summary and synthesis of workshop …...summary and synthesis of workshop break out group...

SUMMARY AND SYNTHESIS OF WORKSHOP BREAK OUTGROUP DISCUSSIONS

Paul H. Whitfield1,2,3, R.D. (Dan) Moore4, and Kevin Shook1

1Centre for Hydrology, University of Saskatchewan, Saskatoon, SK, S7N 5C8 2Environment Canada, Vancouver, B.C. V6C 3S5

3Department of Earth Sciences, Simon Fraser University, Burnaby, BC4Department of Geography, University of British Columbia, Vancouver, B.C.

With input from the workshop Group Leaders:

Chris Spence5, Sarah Boon6, Ross Woods7, Carmen de Jong8, Timothy Link9,David Garen10,,Julie Kiang11, Denis Hughes12, Ian Littlewood13,

James McPhee14, and Anil Gupta15

5Environment Canada, Saskatoon, SK6University of Lethbridge, Lethbridge, AB

7National Institute of Water & Atmospheric Research, Christchurch, New Zealand8Mountain Centre, University of Savoy, Le Bourget du Lac, France

9Department of Forest Resources, University of Idaho, Moscow, Idaho, USA10US Department of Agriculture, Natural Resources Conservation Service, Portland, Oregon, USA

11US Geological Survey Reston, Virginia, USA12Institute for Water Research Rhodes University, Grahamstown, South Africa

13IGL Environment, Didcot, UK14Universidad de Chile, Santiago, Chile

15Alberta Environment, Calgary, AB

20.1 ABSTRACT

The workshop discussions about making predictions in ungauged basins indifferent hydrological landscapes with different states of data availability aresummarized. While the science underlying hydrological prediction hasadvanced considerably during the past decade, implementing the new

271

20

science into practice remains a challenge. The workshop participantsidentified a number of opportunities for addressing this challenge andrecommended developing a watershed classification system with ageneralized diagnostic to facilitate the transfer of results from researchcatchments and other gauged watersheds to ungauged systems. These toolsshould be developed adopting an open source framework with betteroutreach to enhance the accessibility and adoption by practitioners.

20.2 RÉSUMÉ

Sont résumées les discussions de l’atelier quant aux prévisions en bassins nonjaugés dans différents paysages hydrologiques comportant différentessituations de disponibilité des données. Bien que la science qui sous-tend lesprévisions hydrologiques ait fait des progrès considérables au cours de ladernière décennie, la mise en pratique de la nouvelle science représente encoreun défi. Les participants à l’atelier ont cerné un certain nombre d’occasions decomposer avec ce défi et ont recommandé l’élaboration d’un système declassification des bassins hydrographiques favorisant un diagnostic généralisépour faciliter le transfert des résultats des bassins de recherche et autresbassins jaugés aux bassins non jaugés. Ces outils doivent être conçus enadoptant un cadre de source ouverte d’une meilleure portée afin d’en accroîtrel’accessibilité et l’adoption par les professionnels en exercice.

20.3 INTRODUCTION

This paper provides a summary of the break out sessions from the PuttingPUB into Practice [P3] meeting in Canmore, Alberta, in 2011. At theworkshop, issues related to prediction in ungauged basins were discussed inrelation to (1) type of landscape (e.g., high mountain, boreal), and (2) dataavailability (data-rich, data-sparse, and data-poor). This summary focuseson commonalities and differences of PUB challenges across landscapes anddata richness. Through this comparison, we seek to share and consolidatebetween and across:

• The PUB themes and working groups,• The variety of regional efforts and perspectives,• The different approaches that maximize the predictive value of

streamflow data, other data, and their use,

272

Putting Prediction in Ungauged Basins into Practice Whitfield et al

• The approaches that maximize the use of physically based theoryand process structure, process variability, and their emergence intopredictive approaches,

• The inclusion of new measurement and information technologies formeteorological inputs, process verification, and catchmentcharacterization, and

• The continuation of the exploration of improved models and tools thatreflect improved hydrological understanding and their use in practice.

One key task was to identify opportunities for future developments or newperspectives that would contribute to the main issue of turning research intoaccessible tools that improve the practice of making predictions in ungaugedbasins. In particular, could approaches developed and validated for data-richareas be adapted as prediction tools for data-sparse and data-poor areas?All of the break out groups addressed common discussion points, but eachfocused on a different hydrological landscape. The participants representeda broad cross section of researchers and practitioners having a range ofexperience and expertise in hydrology and in different landscapes. Table21.1 contains the guidance that was provided to each of the six break outgroups over the three days of the meeting. A list of the workshop participantsis provided in the Appendix to this volume. This summary begins with a brief description of the attributes andmonitoring issues associated with the six hydrological landscapes uponwhich the discussions were focused. Then, the scales and methods ofhydrological analysis being used in predictions in ungauged basins areaddressed. The barriers to adoption and implementation of new methods asidentified by the participants are described. Following on thesecommonalities, the specific issues characterizing the hydrologicallandscapes are provided. The summary ends with a series of recommendedactions, research, and tools for future work.

20.4 HYDROCLIMATIC / LANDSCAPE REGIONS

The key attributes of the hydrological landscapes considered by the breakout groups are given in Table 21.2. The key types of predictions required inungauged basins are also identified. Water supply is an issue common to alllandscapes, but each landscape has a specific set of relevant hydrologicalattributes.

273

20 – Summary and Synthesis of Workshop Break out Group Discussions

274


a) Data-Rich

b) Data-Sparse

The PUB issue in “Data-Rich” areas is not whether good predictions can be made

for ungauged sites as both empirical and deterministic modelling approaches are

likely to be successful. The challenge in these situations is to learn about how to

use those methods to reduce the uncertainty of predictions in ungauged basins.

1) How can the various approaches for hydrological prediction in this

hydroclimatic region be implemented given the availability of

meteorological and catchment data and current understanding of

hydrology?

-small spatial scales, short time scales

-large scales, longer time scales

2) How can PUB predictive approaches be improved? Is there additional

process understanding required or additional data required?

3) How can the available hydrological tools contribute to products usable

by practitioners?

4) How can information gleaned from data-rich regions be applied to more

data-poor regions?

The PUB issue in “Data-Sparse” areas is whether good predictions can be made

for ungauged sites using modelling approaches that are ‘likely’ to be successful.

How can empirical information be used to extract sufficient information from the

limited observation records to validate models where data are sparse? The

challenge in these situations is to learn about how to use those methods to

understand the uncertainty of predictions in ungauged basins.

1) How can the various approaches for hydrological prediction in your



hydrology?



2) How can PUB predictive approaches be improved? Is there additional

process understanding required or additional data required?

3) How can the available hydrological tools contribute to products usable by

practitioners?

4) How can information gleaned from data-sparse regions be applied to

more data-poor regions?

How can we address practitioners’ needs for toolsto do PUB?

How can we extend the information based upondata-rich PUB to other areas?

How do we address practitioners’ needs for tools todo PUB?

How can we extend the information basedupon data-sparse PUB to other areas?

Table 20.1 The guidance provided to each of the break out groups during the workshop.The three parts took place on successive days.

Arid and semi-arid

A semi-arid (sub-polar) region receives precipitation equal to or less than thepotential evapotranspiration. Many semi-arid areas are characterized byhigh spatial variability in rainfall making it difficult to quantify areal rainfallinputs into hydrological and water resource simulation models (Paturel etal., 1995; Andréassian et al., 2001; Fekete et al., 2004). Semi-arid areas exist in both cold (e.g. Tundra) and warm (e.g. Prairie)regions, and groundwater may be more important than rivers as a waterresource. Temporary streams are common in semi-arid regions (Buttle et al.,2012). Intensive agriculture and irrigation are often widespread in warmsemi-arid regions. Hence, in semi-arid regions the effect of land use mayovershadow those of climate and weather, increasing the complexity ofmodelling and analysis.

