students at all levels.
Hybrids of experiments and models represent reality in a variety of ways that are insightful to compare, such as simplification and degree of materiality (Morgan, 2003).Consider experiments and models on flow in pores.A volume of soil can be transferred to the laboratory and subjected to a variety of boundary conditions (de Rooij, 1996).Nondestructive three-dimensional mapping with Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) could be used to construct a digital representation (Lehmann et al., 2006;Kleinhans et al., 2008) for use in a flow model.In this case the experiment and the model are representative of soils with pores in reality.Alternatively, the pores could be incorporated in an experiment by artificial structures and sediments, such as glass beads and small porous pipes, and incorporated in a flow model by a network of channels with sizes according to a certain distribution (Lehmann et al., 1998;Joekar Niasar et al., 2009).In this case the experiment and model are representative for real soils.Such a variety of setups could test different hypotheses about the effect of pore structure and network on groundwater flow.

Experiments in hydrology
To be sure, hydrology has its experiments, albeit few.Three examples are given (also see Hopp et al., 2009).Salehin et al. (2004) studied the basic effects of sediment structure on hyporheic exchange with only eight experiments on two experimental heterogeneous sediment beds in a very small laboratory flume.They found that solute penetration was confined to a shallower region and led to faster near-surface transport compared to homogeneous beds.This was confirmed with modelling applied to heterogeneous and homogeneous beds.The artificially emplaced sand layers could easily be made more naturally in Introduction

References
Tables Figures

Back Close
Full Screen / Esc Printer-friendly Version Interactive Discussion larger flumes by having them formed by the flow and this is fruitful to explore.But the point is that this small-scale but careful experimentation revealed an explanation for a natural phenomenon by focussing on the contrast between homogeneous and heterogeneous beds.Michaelides and Wainwright (2008) compared hillslope-channel flow coupling in a Froude-scaled experiment to a numerical model applied at the experiment scale.Hillslope angle, channel angle, hillslope discharge and channel discharge were systematically varied in a setup with side-slopes to a straight channel.Effects of hillslope flow on the channel routing were identified by using more than one statistical measure for model performance, which is significant because concurrent changes in variables in the channel and on the hillslope propagate errors in the model resulting from process representation and/or model structure.In short, not only obvious parameters such as channel discharge and gradient determine flow velocity and depth, but also hillslope gradient and discharge.As in the previous example, a process relevant in nature was identified and understood by comparing results between experiments and between some experiments and a model.A large hillslope experiment is being designed in a Biosphere 2 dome (Hopp et al., 2009), where the temporal and spatial scale of the facility allows for interactions between vegetation, soil, water chemistry, subsurface and overland flow.The experimental scale (30×15 m) is large enough to allow natural evolution of spatial patterns and variability of these properties.Hopp et al. (2009) discuss the intricacies and limitations of the design of this large-scale experiment.Following the detailed study and modelling of this microcosm, hypotheses are expected that certain phenomena in the real world may have the same explanations as in the experiment.
In general these examples illustrate how the three pillars of the geosciences, fieldwork, experimentation and modelling, are complementary and can be combined in various ways (see Kleinhans, 2009b, and

Scale problems in experiments
Of course there are scale problems, unless the prototype system in nature is so small that it will fit in the laboratory.But this is as much an argument not to experiment as equifinality is an argument not to model.On the contrary, given underdetermination problems in field data and modelling, any addition to our toolbox to explore nature is welcome.Experiments may lead to new understanding if they are allowed to differ quantitatively from a prototype as much as an uncalibrated physics-based model would do.It also matters which aspects are under scrutiny whether scale effects are really problematic.
For instance, it could be argued that vegetation cannot be scaled to the laboratory.
Against experiments with vegetation on braided rivers it can be argued that the stems of the plants scale like Sequoia Gigantea to the Rubicon river.In terms of size this may be correct, but size does not matter here.The relevant properties of the plants are their hydraulic resistance as well as the strength their roots provide to the sediment (Tal and Paola, 2007).This added strength can additionally be quantified systematically with geotechnical and other experiments (Kleinhans, 2009b).Surely there are some species among the millions on Earth of which the sprouts under some controlled conditions produce the required vegetation effects for hydrological experiments.
In geomorphology, experiments have yielded many new insights, new empirical relations and new physics-based models ranging on spatial and temporal scales from sand particle interaction and turbulence in milliseconds to landscape (largest catchments) and sedimentary basin formation (entire seas) over millions of years.Typical scale problems, also relevant for potential hydrological experiments, are that a sand particle in the experiment represents a much larger volume of sediment in reality (Postma et al., 2008), or that very thin experimental surface flows have relatively low Reynolds numbers and relatively high Froude numbers and surface tension is important (Peakall et al., 1996) (although no-one really knows how much) compared to nature.Yet these experiments recreated natural phenomena at a small scale and led to new hypotheses that explain natural phenomena.Introduction

