sustainability designing river flows to improve food...

RESEARCH ARTICLE SUMMARY

SUSTAINABILITY

Designing river flows to improve foodsecurity futures in the LowerMekong BasinJ L Sabo A Ruhi G W Holtgrieve V Elliott M E Arias Peng Bun NgorT A Raumlsaumlnen So Nam

INTRODUCTION TheMekong River providesrenewable energy and food security for a popu-lation of more than 60 million people in sixcountries ChinaMyanmar LaoPDR ThailandVietnam and Cambodia Seasonal rains floodthe riverrsquos floodplain and delta This flood pulsefuels what is likely the worldrsquos largest fresh-water fishery in Cambodiarsquos Tonle Sap Lakewith gt2 million tonnes of annual harvest val-ued at ~$2 billion Hydropower developmentis crucial to the regionrsquos economic prosperityand is simultaneously a threat to fisheries andagriculture that thrived in the natural-flow re-gime The Mekong is testament to the foodenergy and water challenges facing tropicalrivers globally

RATIONALEWehypothesized that high fish-eries yields are driven by measurable attributesof hydrologic variability and that these relation-ships can be used to design and implement fu-ture flow regimes that improve fisheries yieldthrough control of impending hydropower op-erations Hydrologic attributes that drive strongfisheries yields were identified using a data-driven approach that combined 17 years of dis-charge and standardized harvest data withseveral time-series methods in the frequencyand time domains We then analyzed century-scale timeseries ofdischargedataon theMekongand associated hydroclimate data sets to under-standhowcurrentdams independent of climatehave changed key drivers of the fishery since theearly 1960s Finally we used estimated hydro-logic drivers of the historical bag net or ldquoDairdquofishery on the Tonle SapRivermdashthe largest com-mercial fishery in theMekongmdashto design betterfisheries futures by comparing designed flows tocurrent and pre-dam (natural-flow) regimes

RESULTS Our analysis identified several fea-tures of hydrologic variability that portendstrong fisheries yield These include two ldquohigh-levelrdquo descriptors flood pulse extent (FPExt)and net annual anomaly (NAA) FPExt whichcombines flood magnitude and duration haslong been hypothesized to drive fisheries yield

in ecosystems subject to flood pulses such asthe Mekong NAA is the annual sum of dailyresidual flows standardized to the long-termaverage hydrograph Hence NAA is a compactmeasure of hydrologic variance and can be fur-therdecomposed intonine shape ldquocomponentsrdquo

Several of these components drive high fisheriesyields including a long low-flow period followedby a short strong flood pulse with multiplepeaks All essential drivers of the flood pulsefishery have been changing since the closureof the first Mekong tributary dam and are in-dependent of changes associated with climateobserved over the past century The directionof these changes is consistent with decliningfisheries yield in the Tonle Sap Projection ofthe fishery driven by a hypothetical ldquodesignerrdquohydrograph capturing the key shape featuresassociated with strong yield improved harvestrelative to current conditions yield was pro-

jected to exceed that of the natural-flow regimeby a factor of 37 This result was robust to theinclusion of density-dependent recruitment inour time-series model

CONCLUSION A data-driven approach re-veals a new perspective on hydrologic driversof fishery productivity in the Mekong The ex-tent of the flood pulse is paramount as pre-

vious literature suggestsbut so are other descrip-torsofhydrologicvariationincluding anomalous lowflows Variance is keymdashspecifically the sequenceand timing ofwithin-year

anomalous high and low flows A focus on var-iance shifts the conversation from ldquoHow muchwater do we needrdquo to ldquoWhen do we need it themost and when can we spare itrdquo Beneficialcomponents of variance in the hydrograph canbe described by a simple Fourier seriesmdashanasymmetric rectangular pulse train A quan-

titative ecological objective function fills a criti-cal gap in the balancing of fisheries harvestwith other important objective functions in-cludinghydropower generation riceproductionand transportation This opens the possibilityof specifying and implementing flow regimesto manage rivers to satisfice trade-offs betweenfishery productivity and other ecosystem ser-vices provided by tropical rivers subject to floodpulses

RESEARCH

Sabo et al Science 358 1270 (2017) 8 December 2017 1 of 1

The list of author affiliations is available in the full article onlineCorresponding author Email johnlsaboasueduCite this article as J L Sabo et al Science 358 eaao1053(2017) DOI 101126scienceaao1053

Subsistence fisher tending nets on the Tonle Sap Lake CambodiaPHOTO

BYJ

LSABO

ON OUR WEBSITE

Read the full articleat httpdxdoiorg101126scienceaao1053

on August 20 2020

httpsciencesciencem

agorgD

ownloaded from

RESEARCH ARTICLE

SUSTAINABILITY

Designing river flows to improvefood security futures in theLower Mekong BasinJ L Sabo123 A Ruhi24 G W Holtgrieve5 V Elliott6 M E Arias7

Peng Bun Ngor8dagger T A Raumlsaumlnen9 So Nam8Dagger

Rivers provide unrivaled opportunity for clean energy via hydropower but little is knownabout the potential impact of dam-building on the food security these rivers provide Intropical rivers rainfall drives a periodic flood pulse fueling fish production and deliveringnutrition to more than 150 million people worldwide Hydropower will modulate this floodpulse thereby threatening food security We identified variance components of theMekong River flood pulse that predict yield in one of the largest freshwater fisheries inthe world We used these variance components to design an algorithm for a managedhydrograph to explore future yields This algorithm mimics attributes of dischargevariance that drive fishery yield prolonged low flows followed by a short flood pulseDesigned flows increased yield by a factor of 37 relative to historical hydrologyManaging desired components of discharge variance will lead to greater efficiency in theLower Mekong Basin food system