275


c) Data-Poor

The PUB issue in “Data-Poor” areas is whether good predictions can be properly

validated. How can empirical information be used to extract sufficient data

coverage from the limited observation records to validate models where data are

poor? The challenge in these situations is to learn about how to use those

methods to understand the uncertainty of predictions in ungauged basins.

1) How can the various approaches for hydrological prediction in your



hydrology?



2) How do predictive approaches need to be improved? Are there options

other than additional process understanding required or additional data

required?

3) How can the available hydrological tools contribute to products usable

by practitioners?

4) How can information gleaned from data-poor regions be better applied

to other data-poor regions?

How do we address practitioners’ needs forpractical tools to do PUB in data-poor areas? Tools for validatingor confirming predictions?

How can we extend the informationbased upon data-poor PUB to provide feedback to thedevelopments made in data-rich and data-sparse areas?

Table 20.1 (cont’d)

276


Type Sub-Type Characteristics & PUB Connection

Arid &

Semi-arid

Mountains

Temperate

Agriculture

Temperate

Forests

Arctic &

Boreal

Tropical

Warm

Cold

Warm

Cold

Warm

Cold

Taiga

Tundra

Humid

Semi-arid

Potential Evaporation exceeds precipitation

Groundwater plays an important role

Extensive grasslands

Drought, flooding, and water supply

Steep elevation and vegetation gradients

Snow, snowpack development, and melt processes

High watershed gradients

Glaciers and glaciation features

Hydropower, fish habitat, water supply, and flooding

Extensive soils

Large conversion of landscape through tillage and

modifications

Hydrological modifications also extensive; storages,

abstractions

Drought, wetlands, and ecological values

Evaporation

Large conversion of original forests

Insects, deforestation, and forest rotation

Sediment transport

Hydropower, fish, aquatic habitats, and flooding

Long, extremely cold winters

Latitudinal vegetation gradients

Low station densities and bias towards very large

basins >100 000 km2

Hydropower, resource development, and

ecological values

High rainfall intensity and depth

Strong seasonal rainfall regime

Seasonally hydrophobic soils

Large surface runoff components and high sediment

loads

Drought, flooding, sediment transport, water supply,

and groundwater

Table 20.2 Summary of the main attributes of the hydrological regions used in thediscussions. The text in italics indicates the predominant needs for predictions inungauged basins in the region.

Streamflow records are often sparse in semi-arid regions as manywatercourses lack continuous flow. In many jurisdictions, gauges areoperated seasonally or only during short periods of the year. Rainfallnetworks are too sparse to adequately observe precipitation processes that arestrongly convective in summer and predominantly snowfall during winter.

Mountains

Mountainous areas generally receive greater precipitation than lowlandareas and have low evapotranspiration, and thus are efficient generators ofstreamflow that is a critical water resource in densely populated downstreamareas. Much of the precipitation is stored as snow or ice for periods of time(months in the case of snow, many years in the case of ice) and is laterreleased during the spring-summer melt period. In many high mountainregions, glaciers play an important role in supplementing streamflow in latesummer and early autumn and regulating the interannual variability ofstreamflow (Fountain and Tangborn, 1985; Stahl and Moore, 2006).Mountainous areas generally have low population densities and are poorlygauged outside of research basins. Weather stations tend to be located in valleybottoms and do not represent the higher elevations due to orographic effects.In addition to this dependence on elevation, key hydroclimatic variables(temperature, precipitation, solar radiation, humidity, wind speed) also varystrongly with slope and aspect and as a result of complex interactions betweenweather systems and topography, such as seeder-feeder precipitationprocesses and rain shadow effects. Extrapolation of hydroclimatic informationfrom weather stations to account for this spatial variability is a key challengein making hydrological predictions in mountain regions.Gauging stations are also predominantly located in valley bottom sites on mainstem channels. As a consequence, streamflow records integrate runoff from abroad range of elevations, and measured streamflow may not represent thequantity and timing in smaller headwater catchments. Due to the combinationof sparse monitoring networks and biased station locations, almost every waterresource analysis in mountainous areas is an exercise in PUB.

Temperate forests

This landscape comprises the forested areas between the tropics and theboreal forest. In this landscape there have been conversions of large areas ofthe original forests through harvesting and subsequent replanting or

277


regeneration for repeated harvesting cycles. These areas are also subject tochanges that accompany natural disturbances such as wildfire and insectinfestations, as well as deforestation associated with conversion of land toagricultural, urban, and other non-forest land use. The hydrological variablesbeing predicted in ungauged temperate forest basins are principally runoffand evaporation. The types of hydrological predictions required in theseareas are wide ranging, from water supply estimates to hydrologicalvariables that support protection of ecological values.As this landscape contains highly populated settlements, temperate forestsare generally better monitored than other regions with respect to bothclimate and streamflow; however, monitoring networks for climate andstreamflow are generally disconnected from each other as the networksdeveloped separately to meet different needs. This often results inpractitioners having access to local observations (e.g., of precipitation andstreamflow) but not from the same watersheds, so that data extrapolation isa key challenge in PUB applications.

Temperate agricultural

Temperate agricultural landscapes are typified by extensive well developedsoils. Typical of these areas is the extensive conversion of landscape throughtilling, draining, and other modifications. Hydrological modifications are alsoextensive. Drainage alterations, drainage of natural storages such as wetlandsand creation of artificial storages, and water abstraction and augmentationsthrough irrigation result in watersheds with altered hydrological responses. Inaddition, drainage areas may become disconnected for periods of timeresulting in variable watershed contribution areas.With these extensive and complex modifications of the landscape and wateron the landscape, monitoring and process-based observations alone will notbe sufficient to model these systems; effective models for prediction inungauged basins must take into consideration these landscape changes.

Arctic and boreal

The boreal forest is a circum-global ecozone dominated by long coldwinters, peat deposits, and coniferous forests. North of this is the arcticregion; above treeline the vegetation is tundra. Evapotranspiration is thedominant hydrological flux in the boreal and arctic region. A key feature ofthe region is the importance of storage of water on the landscape both in

278


shallow lakes and in frozen form as seasonal snowpacks, and in somemountainous catchments, as multiple-year to decadal storage as glaciers.Outside of research basins, forcing data are only available at a coarse scale.Hydrometric gauge density in the Arctic Ocean drainage basin has remainedstatic at ~1 per 104 km2 (Shiklomanov et al., 2002). While ~70% of the pan-arctic area is gauged, most of that area is observed only at the mouths of verylarge rivers, such as the Mackenzie (1.8 million km2) and Lena(2.5 million km2) basins (Prowse and Flegg, 2000). Even though most of thelargest basins are gauged, data collected at the mouths of these major riversystems does not adequately provide a representative picture of the flow regimecharacteristics of the smaller watersheds that feed into these large systems.Processes operating at small spatial scales generally differ from those foundwithin the large river systems. The fraction of small catchments (i.e. less than10 000 km2) that are gauged remains unacceptably small (e.g. 0.8% in theMackenzie). Region-specific flow characteristics that are critical to ecosystemhealth (Poff et al., 1997) are masked in the flows of the larger system and oftenrequire decades of times series data to understand (Burn et al., 2008). Without the benefit of local, region-specific monitoring, important annualand inter-annual flow variation of the smaller river systems is not wellobserved and is often not understood; this can have substantial importanceto local and regional communities. Because few small basins are monitored,the current sample of gauges cannot be assumed to be representative of therange of basin characteristics across the arctic and boreal landmass – thusthere remain many catchment types (particularly those that are small and/orglacier fed) for which there is little to no data (Spence and Saso, 2005;Spence and Burke, 2008). The small sample and inherent spatial andtemporal variability in runoff response from these catchments increasesuncertainty in hydrological prediction for this region.