Conclusions References
Tables Figures

Back Close
Full Likewise for hydrology, it could be argued that it is at presently unknown how to scale from small catchments to large catchments, let alone from microscopic experimental catchments to natural catchments.Yet, whilst the same scale factors applied to the geomorphological examples, and for larger time scales too, this did not stop geomorphologists from doing such experiments with good results.A frequently applied safeguard against severe scale problems is applying basic laws of relevant processes for order-of-magnitude predictions at the experimental scale to help design experiments.For instance, if a certain ratio of subsurface flow and surface runoff is required in the experiment, a good experimental sediment with a certain conductivity and hydraulic roughness can be chosen from calculations with Darcy and Manning.Necessarily, this involves simplifications.In this particular experiment macropores and strong channel turbulence are ignored.It will not do, and it is unnecessary, to recreate a microcosm in which as many processes and details are included as possible.As in explanatory modelling, experiments are simpler than reality.
Scale problems can lead to quantitative bias and even different processes unrepresentative of nature, but apparently these do not fatally preclude the derivation of explanations for natural geomorphological phenomena (Peakall et al., 1996;Paola et al., 2001;Postma et al., 2008).Many phenomena are in fact nearly without scale and it is difficult to distinguish between erosive landscapes and alluvial fans in nature and in the laboratory if no scale is provided (e.g.Rodrigues-Iturbe and Ronaldo, 1997;Paola et al., 2001).This scaling has in fact been proposed as another alley to explanation of field data (Rodrigues-Iturbe and Ronaldo, 1997;Sivapalan, 2003;McDonnell et al., 2007) and could very well include experiments.In fact, it could be argued that a phenomenon is truly understood when it has been subsumed under the laws of nature and has been reproduced experimentally.
In summary, experiments have great potential to gain understanding in hydrology.The problems of scaling down to laboratory size are no reason not to try, as such problems have not impeded progress in geomorphology on similar spatial scales and longer time scales.Introduction

Conclusions References
Tables Figures

Back Close
Full Many different geosciences tend to focus mostly on one or two of the three pillars of the earth sciences.Hydrology needs no introduction to this readership, and the unemployed potential of experimentation has been remarked upon.By comparing some geosciences to hydrology we will illustrate why and how we can fruitfully combine fieldwork, experimentation and modelling, and where ignoring one of the pillars is damaging to science.A study on the reasons for differences in approach between the geosciences -is it the nature of the subject?Is there a socio-historical reason?Is the practice of these approaches so different that they diverged into different disciplines?-Is much beyond the scope of this paper.

Planetary science: comparing Mars and Earth
Planet Mars has been studied mostly by photogeologic interpretation.This is related to the fact that most questions on Mars are about a distant past, for instance how the planet developed tectonically, or how wet and warm the climate was billions of years ago.These questions are most directly addressed with interpretation of the surface features (Kleinhans, 2009a).Spectral remote sensing and ground-penetrating radar led to the first geochemical analyses and subsoil mapping in the past decade.Both terrestrial analogues and concepts are borrowed from geology, geomorphology and so on, but are sometimes over-interpreted and unconfined by available data and physicsbased modelling.
Physics-based modelling is rarely applied except for impact cratering; tectonic modelling has been applied as well as global aquifer modelling, climate modelling and some landscape evolution modelling.All these model exercises were explanatory, in part because most of the required input data are unavailable and in part because the focus is on major questions rather than detailed quantitative hindcasts. of the Missoula Flood events and similar events in Siberia at the end of the last glacial (Baker and Milton, 1974).Experiments applied to Mars are rare: near-surface atmospheric conditions and soil properties have been simulated to assess possibilities for the emergence and survival of amino acids (with implications for primitive life forms) (Ten Kate et al., 2005), box canyon were formed experimentally by groundwater sapping (Howard et al., 1988) and deltas formed in impact crater lakes have been recreated experimentally at a small scale (Kraal et al., 2008).
There is considerable scope for rapid progress in planetary science by modelling and experimentation borrowed from other sciences, including hydrology (see Kleinhans, 2009a, for review).For hypotheses and ideas established disciplinary boundaries are already fruitfully ignored.