Hydropower is currently the dominant sourceof renewable energy worldwide (1ndash4) andis poised to power some of the poorest pre-dominantly rural populations in the world(5 6) Storage of streamflow in reservoirs

and dam operations often dampen peak flowsdeliver higher base flows and change the rangeand frequency of discharge variability in riversworldwide (7ndash11) This hydrologic alterationmdashfrom dams and other stressorsmdashpromotes theinvasion of non-native aquatic species (12 13)alters food web structure (14) and reduces bio-diversity by homogenizing regional freshwaterfaunas (10) Traditional river conservation underthe natural-flow regime paradigm (15) posits thatthe restoration of pre-dam flow variability max-

imizes ecological outcomes (15ndash18) Althoughmany empirical studies support this hypothesisremoval of large dams and restoration of pre-dam conditions is rarely possible especially forsystems with large hydropower dams or storagereservoirs that have long lifespans and are centralto socio-environmental water systems (19)In tropical basins construction of mainstem

and tributary dams is imminent and requires fore-sight about how these new facilities are imple-mented In these cases strategic operation ofexisting and future infrastructure to deliver theright dose of variation via smart control (20ndash22)may be a more effective and immediately de-ployable strategy to deliver human and ecologicaloutcomes This recent paradigm shift is fuelingmuch research on algorithm development to pre-scribe ecological objective functions [eg (23)]and optimize these with other competing objec-tive functions to harness water for power genera-tion irrigation and other human uses [reviewedin (24)] The framework for optimization is richbut the algorithms for prescribing an ecologicalobjective function are generally less well de-veloped Here we address this limitation usinga data-driven time-series approach relating hydro-logic variation to fishery yields and we providetangible alternatives in the context of hydropowerdevelopment and inland fishery production inthe Lower Mekong Basin (LMB)The Mekong is the eighth largest river by dis-

charge and hosts one of the largest inland fish-eries in the world (25) The proposed scope ofhydropower development in the Mekong andother tropical rivers in Asia Africa and SouthAmerica poses pronounced trade-offs for bio-

diversity and the production of freshwater fishfor food (26 27) In the LMBmonsoon rains drivea flood pulse that inundates floodplain habitatsthroughout the basin In flood-pulse fisheriesthe duration andmagnitude of the flood pulsemdashwhich controls the spatial extent of inundation(28)mdashis a key driver of production (28) but thereis some support for the notion that flood timingespecially the relative sequence of weak andstrong flood pulses is important (29 30) Thissuggests that regulation of the timing and mag-nitude of the high and low flows could be har-nessed to improve the fisheries However theexact features of a hydrograph that could pro-mote fishery yields in the LMB and how thesemight be translated into actionable design prin-ciples remain unknownWe tested the historic role of multiple var-

iance components of the LMB hydrograph indriving harvest of fishes (total annual harvestin kilograms standardized by effort) from thebag net or ldquoDairdquo fishery on the Tonle Sap RiverThe river connects the Mekong to the Tonle SapLake the largest lake and wetland in SoutheastAsia and a nursery habitat for as many as 300species of fishes (25 31) many of which provide akey source of animal protein and vitamin A forrural subsistence fishing and agricultural com-munities (32ndash34) The Dai fishery is one of themost valuable and productive commercial fish-eries in the LMB (35 ) and has the longest andmost complete record of fishery catches in theregion with biomass and species compositiondata spanning 17 years (1996ndash2012) from 64 Daisat 14 locations along the Tonle Sap RiverOur analysis expands on previous research in

at least three ways (i) Our data-driven ratherthan mechanistic modeling approach allows usto connect key aspects of hydrology to fisherydynamics while minimizing assumptions arisingfrom specified eco-physiological parameters (36)Briefly we leverage observed relationships be-tween variance components of recent hydrologyand harvest to identify shape features in thehydrograph that should maximize desired out-comes for the future fishery Hence our analysisprovides a path for scientifically informed pro-active management of a flood-pulse fishery insystems in which a reasonable understanding ofthe biological mechanisms at play is not withinreach (ii) We focus on design of future hydrologyvia ecologically informed dam operations Thisfocus is timely in the Mekong given the im-minence of extensive hydropower in the LMBand the potential to undertake this developmentin ways that minimize the impact on the fishery(and thus on regional food security) The methodfor design that we propose leverages a highlyflexible and well-understood spectral toolboxthat would have direct transferable applicabilityto systems engineers who operate dams In shortthe design provides a potential tool for manag-ing mainstem hydrology for fisheries as newdams come on line (iii) We articulate a set ofecological design principles that achieve socio-ecological goals by modifying the variance withinan annual hydrograph without increasing mean

RESEARCH

Sabo et al Science 358 eaao1053 (2017) 8 December 2017 1 of 11

1Future H2O Knowledge Enterprise Development ArizonaState University Tempe AZ 85287 USA 2Julie Ann WrigleyGlobal Institute of Sustainability Arizona State UniversityTempe AZ 85287 USA 3Faculty of Ecology Evolution andEnvironmental Science School of Life Sciences Arizona StateUniversity Tempe AZ 85287 USA 4National SocialEcological Synthesis Center (SESYNC) University ofMaryland Annapolis MD 21401 USA 5School of Aquatic andFishery Sciences University of Washington Seattle WA98105 USA 6Betty and Gordon Moore Center for ScienceConservation International Arlington VA 22202 USA7Department of Civil and Environmental EngineeringUniversity of South Florida Tampa FL 33620 USA 8MekongRiver Commission Secretariat Chak Angre Krom MeancheyPhnom Penh Kingdom of Cambodia 9Department of BuiltEnvironment Aalto University 00076 Aalto FinlandCorresponding author Email johnlsaboasuedudaggerPresent address CNRS Universiteacute Toulouse III Paul Sabatier ENFAUMR5174 EDB (Laboratoire Eacutevolution amp Diversiteacute Biologique)F-31062 Toulouse France DaggerPresent address Mekong RiverCommission Secretariat Ban Sithane Neua Sikhottabong DistrictVientiane 01000 Lao PDR

on August 20 2020

httpsciencesciencem

agorgD

ownloaded from

annual discharge A focus on variance shiftsthe conversation from ldquoHow much water dowe need for the fisheryrdquo to ldquoWhen do we needit the most and when can we spare itrdquo (37)