Tropics

The humid tropics and semi-arid tropics have distinctive hydrologicalcharacteristics that distinguish them from each other and from otherhydroclimatic zones. The humid tropics are characterized by high rainfallintensity and depth, generally with a well-defined seasonal rainfall regime.These characteristics can lead to high volumes of surface runoff, highsediment loads, and seasonally hydrophobic soils. Temperatures are warmyear-round. The semi-arid tropics are much drier than the humid tropics, but

279


also exhibit strong rainfall seasonality. In the humid tropics, large scale landuse change is one of the major water management issues (Eden and Parry,1996), as conversion of forested lands to agricultural use can result inchanges to the seasonal distribution of runoff. This, in turn, can havesignificant effects on water use and availability, water quality, sedimentloads, and ecology. In the semi-arid tropics, the main water managementissues are somewhat different, instead focusing on drought prediction andmanagement (Mishra and Singh, 2011). Prediction of droughts and droughtfrequency, estimating yields from reservoirs or other small scale waterresource structures, and efficient conjunctive use of surface water andgroundwater are all important water management objectives. Hydrological processes of particular importance in the tropics include:interception loss, especially within the dense multi-layered canopy oftropical rain and cloud forests; vegetation-atmosphere, surface water-groundwater, and soil-atmosphere interactions; and changing land use,especially given very high rainfall intensities. The latter is important in thesemi-arid tropics as mentioned previously. Understanding these importantprocesses in the catchment of interest goes hand in hand with data collectionsince the data help to illuminate important processes, and an understandingof the important processes can inform what data are needed. Throughout the tropics, data availability is limited and the reliability of datacan be problematic, making prediction of flow difficult for large ungaugedareas (Hughes, 2006). Data-rich areas are generally limited to a smallnumber of well-resourced areas, and extrapolation to the majority of thecatchments which are ungauged represents a serious challenge for bothscience and practice. Many areas also lack centralized water resourcemanagement institutions, which further exacerbates the problems of dataaccess and contributes negatively to the sharing of expertise in waterresources estimation methods.

20.5 METHODS FOR PUB APPLICATIONS

The methods currently in use for prediction in ungauged basins are varied.Simple approaches that are in common use include data transfer methodsand techniques such as rule curves and the rational method, in addition tofield-based methods based on channel morphology. In areas with sufficientdata, statistical approaches such as regional regression analysis are popular

280


among practitioners. Process-based models, both conceptual and physicallybased, are rarely applied, although substantial research effort has focused ondeveloping and testing them.

Data transfer

A common approach is to identify a gauged catchment that is judged to besimilar to the target catchment, and then to adjust flows to account fordifferences in drainage area. Back-of-envelope adjustments can be made toaccount for differences in glacier cover or other attributes. This approachcan be improved upon by installing a short-term gauge on the target streamto verify any empirical relationship between the target stream and thegauged stream. In regions where the basin contributing areas are notconstant, this method can be very difficult to apply.

Geomorphic approach

Where no usable streamflow or weather records are available, a geomorphicapproach can be used to estimate design peak flows. In this approach, achannel survey is conducted to determine bankfull channel geometry androughness. Manning’s equation is then used to estimate velocity and bankfulldischarge, which is often assumed to coincide with a return period of twoyears; however, in some landscapes the bankfull return period may be greater.

Generalized rainfall-runoff relations

A number of approaches have been developed to predict runoff responsefrom rainfall at a range of temporal resolutions. Many of these methods arepopular in engineering applications, and involve the use of tables ordiagrams to estimate parameters based on catchment characteristics such astopography and vegetation cover. For example, synthetic unit hydrographscan be used to estimate stormflow response during an individual stormevent. The rational method is commonly used to compute design floods incases where no usable streamflow records are available, but rainfall intensityhas been recorded at a weather station in or near the target catchment;however, because most precipitation gauges are in valley bottoms, measuredrainfall will typically underestimate actual rainfall over the catchment. Afundamental criticism of these rainfall-runoff relations is that peak flows inregions with cold winters typically occur as a result of spring-summersnowmelt or mid-winter rain-on-snow events and not simply rainfall.

281


Statistical modelling

Where a sufficient number of gauges are available, it may be possible toderive statistical relations between streamflow metrics or flow durationcurves and predictor variables based on catchment characteristics such asdrainage area, geology, land cover, and elevation. These relations arecommonly derived using techniques such as multiple regression. In addition,geostatistical methods such as kriging can be used either on their own or inconjunction with multiple regression (e.g. by using kriging to interpolateprediction errors to account for any spatial autocorrelation). In BritishColumbia, for example, Eaton et al. (2002) used geostatistical interpolationto map a “k factor” computed from gauged basins as k = Qma/A0.75, whereQma is the mean annual flood and A is the drainage area (km2). The k factorrepresents the mean annual flood for a catchment with a drainage area of1 km2. To compute the mean annual flood for an ungauged basin, the kfactor is extracted from the map and then multiplied by the drainage arearaised to the power 0.75. An advantage of these statistical methods is thatthey can provide estimates of the prediction error.

Process-based modelling

Data-based and field-based methods as described above cannot account forchanging climatic conditions or changes in land cover. For example, forestrecovery following the extensive tree mortality associated with the recentoutbreak of mountain pine beetle in western North America willfundamentally change the water balance of affected catchments over thecoming decades. In contrast, process-based models have the potential toaddress all of the weaknesses associated with currently used methods. Theirtemporal resolution can match the resolution of available forcing data, andmodels can, in principle, explicitly represent the effects of changes in landcover (e.g. Koboltschnik et al., 2007). Despite their potential advantages,process-based models are not routinely used in water resource analyses. Amajor challenge to the use of process-based models in PUB applications istheir need for input data. A complete suite of weather forcing data forphysically based simulation of melt and evapotranspiration would includeair temperature, precipitation, humidity, solar radiation, and wind speed. Atminimum, process-based models require air temperature and precipitationat daily or higher temporal resolution; humidity and solar radiation can beestimated from temperature and precipitation, if required (Walter et al.,

282


2005). In addition to weather forcing data, state variables such assnowcover and snowpack water equivalent can be valuable for modeldevelopment and testing.