Geomorphology
Geomorphology has combined field data analysis and experiments since before the term was coined.To be fair, the emphasis has always been on fieldwork while experiments usually were small-scale.Modelling was introduced with the advent of computers as in all natural sciences, but remained at a relatively low level of complexity compared to the models of meteorology and hydrology.Large-scale Froude-scaled experiments in fluvial engineering have largely been replaced by numerical modelling for cost reasons and because of high faith in models outside the modelling community.
Within the community there are model comparisons, not to determine which is the best, but to learn about nature from differences between the models and their outcomes (e.g.Davies et al., 2002, in coastal science;e.g. Nearing et al., 2005, in soil erosion).
The relation between fieldwork, experiment and modelling in geomorphology is not easy.Although geomorphology is identifiable as a discipline by a number of conferences and journals, it lies on the overlap between more field-oriented geology (Quaternary geology, sedimentology) and more model-oriented sciences (civil engineering, geophysics).Apart from the contrasts in quantitative and qualitative approaches there 6597 Introduction

Conclusions References
Tables Figures

Back Close
Full is also the usual misunderstanding between their fundamentally different questions: "what was the cause in the past" of the geologists versus "what were the laws involved; how does it work" of the process-oriented sedimentologists and morphologists (Baker, 1996).
There is much scope for fast progress in geomorphology by combination of experiments in large-scale facilities as well as more sophisticated modelling on supercomputers as is common in meteorology (see Kleinhans, 2009b, for review on fluvial morphology).

Meteorology
Meteorology has a long tradition of observation, and has probably the largest and most diverse data collection system on Earth based on a high density network of sensors as well as remote sensing.
Several sophisticated models are run continuously on supercomputers for ensemble forecasting, while their initial and boundary conditions are continuously updated by data.The models have not and could not have been developed by individuals; they are community models that are often compared against each other to learn from the differences.Also climate modelling proceeds in this manner, and explanatory modelling as well as scenario modelling is a key activity.The obvious societal relevance is one reason why this science has much more resources than other geosciences.
Weather prediction by models did not improve gradually over the past decades.
Some improvements of the models, such as the replacement of parameterisations by more physics, actually led to deterioration of the predictive capacity.This apparent paradox led to the search for other model elements that needed improvement, and in the longer term the models improved considerably (G.Komen, KNMI, personal communication, 2004).An experimental approach to modelling is common; model improvement was not obtained by extensive calculations of model uncertainty but by twisting the model's tail to "experimentally" determine whether a certain model element is responsible for a certain phenomenon.Experiments are less common in meteorology and related sciences.One reason perhaps is that the intrinsic scales of weather are large (shower, cyclone) and hard to scale back to the laboratory.Another reason is that theory derived from experiments is borrowed from other sciences, for instance in wind tunnel experiments some boundary layer descriptions and turbulence closures are directly based on fluid dynamics.The same can be said of oceanography and glaciology, which collaborate with meteorology in climatology.The border between such sciences is obviously arbitrary but results from the disciplinary boundaries defined by tradition.So, if we forget these for a moment, then meteorology does indeed use experiments for the study of fundamental processes in controlled conditions.Furthermore, we note that meteorology is interdisciplinary, as is hydrology (Klemes, 1986).
Future progress at the border between hydrology and climatology can perhaps be made in combining experimental work on microclimate, vegetation and landuse with the large-scale forcing by vegetation of precipitation and evapotranspiration forcing (Pielke Sr., 2008).