Defining the flow objective function forthe Tonle Sap fishery

We used a data-driven time-series approach thatallowed us to quantify interannual variationin discharge variance and connect this to changein harvest of the fishery Hydrologic variancewas quantified and decomposed into nine shapecomponents describing departure from the long-term trend using discrete fast Fourier transform(DFFT) methods (38) These metrics of hydro-logic variance were then tested in terms of theirability to explain variance in fishery harvest ac-cording to a multivariate autoregressive statespace (MARSS) modeling framework (39)

Discharge variance

We measured recent discharge variability andcentury-scale hydrologic change (see below) usingDFFTmethods (38) (Fig 1) on daily average stage(water level) data for theMekongRivermainstem(Stung Treng Cambodia) The DFFT is a type ofspectral analysis that converts a time series intoa vector of amplitudes (power) and phases asso-ciated with all frequencies in the data set (typi-callyN2whereN is the length of data in the timedomain) In a hydrologic context a small numberof these characteristic frequencies [lt3 in mostcases (38)] constitute the characteristic signal ofthe discharge time series seasonal variation in thehydrograph We extracted this characteristic sig-nal for the time series spanning 1993ndash2012 (con-sistentwith the record of fishery yield 1996ndash2012)Using the characteristic signal for 1993ndash2012

we then identified low- and high-flow deviationsfrom the characteristic signal These deviationsare called anomalies and they sum to zero acrossthe entire time series However there is consid-erable interannual variation in the sum of within-year anomalies Hence the sum of all positive andnegative anomalies in a year or the net annualanomaly (NAA) provides a simple compositemeasure of annual discharge variance ThereforeNAA describes anomalous or atypical wetness(positive NAA) or dryness (negative NAA) in anygiven year (40 41)TheNAA in turn consists of nine components

that define the sequence of deviation (ie theshape) of the annual hydrograph in terms ofmagnitude duration and frequency of low andhigh anomalies (Fig 1) Noteworthy among theseare the interflood interval (IFI) whichmeasuresthe number of contiguous days between maxi-mumhigh-flow anomalies in adjacentwater yearsand the equivalent interdrought interval (IDI) thatmeasures contiguousdaysbetweennegative anom-alies Spectral anomaly magnitude and frequency(SAM and SAF) respectively quantify the strengthand number of independent observations ofwithin-year anomalies Thus these metrics canbe negative (low or LSAM) or positive (high orHSAM) relative to the long-term average condi-tion The timing of LSAM andHSAM can bemea-

sured against the expected long-term peak signal(in days) Finally transition time measures thenumber of contiguous days between the daywiththe greatest absolute magnitude LSAM and theday with the greatest HSAM (Fig 1) See table S1for quantified annual measures of NAA and itsnine component measures of variance for therecord spanning 1996ndash2012We further quantified a more standard mea-

sure of the magnitude of the flood pulse floodpulse extent or FPExt FPExt is defined as theaverage deviation between stage and baseflowmultiplied by the number of days in which flowexceeds baseflow (ie average magnitude times du-ration) Thismetric has been positively related tothe spatial extent of flooding and fish harvest inflood-pulse fisheries (28) In our analysis of hy-drologic drivers of the fishery we divided these11 metrics into two groups of covariates ldquohigh-levelrdquo (HL) including FPExt andNAA and ldquoNAA-componentrdquo including IFI IDI HSAM LSAMHSAF LSAF timing of HSAM and LSAM andtransition time

Flow-fish relationships

Flow-fish relationships were developed usingcatch (biomass in kilograms) and effort data (Daidays equal to number of fishing days across allDais in a given Dai row and year) collected bythe Inland Fisheries Research and DevelopmentInstitute of Cambodia and maintained by theMekong River Commission The data set includes

monthly species-specific harvest allowing thecalculation of catch per unit effort (CPUE) for 64Dais in 14 locations on the Tonle Sap River Wepooled catch data within location for each yearculminating in a data set of 14 time series of totalCPUE each 17 years longThe relationship between annual HL and

NAA-component drivers and total CPUE wasdetermined using a multivariate autoregressivestate-space (MARSS) framework Multivariateautoregressive models are common in econome-trics where they are referred to as vector autore-gression In ecology they have been used mostoften to identify drivers of community and foodweb dynamics [reviewed in (36)] Because theydo not require eco-physiological parametersMARSSmodels can be less challenging to param-eterize than mechanistic models instead ofmeasuring the parameters those parameters areinferred from observational time-series data ac-cording to theory about temporal correlation pat-terns (42)More specifically abundance time seriesare used to estimate parameters associated withpopulation growth biotic interactions or envi-ronmental stochasticity (42)Moreover the state-space approach enables the separation of oftensubstantial observation error from true fluctua-tions in yield allowing for amore accurate identi-fication of drivers (43 44) We used the ldquoMARSSrdquoR-package (39) which provides support for fittingMARSS models with covariates to multivariatetime series data via maximum likelihood using


Fig 1 Illustration of variance components of a hydrograph including the net annualanomaly (NAA) and its components The spectral anomaly magnitude is the largest negative(1 LSAM) or positive (2 HSAM) anomaly within an 18-month hydrologic window that includestwo peaks and troughs in the long-term seasonal signal (cyan solid sinusoidal line) Timing ofHSAM and LSAM (3 and 4 in days timeH and timeL respectively) is determined with referenceto the ordinal day of the peak in the long-term seasonal signal (star) Transition time (5) is thenumber of days separating LSAM and HSAM in the same 18-month window Finally the durationsof consecutive strings of low-flow and high-flow anomalies (in days) are the IFI (6) and IDI(7) The spectral anomaly frequencies (HSAF and LSAF not illustrated here) are numbers ofindependent high- and low-flow events in the same 18-month window where independence isdetermined by intersection with the long-term seasonal signal (38)