Considerations in the choice of PUB method

Depending on the application, a range of prediction targets may be ofinterest. For example, broad scale screening assessments may require onlymean annual runoff. For the planning of large reservoirs, annual runoff andits interannual variability may be relevant. For many infrastructure designneeds, a flow extreme associated with a specific return interval may suffice,such as the 200-year flood or the 10-year 7-day low flow. In other cases,time series of discharge at daily or shorter time intervals may be required. In principle, a process-based hydrological model that runs at a daily or sub-daily time step, in combination with an appropriately long time series ofinput data, could be used to generate the full range of prediction targets. Inpractice, however, simpler methods that are less expensive to apply could beappropriate if their predictions were sufficiently accurate for the projectrequirements. As an example, consider mean annual runoff as a predictiontarget. It can be predicted using statistical relations with drainage areadefined using a regional monitoring network, which can generally provideestimates within an order of magnitude. While this approach is efficient, itdoes not necessarily provide estimates within an acceptable level ofuncertainty, nor does it provide information about seasonal patterns(Whitfield and Spence, 2011). In order to provide bounds for these estimatesand extreme values, we can use a combination of traditional knowledge(Woo et al., 2007), hydraulic geometry measurements (McNamara andKane, 2009), and paleo-records (Fortin and Lamoureux, 2009). Vegetationand animal species diversity regimes can also be indicators of floodplainextent and, in turn, of the extent and duration of extreme high flows.Statistical regression techniques can be robust (Lee and Ouarda, 2010), butif the regional monitoring network does not include physioclimaticallyand/or hydrologically similar gauging sites to those of the target catchment,the results may be dubious (Spence and Burke, 2008). An alternative is touse catchment classification indices (Quinton et al., 2003). Where time andfinancial resources permit, short term gauges can be installed in the basinsof interest and the resulting datasets can be used to develop relationshipswith long-term operational gauges.

283


In New Zealand, mean annual discharge is estimated from maps of meanannual precipitation and evapotranspiration. This estimate is checked forconsistency with nearby gauged catchments, and is then prorated to monthlyflows on the basis of maps of monthly flow proportions. Deterministic models are often used by electrical utilities for short-termforecasting (e.g. St. Hilaire et al., 2010) and to simulate streamflow underclimate change scenarios (Kattsov et al., 2007); however, they are notnecessarily the most feasible tool to determine long-term streamflow regimesbecause the length of climate data required to force them rarely exists.Stochastic weather generators (Srikanthan and McMahon, 2001) that mimicobserved or potential climate regimes have been used as an alternative.

Snowcover and snowpack water equivalent

The availability of snowcover information from the Moderate ResolutionImaging Spectrometer (MODIS) platform and other distributed snow productsare potentially valuable targets for model calibration. Finger et al. (2011) usedMODIS snowcover information along with streamflow to calibrate acatchment hydrology model. Boyle et al. (this volume) found that the use ofthe SNODAS product as the sole calibration target generated a parameter setthat also performed well for simulating streamflow. A weakness with MODISand other optical remote-sensing products is the effect of cloud cover, whichcan severely limit the completeness of snowcover scenes in mountain regions. Snow water equivalent is more difficult to sense than the snow extent.Natural gamma emissions measured from low flying aircraft can provideestimates of SWE, although the measurements can also be affected by thepresence of ice lenses or liquid water in the snowpack or underlying soil(WMO, 2008). In North America, NOAA conducts airborne gamma snowsurveys over many northern states as well as portions of Canada.Other passive measurements, generally made from satellites, are also usedto estimate SWE. These measurements may use a wide variety ofelectromagnetic frequencies, including microwaves. The condition of thesnowpack (crystal size, wetness) and blocking/shading due to vegetation ortopography can cause large errors in the magnitudes of the estimated SWE.

Evapotranspiration

Evapotranspiration is often only treated in terms of estimations and generalclasses. There are problems of availability of validation data, in particular at

284


different temporal scales, e.g. monthly versus daily. Methods are available forregionalization in space and time, e.g. via PRISM. It is possible to modelsolar radiation reliably for monthly data as an input to evapotranspiration, butcloud cover remains an obstacle at shorter time intervals. Sometimes this canbe overcome by using weather satellites. Alternatively, daily air temperaturerange can be used to model the effects of cloud cover on incident solarradiation at a daily time-step (e.g. Bristow and Campbell, 1984). It is alsoimportant to define actual versus potential evapotranspiration. Techniquessuch as the scintillometer are available for this. Global estimates ofevapotranspiration are available at the 1 km grid, for example from MODIS,AIRS, and CERES. Although this resolution is too coarse for complexterrain, it may be useful in semi-arid regions without too much variability invegetation or in any region without complex topography. The effects ofclimate change on evapotranspiration remain difficult to evaluate, inparticular when long-term projections on the variability of evapotranspirationare required. Projecting vapour pressure changes in the atmosphere remainsa substantial challenge.

20.6 COMPARISONS ACROSS LANDSCAPES

Process Understanding

The need for process understanding is common to all landscapes. Beforeselecting any conceptual model, a basic understanding of the importantprocesses that play a dominant role within the catchment is needed (Weilerand McDonnell, 2004; Abesser et al., 2008), along with knowledge of howthe dominant processes change between the seasons and how they varyspatially. Put simply, scientists and practitioners both need to understand thewater balance, in particular how water storages, fluxes, and pathways areaffected by landscape factors including:

• seasonality of precipitation and evaporative demand • groundwater • surface water impoundments• soil moisture and storage• wetlands• vegetation – interception, transpiration• other development, including urban areas

285


Data-rich regions

Spatially, the definition of “data-rich” depends on the density ofmeteorological and hydrological (surface) as well as groundwater(subsurface) stations. It also depends on the diversity of data types and thedegree of connection between different scales of data. Temporally, “data-rich” depends on the time-step between observations in relation to the timescale of the hydrological processes, and the length and resolution of datarecord. It is important to consider the applicability of data from data-richcatchments and whether the full range of available variables can beexploited in models, including the following:

• weather stations at high and low elevations• streamflow• topography and land cover • soil information• glacier mass balance• snowpack SWE

We need to consider that developing understanding in the application ofconceptual or process-based models in a data-rich situation shouldcontribute heavily to predictions in ungauged basins where data are notavailable. While practitioners in data-rich areas may face less uncertaintythan those in other areas, they have the opportunity to document the limitsof statistical, conceptual, and process models and thus can potentiallycontribute to quantifying uncertainty in the data-sparse and data-poorcontexts. Restated, models developed in data-rich areas need to be fullyexplored and exploited before applying them to areas where supporting dataare not as available. In all landscapes, data-rich areas are primarily associated with experimentalwatersheds or long-term ecological research areas. These research sites aregenerally small in area (i.e. less than 100 km2) but frequently have data forlong periods of time for many relevant variables.

Data-sparse regions

Any definition of “data-sparse” depends on the type of landscape beingconsidered, e.g. homogenous plains vs. highly variable mountain terrain; italso depends on the representativeness of the scale of variability. Spatially,

286


“data-sparse” was considered to include only one or two observationlocations within a basin, thereby excluding the ability to resolve spatialvariations. Temporally, in “data-sparse” areas, observations cover only shortperiods of time (i.e. one or two years) or are inhomogeneous data serieswhere data overlap only during relatively brief periods. Measuring stationsare spaced at intervals of at least 400 km or are situated out of the basin insimilar regions. Only limited variables are available and the data areinsufficient for both calibration and validation of predictive models. Thedata-sparse classification could also apply to basins that are covered byremote sensing data, but no monitoring stations to ground truth these data. In data-sparse regions, complex questions cannot be answered usingavailable data and current PUB methods. Management decisions, however,do not always require complicated models and equations; rather they rely onconfident predictions. Scientists are often asked how to reduce uncertaintyof forecasts or predictions. This is frequently a question of the amount oftime or money that can be invested with relation to actual improvement. Thereduction of uncertainty also depends on how to determine the best locationfor answering the question. Managers in these situations need to makedecisions that consider or tolerate irreducible uncertainty. From thisperspective, social adaptation or risk based decision making may be a betteralternative. Understanding and modelling in data-sparse areas may be difficult tovalidate as fully as might be possible in data-rich situations. Here thepractitioner needs to accept being dependent upon a combination of goodjudgement in selecting models and reasonable scales (temporal and spatial)and in choosing how to use the available data, either in calibration orvalidation. At the same time, a similar level of judgement needs to beapplied to interpretation of the results. The quality of predictions in data-sparse applications is necessarily of lower confidence and resolution than indata-rich applications. It would be useful to develop a nomograph or indextable that would assist the practitioner in this effort; such an approach mightrelate the hydrological attribute to the available data. Simply put, the annualmean flow can be predicted in a data-rich area with more precision than ina data-sparse area; monthly or daily flows that could be predicted with someaccuracy in a data-rich application might be unreasonable to generate whereinsufficient data exists.