How we explain: geo-logic and hypothesis conception
So far the reasoning why hydrology could benefit from experiments proceeded by analogy with other geosciences and qualitative arguments such as materiality in experiments.In the following section we explore in more depth how earth scientists arrive at hypotheses and explanations through three elementary pieces of logic, and why and how experimentation plays such an important role in the generation of hypotheses.All three are employed in various combined ways in fieldwork, modelling and experimentation.Introduction

Conclusions References
Tables Figures

Back Close
Full

Deduction, induction, and abduction
When scientists are asked how their science works, they commonly and rightly refer to induction and deduction.Statistical generalisations such as Hack's Law were obtained by induction, whereas physics-based prediction is a typical deductive exercise.But the practice of science hardly involves following the recipes of deduction and induction.In fact, there is not agreement yet among philosophers of science as to what amounts to an explanation and what exactly is understanding (see, e.g., Lipton, 1991).Three applicable types of logical reasoning are based on causes, effects and laws, two of which are necessary to arrive at the third (Fig. 2).With deduction and induction, the third possibility is abduction.They are not merely alternatives but answer different types of questions.Neither are different sciences limited to one of them, but all three are employed in all sciences, including hydrology (Kleinhans et al., 2009c).New hypotheses are conceived through abduction (this is not an entirely complete and correct account but sufficient for now; for an authoritative account see Lipton, 1991).The term abduction was coined by the american philosopher C. S. Peirce more than a century ago but has surprisingly remained unknown by most scientists (except Baker, 1996).Some are clearly ill at ease with it or ill at ease with the fact that something is missing from our vocabulary to describe what we do as scientist; for instance Savenije (2009) called it the "art" of science.We will explain that it is a form of logic in its own right and we will argue that experiments are ideally suited for abduction, which is why experiments may yield many new hypotheses about the world.

Deduction
For deduction, the initial conditions (causes) are combined with laws of nature to explain or predict the effects (Fig. 2).This is what happens in analytical solutions for linear stability analyses and physics-or chemistry-based modelling to solve boundary value problems (e.g.Freeze and Harlan, 1969, in hydrology).For specific sites it has obvious relevance such as flood forecasting.Deduction is a solid form of logic compared to induction and abduction.Its Achilles heel is in the choice of relevant laws and the common use of generalisations rather than laws, and the initial and boundary conditions which must be based on measurements that may be incomplete or contain errors (Oreskes et al., 1994) (also see Sect.2.2).

Induction
Induction leads to (statistical) generalisations based on both causes and effects (Fig. 2).Interpolation and extrapolation are induction too (see examples in Klemes, 1986).Induction yielded useful generalisations in hydrology and in the other geosciences, such as Hack's "law".
The problems of induction are well known: the validity range of empirical relations is determined by the range and bias of the data included, and the amount of data is obviously never large enough to create a universally valid generalisation, that is, law.
Nevertheless empirical relations somehow contain information about reality and have shown the way to underlying mechanisms in the past, but not infallibly so.

Abduction
In abductive inference, final conditions, facts and so on are (often implicitly) combined with laws or generalisations of nature, to arrive at the best of a limited number of hypotheses that explain the observations.It starts with a surprising observation, followed by an insight how the phenomenon may have come about.Thus abduction leads to hypotheses, including hypotheses about conditions or events in the past that led to the present phenomenon under observation.Such hypotheses can then be tested by modelling as in the downward approach, in diagnostic analysis, in explanatory modelling and in an experimental approach to modelling.The hypotheses can also be tested in experiments, including generalised simplified setups and scaled experiments dedicated to a unique place.The major limitation of abduction is that one cannot be certain that all possible hypotheses, including the correct one, have been conceived.The right hypothesis might be one that no-one thought of.For example, several clues from geological investigations, combined with a law that iridium must come from outside the Earth, led to the hypothesis that dinosaurs became extinct after a comet impacted.This is an example of abduction, which earth scientists and also detectives commonly employ (Baker, 1996;Kleinhans et al., 2009c).Thus abduction led to one process-oriented narrative of what happened, but alternative hypotheses have also been formulated and the jury is still out.The inference of a perceptual model, a perspective view of the watershed's functioning (Sivapalan, 2003(Sivapalan, , 2009), is also an example of abduction from end results (observations) to relevant laws, generalisations and boundary conditions.
Abduction is the most interesting for geoscientists even though it is perhaps the most fallible of the three, precisely because it leads to new ideas and hypotheses.In some post-war philosophy of science the practice of science was divided between the context of justification or falsification of theories, that is, the science, and the context of discovery, that is, the magic.The former was considered open to philosophical and logical analysis, whereas the latter was considered the realm of psychologists.
The account of abductive inference (C. S. Peirce and later work, see Lipton, 1991;Baker, 1996;Kleinhans et al., 2009c) puts the conception of ideas back into science where it belongs.New ideas are often inspired by observation in the field, by experimental playing with models and, importantly, by experimentation.In the cases of modelling or experimentation, abduction follows: a result may resemble a natural phenomenon of interest, from which it can then be abduced that the natural phenomenon is perhaps also explained by the same mechanism and initial and boundary conditions as in the model or experiment.
At the lower level of the experiment itself, abduction also plays a major role.Experiments may yield many surprising results.Given the mechanisms in play and the results, the causes must then be abduced, so that hypotheses are generated how the surprising result came about.Contrary to field observations, the initial and boundary Introduction