RESEARCH | RESEARCH ARTICLEon A

ugust 20 2020

httpsciencesciencemagorg

Dow

nloaded from

an expectation maximization algo-rithm In our case the covariates areannual measures of hydrologic var-iation extracted fromDFFT Inmatrixform theMARSSmodel can be writ-ten as

xt frac14 Bxt1 thorn Cct thornwt eth1THORN

yt frac14 Zxt thorn vt eth2THORN

where wt ~ MVN(0 Q) and vt ~MVN(0 R) The fishery data areharvests of all species in one yearfrom a single row of Dais (hereafterldquoDairdquo) standardized by effort (CPUE)where effort is Dai days in a yearThese data enter the model as y inEq 2 where yt is the log-transformedCPUE of all species at each Dai WethenmodeledCPUEas a linear func-tion of the ldquounobservablerdquo or truecatch (xt) and vt a vector of obser-vation errors In the state process(Eq 1) B is an interaction matrixthat can be used to model the ef-fect of abundances on each other(eg density dependence) C is thematrix whose elements describe theeffect of each hydrologic covariateon abundance at eachDai andwt isa matrix of the process error withprocess errors at time t beingmulti-variate normal with mean 0 andcovariance matrix Q In our casethe covariate data (ct) comprised asuite of 11 possible metrics of dis-charge variation for the current har-vest year described above (Fig 1)thereby allowing us to measure theimpact of hydrology at time t (iethis year) on the change in harvestbetween t ndash 1 and t (ie last yearand this year)The number of covariates is large in number

relative to our data set more important manyof the covariates are likely correlated (eg HSAMand LSAM should be negatively correlated withina year) In addition to this not all covariates arelikely important To address potential estimationbias associated with collinearity and model over-specification we devised a three-stepmodel selec-tion routine (i) We divided our drivers into HLand NAA-component groups and carried out se-parate MARSS analysis on the same CPUE datausing either HL or NAA-component drivers ascovariates HL drivers are highly correlated (figS9) but not collinear [variance inflation factor(VIF) lt 5] Hence the HLmodel included bothNAA and FPExt The observation that NAA andFPExt are correlated underscores the positiverelationship between high-flow anomalies cap-tured in NAA and the extent of the flood pulseThe lack of collinearity reinforces the indepen-dence of low-flow events and FPExt (ii) For theNAA-component model we first screened driverswith potentially high collinearity eliminating all

drivers with VIF gt 5 (iii) This VIF screen wasfollowed bymodel selection using an all-possible-subsets (APS) routine on the model with all re-maining noncollinearNAA-component covariatesIn this APSwe considered onlymain effects (notinteractions) of the remaining independent NAAdrivers Alternative models were tested using aninformation-theoretic approach [Akaikersquos infor-mation criterion corrected for small sample sizeAICc (45)] that represented different hypothesesconcerning which drivers best explain observedfishery yields Included in model selection wereall possible combinations of hydrologic variables(after culling for strong collinearity) Ranges ineffect sizes of all parameters are presented as 95confidence intervals (CIs) basedon 1000parametricbootstrap samples (41)

Result 1 An ecological objective functionfor the Mekong

We found that FPExt had strong positive effectsizes and that NAA had strong negative effectsizes indicating that flood pulse extent and low

flows (negative NAA) are almostequally strong and significant driv-ers of the fishery (Fig 2 and Table 1)Of the many shape components ofNAA (41) four were positively relatedto fish catch the interflood interval(IFI) the maximum magnitude ofspectral anomalies (HSAM) the fre-quency of events with large positivespectral anomalies (HSAF) and theminimum (negative) magnitude ofspectral anomalies (LSAM) Of thesehigh-level drivers IFI had the highest-magnitude effect followed closely byHSAM However the effect of LSAMwas significant inmagnitude indicat-ing that low flows (ie IFI LSAM) areas important in determining catchas the magnitude of the flood pulseitself (HSAM Fig 2) Hence the de-sign principles for productive fish-eries in this system include a long dryperiod (large IFI large-magnitudenegative NAA and LSAM) punctu-ated by a strong flood pulse (largeFPExt and HSAM) characterized bymultiple storm peaks (high HSAF)Variance in addition to the magni-tude of the flood pulse is paramount

Density dependence

Our initial analysis assumed densityindependence (ie a Bmatrix withones in the diagonal and zeros inthe off-diagonal elements in Eq 1)based on the assumption that highfishing pressure shouldmaintain totaldensity below the level triggeringnegative feedbacksDespite this likelyscenario we also analyzed HL andNAA-component models assumingdensity dependence (ie B diagonalestimated by MARSS rather thanfixed at 1) For HL models the sign

and relative magnitude of coefficients for FPExtand NAA were preserved the flood pulse andnegative NAA had positive effects on the fisheryin the density-dependentHLmodelHowever theeffect size for NAA was nonsignificant under theassumption of density dependence likely owingto lower power in a more heavily parameterizedmodel Similarly for NAA-component modelseffect sizes generally were in accord betweendensity-independent and density-dependent models(Fig 2) NAA-component models under bothassumptions featured significant effect sizes ofIFI andLSAMOne significant difference betweenthese twoNAA-componentmodels was a strongerinfluence of low flows in the density-dependentmodel (high-magnitude effect size for LSAM andnonsignificant effect size for HSAM and HSAF)We note that one of the NAA-component density-dependent candidatemodels producedanull (zero)process error variance estimate thereby precludingpropermodel averaging andprescription of robustdensity-dependent designs This was likely causedby limited data relative to estimated parameters


Fig 2 Weighted coefficients (effect sizes) and 95 confidenceintervals from the MARSS models quantifying the hydrologic drivers oncatch per unit effort in the Dai fishery on Tonle Sap River (A) Resultsfrom the density-independent model (B) results from a model thatincludes density dependence The fishery has been monitored annually at14 Dais since 1996 we analyzed quality-controlled data from 1996 to 2012Hydrologic drivers are estimated from daily stage of the Mekong river atthe Stung Treng gage (record 1993ndash2012) Drivers are defined in Fig 1Where confidence intervals cross the dotted line (zero) effect sizes arenot statistically significant


ugust 20 2020


Dow

nloaded from

which was not the case for any of the density-independent candidate models Nonethelessregardless of the assumption about density de-pendence the flood pulse and low-flow anom-alies are important (positive) predictors of harvest(CPUE) (41)