287


Data-sparse areas are often the result of the uncoordinated interests ofresource and management agencies. Frequently, meteorological and climatenetworks were located in relation to settlements and to aviation facilities.Water networks, specifically for water levels and streamflows, developed inrelation to resource development (railways, hydropower) and concerns overflooding (floodplain settlements). Soils and soil moisture were observedwithin agricultural and forestry agencies. The net result is that in most data-sparse landscapes the lack of coordination and exchange is a factor in theavailability and/or complementarity of data.

Data-poor regions

While data-poor regions, by definition, lack sufficient ground-basedinformation for hydrological prediction, there is potentially a wealth ofusable information from remote-sensing platforms (particularly in the formof digital elevation models and maps of land use and land cover) and outputfrom global and regional climate models. The main obstacles to providingthis information to users, beyond lack of data availability, are restrictions indata processing capacity and the lack of graphical interfaces, user manuals,and documentation. For example, remote sensing data such as MODIS arewidely available but often lack an interface to help with transferring it intoGIS for use by practitioners. The North American Regional Reanalysis(NARR) product provides spatially distributed weather information at dailyor sub-daily time scales, but is difficult for practitioners to use since itrequires an interface or specialized programming skills to access andmanipulate. One reason for the difficulty is that gridded data are generallyorganized as temporal snapshots and consequently require the user todownload, disaggregate, and combine many large individual files to producetime series for specific locations or catchments.It is an open question whether the practitioners or the scientists should beresponsible for developing easy-to-use tools for accessing and processinggridded data sets. Scientists generally do not feel responsible for translatingtheir research results into practical tools or lack the motivation to do so as theirreward system is based on scientific publications, not practical application; buton the other hand, practitioners often do not have the money or time to investin developing tools for accessing and processing the available information.There is a clear need for someone to operate at the interface between scientistsand end-users, e.g. from technical colleges and governments.

288


Data-poor areas exist within all landscape types but are particularlyprevalent in arctic, arid, and semi-arid regions, which are also commonlywater-poor unpopulated regions. This leads to a reliance on predictions inungauged basins in these water sensitive landscapes, which are increasinglythe focus of resource development. Lack of information and data leads togreater uncertainty for any predictions made for these sensitive regions.

Differences among landscapes

Two factors strongly affect how much data are available in a given country:the density of population and the economic status. There are generally moredata available in populated areas of developed countries than in unpopulatedareas or developing countries. Within any of these contexts, data are likely tobe more available in areas where water issues such as flooding and droughtare recurrent. There may also be a cultural bias, as data seem to be moreavailable where there has been an influence from the historical situations;British Empire military installations, for example, often provided detailedobservations of weather and climate. The distribution of data networks is alsoaffected by economic interests related to potential development, such as windpower and hydropower, and potential risk, such as flood prone areas. In manyareas there are more monitoring data available where there are more peopleand water than in areas where there are fewer people and less water.

20.7 BARRIERS BETWEEN RESEARCH AND PRACTICE

The break out group discussions illustrated a divide that presently exists:researchers are focused on methods that are dynamic, detailed, information-rich, and rely on extensive observations; practitioners are using simple user-friendly methods that can be applied in areas with sparse monitoring, and thatsometimes involve soft data or subjective judgement. The time scales of thepredictions are also often different. The scale of hydrological predictionsrequired ranges from simple annual values to seasonal, monthly, and finertime scales and also includes extreme events. The types of predictions thatare needed or expected also vary from water rich to water poor areas. Thevariables to be predicted can be fluxes, states, or storages.Researchers continue to contribute to all areas needed to improvepredictions in ungauged basins, including process understanding anddevelopment and testing of empirical, statistical, conceptual, and processmodels. These developments continue, not always linearly, but are highly

289


influenced by observation technology and computer capability and capacity.While these tools have promise for improving predictions for ungaugedbasins, their uptake by practitioners does not automatically follow.Practitioners need to be provided with modelling tools and once researchershave developed effective tools, there needs to be uptake by the usercommunity. Users need to be provided with the tools and required dataproducts in a manner that minimizes their investment of time and resources.In research, data sources such as gridded fields of driving variables can beaccessed via the internet, sometimes along with software tools for extractingthe information required in a specific application and exporting it to a formatrequired by a model. Such tools could be better suited for use by practitionersin the form of a graphical user interface or a simple scripting language;however, practitioners are not in a position to adopt tools that requireintensive relearning and technical support, while the research community isunlikely to put a high priority on developing such tools or to provide trainingto the practitioner community. One existing solution would be researchpartnerships, possibly funded by regulatory or government agencies, focusedon technology transfer that supports implementation of research anddevelopment, and on providing a level of training commensurate with theneed for practitioners to use the appropriate tools.An important issue with obtaining data from the internet and agencydatabases is fitness for purpose (Whitfield, 2012). Data which are availablemay, or may not, be suitable for supporting predictions in ungauged basins.Practitioners need tools and guidance that support them in making effectiveand appropriate decisions about the use of such data in their application. Oneconsequence of data becoming widely available on the internet has been adecline in the professional guidance available to ensure the user understandsthe nature of the data. Tools that better communicate metadata and inform theuser of data quality and its representativeness are needed.The adoption of new approaches, including process-based models, isgenerally constrained by cost and aversion to change: clients are typicallyunwilling to invest in new tools when simpler methods are acceptable orrequired by regulators. There is a need to weigh the lower cost of analysisassociated with simple methods against their risk of failure. It is conceivablethat, in some cases, the risk of an erroneous analysis may be sufficiently lowthat there is no economic incentive to pursue application of model-basedapproaches. In other cases, there may be clear economic arguments in favour

290


of model-based analyses; however, it is expected that everyone would rely ongetting the right answer for the right reasons. Other barriers expressed bypractitioners include the following:

• Lack of awareness of emerging approaches and modeldevelopments. Some of these are related to the cost of accessingjournal articles, as well as limited time to devote to professionaldevelopment activities, such as attending conferences

• Cost of purchasing another model and the cost of retraining• Aversion to the risk of investing time in learning an approach that

may not be widely accepted, especially to regulators• Familiarity with existing models and tools and project timelines that

are too tight to allow for alternatives to be implemented

One possible approach that was identified is not to build a model and thencustomize data inputs and model outputs to it. Rather the focus should be todevelop a platform where data and output handling is conducted through astandard interface. Models should be supported by complete documentationand either a graphical user interface and/or a straightforward scriptinglanguage to facilitate training and application. An example of a relativelyeasy-to-use platform is Green Kenue. Green Kenue™ (formerly EnSimHydrologic) is an advanced data preparation, analysis, and visualization toolfor hydrological modellers. It provides a platform that integratesenvironmental databases and geospatial data with model input and output.Green Kenue provides pre- and post-processing for the WATFLOOD andHBV-EC hydrological models. It can be downloaded without cost at:http://www.nrc-cnrc.gc.ca/eng/solutions/advisory/.Potential barriers to the adoption and development of these resources bymodel developers include the following:

• Lack of skill in software applications implementation• Lack of awareness of user environment and needs• Lack of forward planning in model development that ensures

linkages with existing tools• Competition and aversion to risk in participating in applied research

which may not be recognized in academic/research careerpromotions

• Lack of a reward system for researchers, thus no motivation to makemodels user-friendly or to offer training

291


Potential barriers to the adoption of these resources by practitioners includethe following:

• Lack of awareness of the existence of the tools• Inertia, particularly when considering new tools• Dedication to familiar tools, particularly those used during

university training, or commonly used for other applications(e.g., Excel, MATLAB, ArcView)

• Vendor lock-in, where the customer is dependent on a commercialprogram, and switching to another program incurs costs

• Lack of technical support in dealing with onerous data requirements,data handling, and requirements for specific programming skills.