Conclusions References
Tables Figures

Back Close
Full conditions are relatively well known, which makes the abduction easier and less speculative.Contrary to models, there is materiality in experiments, which enhances understanding, the likelihood of surprises and the relation to reality.Hence, if a hypothesis is derived from experiments, it is not unlikely that the same hypothesis applies to a realworld situation.Thus experiments are instrumental in the generation of hypotheses and understanding.

Conclusions: a bright future for hydrology
Based on a comparison between different geosciences we submit that much progress can be made where one of the pillars of the geosciences (fieldwork, modelling and experimentation) was less frequently employed.Hence, hydrology (and teaching in hydrology) would benefit from more experimental work: -Experimentation leads to novel ideas and hypotheses for major questions of hydrology, mostly through abduction.
-Feeling and manipulating the material in experiments adds insight to observation and modelling like playing music adds insight to listening and studying mathematics of music.
Experimentation is an art that needs as much work as modelling or fieldwork to master (like learning to play a musical instrument).Furthermore, investment is required: first in small and cheap experimental facilities for fast exploration and hopefully at some point in large facilities to overcome (or prove negligible) certain scale effects and allow larger systems and more detailed measurements.These practical problems are surmountable through collaboration with the experimenters of, e.g., civil engineering and geomorphology in existing facilities.
It is well known that a simpler calibrated model may be more accurate and much cheaper to deduce predictions for a unique location than a complicated reductionistic Introduction

Conclusions References
Tables Figures

Back Close
Full model given computational cost, required level of detail of initial and boundary conditions and the fundamental problem of model verification and validation.However, reductionistic models are extremely useful for gaining understanding and testing viability of hypotheses -in short, by "experimentation".An experimental attitude to modelling complements the proposal by Sivapalan (2009) for pooling of data on large watersheds and the abductive analysis for major questions of such data.Physics-based models can be applied explanatorily to theoretical and more practical cases to learn by induction what general trends can be found from the incorporated laws of physics.Physical, chemical and biological processes can be introduced step by step to assess their effect on the trend.
There is much scope for experiment design and explanatory modelling that do not focus solely on hydrology but also on the coeval morphology, soil, ecology, microclimate and so on.The landscape evolution then includes the spatiotemporal variation and its effects on the hydrology.Also the results of landscape models, or ecological models, or network, pore and connectivity models could be used as input for hydrological models to compare general hydrological trends for contrasting inputs and compare these to analytical solutions and experimental results.The aim of such exercises is not to fit the hydrograph, so to speak, but to abduce and test hypotheses and gain understanding of major hydrologic issues.Handling water and dirt, even in small experiments, will enhance this understanding considerably.Introduction

Conclusions References
Tables Figures

Back Close
Full . Given the chosen laws in the model, it can be studied what result these laws in this model predict if a certain set of initial and boundary conditions were the case, or whether an emergent (statistical) relation exists between initial and boundary conditions and model outcomes, or what initial Introduction Several studies on Mars helped to understand Earth better.For instance, tremendous floods left clear marks on Mars and helped to interpret less clear marks on Earth Introduction is fundamentally impossible.In the case of hydrology, much of what Introduction