Long-term hydrologic change vis-agrave-visdesign principles

Our observation that the ideal hydrograph forthe fishery is one with a long dry period punc-tuated by a large-magnitude flood pulse is troubl-ing given observed trends in discharge variationin the LMB Substantial work suggests that damsmute variation dampening the flood pulse andaugmenting low or base flows (46ndash48) We devel-oped a spectral approach to quantify century-scale change in discharge variance and we usedthis new tool to measure change in high-level andNAA-component drivers of the Mekong fisherywith reference to the closure of the first dam onthe Mekong (Ubol Ratana dam in the head-waters of the Mun River Thailand in 1964) Thisdam was chosen not to imply that all potentialchange in hydrology derives from a single struc-ture (Ubol Ratana) but rather to pinpoint thestart of extensive and continued developmentin the basin

Century-scale hydrologic change

We constructed a detrended hydrologic baselineusing a windowed average DFFT analysis acrossall 37 consecutive 20-year time series from 1910to 1964 (eg 1910ndash1919 1911ndash1920hellip 1945ndash1964)predating the closure of Ubol Ratana In eachwindow we estimated the trend characteristicphase amplitude and frequency of the signalThis produced a sample of 37 point observationswhich was bootstrapped to generate unbiasedmeasures of central tendency for these historicalparameters Bootstrapped values for frequencyamplitude and phase of significant signals allowus to reconstruct a detrended historical baseline(hereafter ldquobaselinerdquo green line in Fig 1)With a detrended baseline in hand we then

compared (detrended) post-dam time series tothe baseline to estimate NAA and its componentsThis was done using a windowed approach as


Fig 3 Century-scale changes in hydrology and climate (A and B) Change in high-level (HL)hydrologic drivers of the Tonle Sap fishery Panels show estimatedmetrics (circles) as raw estimates (A)or as deviations from the average long-term pre-1964 trend (B) In both cases hydrologic drivers areestimated from a 20-year window of data and plotted by the midpoint of this window and estimates areoverlaid with nonparametric broken-stick regression linesWide light-gray boxes indicate the hydrologydata window that includes closure of the first large dam in the basin (Ubol Ratana Mun River basinThailand) Vertical solid lines indicate statistically significant breakpoints in the data as quantified bynonparametric multiple change point analysisWhere multiple breakpoints exist thickness of linesindicates the order of breakpoint identification with thicker dashed lines indicating first breakpoints(C and D) Change in basin-wide hydroclimate Hydroclimate change is estimated as basin-wide PalmerDrought Severity Index (PDSI) reconstructed from tree rings (climate) Climate is a point estimate forthat year (no window) raw estimates are overlaid with nonparametric broken-stick regression lines as in(A) Narrow darker gray boxes indicate the climate data window that includes the same closureWidelight-gray boxes and vertical solid lines (and thicknesses) are as in (A) and (B) (E to G) Change inNAA-component hydrologic drivers of the Tonle Sap fishery NAA components capture the shapefeatures of discharge variance (ie NAA) Change in metrics was estimated as in (B) by estimation ofdeparture from long-term pre-1964 trend All symbols lines and boxes are as in (A) and (B)

Table 1 Parameter estimates for density-independent MARSS models on historical hydrology and catch data All variables except the net annual

anomaly under density dependence show significant effects (ie 95 confidence intervals obtained via bootstrap nonoverlapping with zero) CL confidence limitParameters from the HL model assuming density dependence are included in parentheses Transition time IDI and timeH were excluded because of collinearity or

nonsignificant effect sizes

Variable AICc weight Weighted coefficient Weighted SE Upper CL Lower CL

Flood pulse extent (FPExt) 1 (1) 1080 (0574) 0313 (02) 1693 (0 96) 0467 (018)

Net annual anomaly (NAA) 1 (1) ndash0813 (ndash027) 0318 (02) ndash1437 (012) ndash0189 (ndash068)

Interflood interval (IFI) 0522 0345 0128 0596 0094

HSAM 0522 0322 0119 0556 0088

HSAF 0478 0187 0070 0196 0177

LSAM 0151 0066 0085 0080 0051

LSAF 0135 ndash0034 0070 ndash0025 ndash0044

Timing of low (timeL) 0286 ndash0055 0069 ndash0045 ndash0064


ugust 20 2020


Dow

nloaded from

above but across all 83 consecutive 20-year timeseries from 1910 to 2012 (eg 1910ndash1919hellip 1965ndash1984 1966ndash1985 hellip 1993ndash2012) Hence bothpre-dam and post-dam data were compared toa smoothed and bootstrapped baseline to obtain83 point estimates of NAA and the componentsof NAA Comparison of detrended post-dam ob-servations to detrended baseline hydrology al-lowed us tomeasure changes in variance (secondmoment) not confounded by changes in the trend(first moment) We also estimated average FPExtnumerically across this same record for all 8320-year time windows in our record We have

thoroughly documented these methods (seebelow)

Breakpoint analysis

With 83-year time series for FPExt NAA and NAAcomponents we identified breakpoints in thetime series using a divisive hierarchical estimationalgorithm for multiple change point analysis(49 50) To further strengthen the inference thatdams andnot changes in climatewere responsiblefor any observed changes in hydrologic variancewe performed the same breakpoint analysis onannualized basin-scale Palmer Drought Severity

Index (PDSI) estimates reconstructed from treering data (Fig 3) (51 52)