20.8 RESEARCH NEEDS AND EMERGING METHODS

Many research and application needs and opportunities were identified inthe workgroup discussions. In this section those which were common toseveral groups are described.

Catchment characterization

One way to strategically choose and collect transferable data is thedevelopment of a classification system for watersheds. Catchmentclassification has been advocated within the Predictions in Ungauged Basinsinitiative for some time (McDonnell and Woods, 2004). While there have beenstreamflow regime classifications for some regions (e.g. Church, 1974), ageneral catchment classification system has not yet been developed. Whilebasin-scale classification systems exist for particular regions, a general systemshould be based on physioclimatic characteristics to enable the prediction ofthe streamflow regime and other hydrological behaviour (Wagener et al.,2007). Basin classification can be based upon several traits, includingtopography (e.g. relief), hydraulic geometry (e.g. channel morphology),vegetation (e.g. NDVI distribution), or response units (e.g. hydrologicalfunction and distribution). Several classification schemes may need to bedeveloped and/or combined so that the most appropriate transfer mechanismsare available for individual indices, parameters, or indicators. While one sizemay not fit all, it will be important to avoid applying an existing regionalclassification too widely. This is a classic hydrology trap; develop locally andthen apply globally.

292


An approach would be to identify several representative basins and developappropriate classification schemes. Building from that, developing and testingindex, parameter, or indicator transferability across a range of space and timescales is needed. This process must incorporate a feedback mechanism toimprove the design and implementation of research basins and the activitiestherein. This will also assist in guiding research and monitoring efforts thatcan continue to support the development of transferable data and information.In every region where data availability is at a premium, parsimony, in bothdata collection and modelling, must remain a key consideration.

Catchment characterization will be an important tool in transferringparameterizations. Characterization schemes will need to account for climaticthermal and moisture regimes (particularly the seasonality of precipitation),land cover, topographic complexity, and geology. Tague and Grant (2009)illustrated the profound influence that the underlying geology can have onstreamflow variability and the response to climate variability. Geological mapsare generalized, and it may be difficult to translate geological information frommaps into hydrologically relevant parameters related to storage and transportdynamics. Research that extends the work of Tague and Grant (2009) to a widerange of geological and hydroclimatic contexts is needed.

For the foreseeable future, semi-distributed models will likely dominateover fully distributed models due to their lower computational demands.Current semi-distributed models use either a Grouped Response Unit (GRU)or a Hydrological Response Unit (HRU) approach.

In the GRU approach, a catchment is normally represented using griddedmaps of various boundary conditions, including elevation, slope, aspect(derived from the elevation grid), underlying geology and/or soil types, andland cover (e.g., forest/open/water/glacier). Individual grid cells arecategorized and cells with similar characteristics are grouped; heat andwater fluxes are then modelled for each GRU rather than each individualcell. The delineation of cells is presently constrained, in part, by theresolution of digital maps. Digital elevation models are increasinglyavailable with grid resolutions less than 100 m. Land cover maps are alsoavailable at increasing resolution. Characterizing land cover, particularlyaccurate land use information, is difficult. Rapid and widespread land use orland cover changes such as agriculture, irrigation, or forest disturbance (e.g.Mountain Pine Beetle) add a considerable challenge.

293


An important difference between HRUs and GRUs is that GRUs do notnormally incorporate any information on lateral interactions between gridcells. On the other hand, HRUs are defined in consideration of their relativeposition within the cascade of lateral water transfers by, e.g., elevation,blowing snow, and subsurface flow, in addition to the types of criteria usedto define GRUs. Delineation of HRUs is normally guided by land cover andgeomorphology that guide the modeller’s understanding of the dynamics ofstorage, connectivity, and thresholds in a landscape. The delineation processis subjective: two modellers are likely to generate different catchmentdescriptions, which contributes yet another dimension to the issue ofequifinality. Clearly, there is scope for research to further develop the understanding ofthe consequences of the HRU/GRU strategies and alternative hybridapproaches. Does the routing present in the HRU approach realisticallycapture the flux of water in the watershed? Does this perform significantlybetter that a GRU approach at some space and time scales? For example,would the GRU approach be suitable for annual fluxes and the HRU forfiner time steps? Are there alternative approaches where the routing can bebetter captured within a GRU type of approach? Can the definitions ofHRUs be made to be more objective, or even automated?

Processes and parameterizations

In principle, gridded models can be transferred in time or space if theyincorporate the appropriate physics. The predominant challenge withcomplex models is that their numerical solution schemes can be costly interms of computer processing time. To address physical processes that areunresolved at the model scale, processes are parameterized; however,highly parameterized models can be sensitive to errors in input variables.Simpler highly parameterized process-based models require fewercomputer resources than detailed processed models; however, robustparameterizations of the processes that facilitate the transferability ofparameter sets must be developed. A globally applicable model may not beappropriate. Instead, the development of modular modelling platforms suchas Cold Regions Hydrological Model (Pomeroy et al. 2007)) and Raven(Craig et al., 2011) may be a more appropriate and useful approach. Suchmodels would incorporate only those processes and representations that arerelevant and limited by available data sources. In such a modelling

294


environment the onus is on the hydrologist to understand the dominantprocesses in the target catchment to ensure they are incorporated in thechosen model.

Calibration or parameter estimation will likely always be required inmodelling applications, either to estimate parameters in processrepresentations or to correct biases in fields of driving variables. Theultimate objective is to develop parameterizations that are transferable intime and space and that do not require site-specific calibration. This wouldbe consistent with the trend for modellers to move away from usingstreamflow as the sole calibration target and incorporating additionalvariables such as glacier mass balance (e.g., Konz and Seibert, 2010;Schaefli and Huss, 2011) and other “soft” data (Seibert and McDonnell,2002). Bias-correcting meteorological fields derived from products such asNARR or MERRA could be approached by using them to drive a process-based snow model and forcing it to reproduce snowcover patterns asderived, e.g. from the MODIS products.

Research opportunities in process-based modelling are extensive. Robustand consistent methods for assessing the suitability of physically basedmodels or parameterizations in model performance for site-specific fieldstudies and for supporting scaling will continue to be a focus (Wagener andWheater, 2006). Research tools are needed to support communicatinguncertainty. New approaches for allocating model resources and supportingappropriate decision making in model implementation (i.e. should theprocess be in the model physics or simply a parameterization) are needed toguide which approach is more suitable.

Scale considerations

Because data products are available over many differing spatial resolutions,converting them to a common scale requires degrading high resolution datato a coarser scale and/or interpolating coarser resolution products to higherresolutions. Degrading the resolution of a more finely resolved product to acoarser scale destroys information. Interpolation of coarse-scale data carriesthe risk of introducing artifacts of the interpolation procedure and may notaccurately represent the true spatial pattern. Research that assesses the trade-offs in these approaches and their influence on model performance is neededas either approach may introduce additional uncertainties and bias.

295


Assessment of model output

Assessment of the uncertainty of model results is essential for understandingthe applicability of the output to water management and the risks associatedwith making decisions with uncertain data (Pappenberger and Beven, 2006).An assessment of the potential to reduce uncertainty is also valuable indetermining where additional resources may be best spent.

Practitioners and researchers would be well served by adopting a commonframework for model assessment. As seen previously, modellers and analystsmake assumptions, parameterize, and simplify as a matter of course, oftenwithout explicit explanation or justification. A standard framework that allowscommon interpretation of these decisions would be widely useful. Sinceeveryone must make assumptions, simplifications, and parameterizations, theuncertainty induced by these simplifications needs to be adequatelycommunicated. As a hypothetical example, a modeller might argue that, sinceglaciers are less than 5% of the given basin’s area, glacier-related processesare not modelled; this choice might induce an uncertainty of < 5% in theannual runoff, but can produce a major error in predictions of late summerdischarge, particularly during hot, dry weather.

Communicating uncertainty is a core issue for predicting in ungauged basins.Frameworks such as GLUE (Generalized Likelihood Uncertainty Estimation)(Beven and Freer, 2001) provide guidance for scientists, practitioners, anddecision makers who need both tools and training that support their use of thisinformation. Too often, decision makers fail to interpret uncertainty in termsof confidence intervals; rather they perceive it as lack of knowledge. At thesame time, a single number may be perceived as “better” than numbers withconfidence intervals or as the “right” answer. Client expectations may resultin practitioners simplifying results (e.g. removing confidence intervals)resulting in information loss.

Research basins

Given that many areas are generally poorly monitored, research basins will becritical in the development and testing of simplified but robust representationsof processes, for determining the appropriate scales for process representation,and for testing alternative approaches to the definition of HRU/GRUs. Withincreasing pressures on hydrometric network managers to reduce costs, it iscrucial to maintain the data-rich infrastructure at research sites, especially at

296


sites with relatively long periods of record, where the effects of climaticvariability can be separated from other impacts and can be more readilyassessed. Reference hydrological networks play a key role in this (Whitfieldet al., 2012; Burn et al., 2012).

Putting PUB advances into wide practice will require research to progresson a number of academic fronts, especially (1) generation of spatial fields ofmeteorological variables, (2) characterization of catchments and delineationof HRUs so as to capture the range of hydrological behaviour in aparsimonious yet robust manner, and (3) development of approaches fortransferring information about processes and parameterizations from data-rich research catchments to ungauged catchments. Researchers shouldengage in a series of PUB emulation exercises to demonstrate theimprovement in predictive capability asociated with newly developedmodelling approaches. The following basic steps incorporate the scientificprinciples associated with predicting flows in ungauged basins, but alsorecognize the limitations that exist in practice:

• basic process understanding • data assessment and compilation• model selection based on processes and data• model parameterization• assessment of model output

An important consideration is that not all applications require or provide thesame level of accuracy; accordingly, there may be a demand for a range ofmodelling approaches to suit the needs and data availability in specificapplications. Researchers will need to continue to engage with practitionersin workshops like “Putting PUB into Practice” to gain a better understandingof their needs.

Dealing with change

Climate change is generally being approached by variability and trend studiesof climate and streamflow, but largely tied to temperature and precipitation.Data from research basins and reference hydrological networks can be usedto define ‘natural’ types and their attributes. Presently, we are not in aposition to provide long term projections without making large simplifyingassumptions such as lack of landscape or vegetative change. Projecting future

297


streamflows is a complex process and there are multiple GCMs, RCMs, andmany equally likely future scenarios that could be considered. Hydrologymodels (statistical, conceptual, and physical) could all be driven by climatemodel outputs, possibly downscaled (dynamically or statistically), for anynumber of possible cases. Research in data-rich areas will need to provideinsight to the direction and scale of potential changes in order for there to beconfidence in the outcomes. PUB in data-sparse and data-poor situations willneed to develop much further before the methods can be considered adequatefor projections that address changes in climate, land use, and vegetation.

20.9 KEY OPPORTUNITIES

Standardized protocols

The hydrological community must recognize and adopt standard protocols orbest management practices for both data collection and data extrapolationacross all regions for hydrological prediction. This would help reduceuncertainty in decision making and data transferability during the toolevaluation process. For instance, the Canadian oil and gas industry iscurrently developing standard protocols for stream gauging in collaborationwith the British Columbia Ministry of the Environment and the Water Surveyof Canada. This example of co-operation illustrates the objective to ensureinter-industry comparability in order to facilitate the more widespread use ofdata already being collected and, thereby, assist the development of morerobust models of water availability for management and allocation purposes.These types of activities help improve the appetite for non-hydrometricservice data and encourage a two-way transfer of information on how tooptimize data collection. In addition, inherent benefits of this multi-agencyapproach are improved data transferability and reduced redundancy amongdifferent groups. Instead, time and finances can be applied to otherknowledge gaps. Important results of the above, of course, are more cost-effective data collection and hydrological model development protocols.

Protocol for a catchment function diagnostic

When faced with a diversity of choices and an even greater range of potentialoutcomes, a tool of increasing popularity is the decision tree (Bosch et al.,1996). This is a simple decision support tool that uses a tree-like graph ormodel of decisions and their possible outcomes, and it can help in the designof a strategy most likely to aid in meeting a specific goal. The decision tree

298


could be used to identify the most suitable approach to hydrological predictiongiven the parameters of a particular situation. Figure 20.1 illustrates a possibledecision tree based on the methods summarized above, that includes thespatial and temporal nature of the question being posed, the level of acceptableuncertainty, and the time and financial constraints for use by hydrologicalpractitioners. This was envisaged as a decision tree approach that would useavailable information, assess process understanding, direct data assessmentand compilation, and guide model selection and model parameterization.Basic process understanding could be determined from available informationon climate, topography, land use, regionally generated streamflow predictions,and experience in data-rich contexts. This information could be incorporatedin the decision tree to identify and differentiate processes and their linkages,and would be needed to develop tools, guidelines, and thresholds. The processmust not be a single entry key; rather it should [1] identify the main processesand pathways including groundwater and landscape storages, [2] identify thelandforms and topography, the existence and extent of wetlands and floodplains, slopes, and drainages, [3] identify the vegetation, soils, includinginterception and evaporation, and [4] address the heterogeneity of the mosaic.Ultimately, any classification needs to be accessible based upon the availableinformation.

299


Basic process

understanding

Assessment of

model output

and uncertainties

Data

compilation

Model selection

&

parameterization

Figure 20.1 Schematic showing the flow of process-based hydrologic modelling underuncertainty.

A well-constructed decision tree available to both academics andpractitioners, indicating both traditional and ground-breaking methodologies,may provide insight into which methods could and should be implemented.We recommend that a decision tree for prediction in ungauged basins shouldbe constructed and made publicly available as it would be a valuable meansof both knowledge and technology transfer, and it would serve as a‘handbook on a page’.