Result 2 Change in hydrologic varianceafter dam closure

Our measure of FPExt has declined since theclosure of the first dam on the Mekong andthere have been similar declines in IFI HSAMand HSAF and countervailing increases in NAA(Fig 3) (41) Thus dam development changeshydrologic variation in a manner exactly oppo-site to the pattern that promotes higher fisheryyields Hydroclimate has changedmultiple timesover the observed discharge record (1910 to thepresent) but identified breakpoints of changein the hydrograph do not coincide with changesin the PDSI However they do coincide with theperiod of dam closure (1954ndash1974) in tributariesof the LMBMany of the initial damswere storagefacilities (41 53) that greatly affect the shape (11)and hence the NAA components of the hydro-graph (Fig 3)

Restoration potential of designed andnatural flows

As with some of the first dams closed on tri-butaries of the LMB many of those slated forconstruction in the Sekong Sesan and Srepoktributaries of the Mekong in Lao PDR Vietnamand Cambodia will create floodwater reservoirsthat have potential to be used for dry-season agri-culture (48) and could thus be used to generatedesigned flows We hypothesized that implemen-tation of controlled releases from these storagefacilities mimicking the key design principlesarticulated above would deliver desired fisherytargets equally or more effectively than restora-tion of the natural-flow regime

Spectral implementation of an ecologicalobjective function

To test this hypothesis we developed a flexiblealgorithm using a signal-processing tool kit thatgenerates target river flows that capture keyhydrologic elements associated with strong fish-ery yield (Fig 2) This algorithm is a compoundsinusoidal generating an asymmetric rectangularpulse train

xt frac14 a0 thornXinfin

nfrac141

an cos2p ftn Xinfin

nfrac141

bn sin2p ftn

eth3THORNwhere a0 = Ad an = (2Anp) sin npd and d = kTHere A d f k n t and T are the amplitudeduty cycle frequency length of the flood pulsewave series number time (in days) and period(365 days) respectively Here we consider anasymmetric rectangular wave (bn = 0) to producea deterministic hydrograph with known charac-teristics that was then corrupted with red noise

ct frac14 rxetht DtTHORN thorn frac12eth1 r2THORN1=2et eth4THORN

where et ~ N(0 s) r = 095 and s = 0042Asymmetry (ie longer IFI) arises in this pulse

train from a short duty cycle (associated with a


Fig 4 Illustration of designed flows including a ldquoGoodrdquo design a ldquoBadrdquo design and thereconstructed historical or ldquonatural-flowrdquo regime Good designs are asymmetric rectangularpulse trains with long troughs (ie long IFI) and punctuated high flows (HSAM see Eq 3)the shape of the Good design emerges from setting the duty cycle (d = kT) of the pulsetrain low (d = 041 k = 150-day flood pulse) and the amplitude high (A = 07) Under theseconditions there is a large negative NAA owing to sustained negative anomalies across thetrough of the pulse train as well as high IFI and HSAM all three of which generate a positiveharvest Bad designs have lower asymmetry leading to positive anomalies in the trough (low IFI)and a weak flood peak The dot-dashed line is the reconstructed historical signal


ugust 20 2020


Dow

nloaded from

brief flood pulse) and the reddened noise spec-trum provides a more natural random walk thatis characteristic of storm sequences We createdtwo designed flowsmdasha ldquoGoodrdquo design (long IFIand large relative amplitude of flood pulse) anda ldquoBadrdquo design (shorter IFI and small relativeamplitude of flood pulse)mdashsimulating the expectedeffects of medium-term hydropower develop-ment (Fig 4)We also created pre-damhydrographsrepresenting restoration of the natural-flow re-gime by resampling a gt50-year record of observedannual metrics for NAA and FPExt from theobserved record before the closure of the firstdam on the LMB in 1964 and strings of NAA andFPExt representing current conditions (Fig 4)We then projected the fishery forward preserv-ing the same MARSS model structure and usingthe estimated coefficients for discharge anomalyeffects (in C) and process error variance (inQ)Trajectories of fishery catch were constrained bythe HLmetrics for each scenario We used catchat each Dai in 2012 as initial values ran 100000simulations per scenario and then pooled abun-dances across Dais by realization to obtain 2525 50 75 and 975 percentile catch levels at eachtime step (8 years ie 2013ndash2020)For each scenario we computed ldquoglobal deltasrdquo

as the difference between median pooled catchat the simulation end (year 8) and initial pooledcatch divided by the initial pooled catch (ie ob-served pooled values in 2012) We similarly com-puted ldquoannual deltasrdquo between successive yearsand calculated a mean annual delta for each sce-nario as a relevantmetric for fisherymanagement(ie the expected percent change in next yearrsquosfishery in response to changes in the flow regime)Finally to test the robustness of our findings underdifferent assumptions we compared the projec-tions obtained in the previously describedmodelto those obtained in a model that incorporatesdensity dependence [that is an aggregate mea-sure of carrying capacity of the fisheries (41)]

Result 3 Engineered hydrographperforms as well as or better thannatural-flow regime

Natural flow restoration (the ldquoPre-damrdquo scenario)produced a 47 annual increase in yield (medianbiomass) relative to the modern fishery (Fig 5A

and Table 2) under the assumption of densityindependence This increase is likely attributableto a stronger flood pulse during the pre-dam eracaptured by the reconstructed covariates usedto project this scenario This increase was nothowever significant (CI ndash072 to 73354) De-spite substantial increases under restoration ofnatural flows designed flows mimicking longIFI and a short but strong flood pulse (the ldquoGoodrdquodesign) produced even greater yields thannatural-flow regimes (76 annual increase Fig 5A andTable 2)The lower bounds of 95 CIs were positive