Better outreach

Developing tools like decision trees or comprehensive handbooks (e.g. Pikeet al., 2010) depends on good and ongoing communication betweenresearchers and practicing hydrologists. Thematic meetings and workshopshave proven to be very successful for information and idea exchange inCanada (Spence et al. 2005, Spence et al., 2008) as they bring together adiversity of attendees, increase awareness and conversation, and developtrust, the value of which cannot be underestimated.In most data-sparse and data-poor regions, academics, practitioners, andgovernments must work together in data collection, research, and thedevelopment of predictive tools. Unfortunately, a good communicationstrategy to advertise the availability of new data or technologies to thepracticing water resource community is generally lacking. A website thatprovides a conduit for research and development notes would be widelyvaluable. Such a website would include required metadata that wouldaccompany the tools and would include details of the technical workingsand the broader relevance for hydrological prediction. This would assistboth technicians and managers in appreciating the value of the information.Active and up-to-date, online resources used as tools for the disseminationof valuable information to the hydrological community would support thetraditional peer-reviewed literature medium for research results by acting asthe key outlet for up-to-date development results. Furthermore, on-linesources support sustainable linkages among academia, government,practitioners, and the public, each of whom have a stake in the developmentand understanding of water resources in ungauged remote areas. Hosting ofsuch a site presents several challenges as there is no single entity that hasthat mandate. A shared approach where individual agencies share theirinformation through a single portal may be a workable system. Most likely,success would rely on an open source approach.

300


Open source solutions

Much of the development of improved tools to date has been top-down, withthe concerns of users often an afterthought. To make the uptake of improvedtools successful, not only do the tools need to be useful to practitioners butpractitioners need to be informed about the existence of new tools, thenature of the improvements, and that the resulting tools implement thesechanges. An advantage of open source methodologies is that they potentiallyincrease the number of people developing the project and those testing thecode, contributing to improvements in development speed and codereliability. The best way to ensure the engagement of users is to involvethem in the development of tools, preferably from an early stage in thedevelopment. In open source projects, this can be done by soliciting theparticipation of end users at the beginning and throughout the project. In allcases there needs to be extensive collaboration between researchers and endusers. Stakeholders can be involved in research projects and in the trainingof students. While financial contributions from the users give them a stakein a project, it is critical to keep expectations clear.

One opportunity is to develop an open-access database of standardizedwatershed variables which is routinely updated. Standardized open sourceclustering of the database records would be updated and ‘published’ on atimely basis perhaps based upon growth in the size of the database. At eachiteration, new records would be tested to see if they “fit” the classification(high similarity); if they do not (low similarity), a rule based new generationwould be generated, reviewed, and released as a new version by a peer team.Each new version would require a description to be published in a publicforum such as a peer-reviewed international journal. The editorial board ofthe journal would need to be approached to create an ongoing relationship.It is expected that the classification and the diagnostic should co-evolve, butbe done separately.

Since every watershed is unique, the classification needs to simultaneouslydeal with the common attributes upon which similar hydrological landscapesare grouped at a high level while allowing more separation based uponadditional attributes. This should provide a system where the major processesand timing are captured for any ungauged basin, but with the additionalinformation available to increase resolution.

301


It is recognized that all new tools need to better incorporate user/practitionerneeds. Projects also need to include better communication plans and training ofend users. Practitioners make it clear that the initial adoption of a newtechnology has both a risk and a cost; hydrology will be well served if newdevelopments in the science are incorporated into a limited number of tools. Alldevelopments being implemented for professional practice need to beaccompanied by training which could include webinars, podcasts, and otherforums where the needs of the users, and not the developers, are best addressed.

20.10 CONCLUDING REMARKS

The lack of data with which to inform any type of predictive model, incombination with the wide diversity of hydrological landscapes, makeprediction in ungauged basins challenging. The PUB Decade has seen thedevelopment of research that has a great potential to advance the practice ofhydrological prediction in ungauged basins, particularly thanks to thedevelopment of gridded hydrometeorological products and researchactivities in relatively data-rich research basins. Support for research basinsneeds to continue as these basins provide the testing grounds for newhypotheses, statistical, conceptual, and deterministic models, and reanalysistools. By clearly determining the scales at which the data and informationproduced would be applicable, possibly through a basin classificationsystem, the value of these sites would be enhanced. Practitioners andmanagers also need such a classification system and other tools designed toenhance the development of transferable data, indices, parameters, andindicators. It is recommended that standardized and generalizedphysiographic information be collected using the same set of tools that arewidely used by practicing hydrologists. The classification system shouldlink landscape attributes to these more intensive and detailed measurements.As predictive tools develop, updated decision trees may prove to be valuableto practicing hydrologists. Development and maintenance of these types oftools require ongoing communication and collaboration among allhydrologists. The existing newsletters, journals, and websites ofprofessional and learned societies are well suited for the spreading of suchinformation and would complement the traditional peer-reviewed literatureconduit for information dissemination. An open source approach to theclassification system and the diagnostic protocol is recommended so thesewill have widespread application and acceptance.

302


20.11 RECOMMENDATIONS AND ISSUES

1. Address the need to maintain continuity. What would be theappropriate mechanism for validating new methods, and bringingthem into practice? It is not obvious who should be responsible forthe ongoing need to convert research into practice. New approachesand improved methods deliberately result from research; movingsuch new developments into professional practice is an apparent gapin our community, although at different times government,practitioners, and software developers have played important roles.Some guidance with respect to this has been provided above.

2. Address the need for interfaces to complex datasets. How can complexdatasets be made widely accessible to practitioners? Researcherscommonly use a diverse mixture of complex datasets from remotesensing, climate models, and reanalysis products. While these datasetsare readily accessible, they are not readily put into practice, and if theyare used in practice, the access methodology is not always clear.

3. Address the need for open source approaches. Are we getting theright answers for the right reasons? With increasing complexity ofmodels and analysis there is a responsibility for transparency. Thecommunity needs to be certain that models and methods properlycapture the science, the uncertainty, and the quality of inputs. This isparticularly true for regulatory agencies that make key decisionsbased upon outputs from models and analysis.

4. Address the need for common operating platforms. Can thecommunity adopt a common framework that would improve uponthe current situation? With the increasing diversity of models andtools there is a growing risk that the users will use tools that theyhave available, that they are familiar with, or that cost less, ratherthan ones where the science matches the information needs. Thegrowing gap needs to be made much smaller. Modellingenvironments such as Green Kenue have demonstrated the value ofcreating a common framework that supports model development,hydrological modelling, and the display of results. Green Kenue wasrecently made freely available with the intent of it being used intraining students in the practice of hydrological modelling.

303


5. Address the need for better outreach. Is the community taking thecorrect actions to support the transfer of science into practice?Adopting an open source approach with a generalized classificationand diagnostic scheme are key steps; making training and supportreadily available to practitioners will also be important. Tools thatare unsupported or left up to the users to discover and apply will beunsuccessful. Practitioners often prefer tools where support isavailable on demand, or just in time, rather than “do it yourself”.

The legacy of the PUB decade includes significant advances in theunderstanding of hydrological processes and development and testing, inresearch settings, of revised or new methods for PUB. The challenge remainsto address the need to adopt standards and globally generalized approachesfor practitioners to make predictions in ungauged basins; the participants inthe workshop portion of this meeting have suggested approaches that willaddress this situation.

20.12 ACKNOWLEDGEMENTS

The authors of this chapter wish to thank all the participants for theircontributions and comments upon which this summary was constructed. Inparticular, the written summaries provided by session chairs and rapporteursare appreciated.

304


summary and synthesis of workshop …...summary and synthesis of workshop break out group...

Documents