(ie the intervals did not include zero) for boththe projected harvest (Fig 5) and the global deltasbased on 10000 resampled realizations (CI 066to 202217 Table 2) This indicates higher cer-tainty in increased yield under this scenario thanunder restoration of natural flows Global deltasby the end of the 8-year forecast under the Gooddesign exceeded those under the Pre-dam sce-nario by a factor of 37 (Table 2) These improve-ments in yield emerged evenwith designed flowsfeaturing an average flood pulsemdashcomparable inmagnitude to that of the observed (current) hy-drological recordmdashby accentuating variation be-tween high and low flows via long IFI and shortand strong HSAM (Fig 4) By contrast short IFIand amuted flood peak (ldquoBadrdquo design Fig 4) ledto sharp (53) annual declines in yield in the Daifishery (Fig 5 and Table 2)The direction of change under each scenario

did not change if density dependence was in-cluded in the model (Fig 5) The differences be-tween theGood design and the Pre-dam scenariowere dampened By the end of the projection theGood design under density-dependent models de-livered total deltas that were higher than those forthe Pre-dam scenario by a factor of 16 versus afactor of 37 difference under density-independentmodels (Table 2) However the Good design wasagain the only scenario delivering a certain(Pgt 95) catch increase by the end of the forecast(Fig 5 andTable 2)Hence engineering the timingof lows and highs relative to the natural-flowregime can produce results similar to or betterthan mimicking the natural-flow regime itselfFinally a sketch of existing and planned dam

locations and reservoir storage with respect to

key spawning areas for Tonle Sap Lake fish tiedto regional food security suggests that designedflows are possible (Fig 6) There is already ade-quate storage in existing storage facilities in Chinaand in the tributaries of theMekong in Lao PDRand Thailand to store and release designed flowsBy contrast further development of mainstemhydropower facilities lower in the basin presentsa trade-off between access to key spawning areasand the flexibility to deliver designed flows fromstorage facilities Specifically the closure of main-stem dams in Cambodia would make designedflows moot as all upstream spawning groundswould be inaccessible to Tonle Sap Lake fishes

Conclusion and outlook for managingthe Mekong

Our results have implications for themanagementof hydropower development in tropical riverbasins many of which provide food security insome of the poorest countries in the world viacapture fisheries In this context previous worksuggests two hypotheses to explain how dams di-minish production of inland fisheries (i) by re-ducing connectivity and dispersal (54) and (ii) byreducingprimaryproductivity through the entrap-ment of sediment supply and associated nutrientdelivery from headwaters to downstream nurseryhabitats (46 55) The importance of a long droughtperiod (high IFI) suggests that changes in thetiming of drought and of the flood pulse influencesecondary production and fish harvest indepen-dent of impacts on connectivity and sedimentsupply Therefore we propose a third hypothesisFlow variation (high and low) may drive produc-tion by controlling redox conditions in floodplainsoils and hence the production of terrestrial nu-trients and organic matter and the transfer of thismaterial to the aquatic ecosystem This hypothe-sis however remains untested Alternatively theimportance of a large flood pulse (large FPExt)may in part be related to increased catchability offish at highwater levels Our state-space approachdistinguishes between real fluctuations in thefishery and fluctuations in catch numbers arisingfromvariability in observation or sampling Thisprovides evidence against this fourth hypothesisand suggests that our conclusion may be robustto flow-related increases in catchability Future


Table 2 Results of the MARSS simulations under the four scenarios Annual deltas are the average percent increase in median catch between

successive years Global deltas represent the difference between the final (year 8) median pooled catch and the initial catch (year 1 of the simulation)divided by the initial catch (year 1 of the simulation) with the result expressed as a multiple of initial catch See Fig 5 for confidence intervals of the

simulations The Good scenario performed best in both density-independent (DI) and density-dependent (DD) forecasts (their deltas are larger than those of

the respective Pre-dam flow regimes by factors of 37 and 16) they are the only scenarios delivering a certain (P gt 95) catch increase by the end ofthe forecast regardless of the model structure used Ranges in parentheses are 95 confidence intervals of global deltas based on 10000 random resamples

of the pooled realizations of Fig 5

Model structure Metric Good Bad Pre-dam Current

DI projections Annual delta () 7613 ndash5338 4727 584

Global delta 5127 (066 to 202217) ndash0995 (ndash100 to ndash083) 1402 (ndash072 to 73354) 0485 (ndash098 to 9992)

DD projections Annual delta () 2166 ndash950 1579 487

Global delta 287 (012 to 1745) ndash0519 (ndash079 to 012) 175 (ndash024 to 1344) 0389 (ndash063 to 584)


ugust 20 2020


Dow

nloaded from

research should attempt to distinguish the rel-ative importance of these four possible mech-anisms linking hydrologic variation to fisherycatchesThe asymmetrical pulse train specified above

(Eqs 3 and 4) represents a flexible design algo-rithm that captures fundamentally importantfeatures of hydrologic variance for fish produc-tion and yield in theMekong Long IFI also drivesprimary production in desert rivers where nitrogenaccumulation within the watershed and subse-quent input are positively related to a dischargesequence including long dry periods (months) fol-lowed by exceptional high flows (56ndash58) Hencethe asymmetric pulse train could be a generaldesign principle with wide application in fresh-water conservation with respect to environmen-tal flow restoration Regardless of its generalityidentifying robust ecological objective functionsearly in the planning process is a necessary pre-cursor to optimization with other competing ob-jective functions (ie hydropower) the pulse trainprovides a viable ecological objective function forsustainable operations of existing and planneddams in theMekong Although designed flows area potentially powerful tool to increase fish produc-tion and therefore fishery yield a designparadigmdoes not guarantee sustainability of the resourceExceptionally little is knownabout the current levelof sustainability of the LMB fishery Nonethelessimplementing designed flows coupledwith inten-sive time-seriesmonitoring of catch can provide ameans to better understand andmanage the fish-ery through adaptive management (59)Because the record used in this analysis post-

dates all major recent dam construction in thelower Mekong our inferences are relevant de-spite any past impacts of dams on upstream dis-persal and downstream sediment transport Futureimpacts of dams on migration and sediment

delivery are however not directly considered inthis analysis The majority of fishes migratingin and out of the Tonle Sap are small explosivebreeders in the familyCyprinidae (HenicorhynchussppLabiobarbus spp andParalaubuca spp totallength lt25 cm) and likely do not migrate longdistances relative to the longer-lived catfishes(eg Pangasiidae) and larger barbs and carp(eg Probarbus Catlocarpio family Cyprinidae)These small cyprinid fishes are nonetheless a staplein the dailymeals of local residents across the LMBand have even given their name to the Cambo-dian currency [riel (60)] This tight connectionbetween fish and human society underscores theneed for formal trade-off analysis between hydro-power and food systems including fish and riceThis task is achievable given the articulation ofan ecological objective function for theMekongas described above and the breadth of approachesin multi-objective optimization already available(24 61ndash63)

Methods

Our data-driven approach connects past hydro-logic dynamics of theMekong to fishery resources(specifically fishery catches standardized by ef-fort) and uses estimated coefficients based onthe historical covariation to designed flows andforecast fishery catches under different scenarios(Fig 7) Each step is further articulated below

Step 1 Quantifying past hydrologicalchange and variability

Discrete fast Fourier transform (DFFT) anddischarge variability metrics

Using available stage data at the Stung Trenggage we identified daily discharge anomaliesusing DFFT (38) Fourier analysis allows for theextraction of the characteristic seasonal profile

of a time series constituted by the recurrent fre-quencies amplitudes and phases With the re-constructed hydrograph (or ldquoseasonal profilerdquo)it is then possible to identify low- and high-flowanomalies or departures from the expected sea-sonal profile We defined a suite of metrics forcharacterizing discharge variability as followsFlood pulse extent (FPExt) (28) is defined as

the product of the average deviation betweenstage and baseflow and the number of days inwhich flow exceeds baseflow (ie average mag-nitude times duration) Here we defined baseflow asthe average stage for the time series To estimateNAA we rely on spectral methods defined in (38)A spectral anomaly is the difference between thedaily observation and the ldquocharacteristicrdquo or20-year seasonal signal obtained from the fre-quency domain viaDFFTand reconstructed in thetime domain (Fig 7) NAA is the net sum of allpositive and negative anomalies in a year Al-though the net sum for the entire time seriesused to estimate the characteristic signal is nullthere is generally ample interannual variation inthis statistic In this way NAA describes anom-alous or ldquoatypicalrdquo wetness (positive NAA) ordryness (negative NAA)NAA (40) is a composite measure of deviations

from an expected hydrograph Its componentsare events that occur outside of the expectedseason (timing component) or are larger thanexpected for a given season (magnitude compo-nent) happenmore often (frequency component)or last longer (duration component) than thoseconstituting the characteristic signal Collectivelythe timing magnitude frequency and durationdescribe the shape of the annual hydrograph andits departure from the norm Hence we developedspectral-basedmetrics for magnitude frequencyand duration [following (38)] as well as newermetrics of timing that are relevant to flood pulse


Fig 5 Projections of fish harvest under designed and naturalflows (A) Density-independent models (B) density-dependentmodels Short-term (8-year) stochastic projections of total biomassharvested from the Tonle Sap Dai fishery forced by hydrologic

conditions that mimic Good and Bad design principles as wellas the Current and Pre-dam (natural-flow) regimes (46) Gooddesigns include high FPExt and negative NAA (high-leveldrivers)


ugust 20 2020


Dow

nloaded from

systems (Fig 1) Noteworthy newmetrics are IFIwhich measures the number of contiguous daysbetween high-flow anomalies and IDI whichmeasures contiguous days between negativeanomalies Finally transition time measures thenumber of contiguous days between LSAM andHSAM (the anomalies with the highest negativeand positive spectral anomaly magnitudes)Using these spectral metrics we set out to

quantify changes in the flood pulse and mea-sures of daily hydrologic variation (from DFFT)across the LMBbefore and during periods of activedam construction (pre-dam 1910ndash1964 post-dam1965ndash2008) The year 1965 marks the closuredate of Ubol Ratana dam on the Phong River inThailand the first dam on the Mekong systemOver the ensuing 6 years (1965ndash1971) four addi-tional dams were built in tributaries resulting infive reservoirs with a total of ~1156 km3 in activestorage the bulk of which is held in Nam Ngum1 Lao PDR (7 km3) Given the large increase intributary storage between 1966 and 1971 we hy-pothesized that filling and ensuing operations ofthese five reservoirs changed the timing of dailyhigh and low flows (within years) and that changesin timing are observable as changes in NAA andits components

Windowing with bootstrapped baseline

The direct beforeafter comparison provides ameans for measuring absolute but not relativechange in key characteristics of the hydrographHere we present a method for prescribing abaseline trend and signal from the historicalrecord (before 1965) that is then used as a refer-ence for comparing current hydrology (after 1965)Our baseline is appealing because it is derivedfrom a distribution of baselines resampled fromdifferent windows of the historical climate recordThis windowing approach requires a long timeseries and the Stung Treng record provided ade-quate sample size to develop a robust baselineWe used the stage data set from the Mekong

River at the Stung Treng gage from 1910 to 2008As above we split this time series into a histori-cal record (1910ndash1964) and a current record(1965ndash2008) To develop a historical baseline wewindowed theDFFT over all 37 consecutive 20-yeartime series from 1910 to 1965 (eg 1910ndash19191911ndash1920 hellip 1946ndash1965) In each window weestimated the trend characteristic phase ampli-tude and frequency of the signal This produceda sample of 37 point observations which we thenbootstrapped to generate unbiased measuresof central tendency for these historical baselineparameters The historical baseline was then esti-mated using the bootstrapped median for thetrend amplitudes frequencies and phases Thisbaseline is a detrended sinusoidal which is thenused as a reference for computing anomalies inthe current recordWith a detrended baseline in hand we then

estimated departures from that historical expec-tation as residuals (anomalies) of detrended ob-servations referenced to the baseline for the entiretime series This differs from direct beforeaftercomparison in that daily discharge is referenced

to the detrended signal from the historical recordrather than the current record As in the historicalanalysis we did this via a window-bootstrappingprocedure in which we compared observationsof daily discharge from all consecutive 20-yearwindows to the historical baseline In each win-dow we estimated all stochastic properties out-lined in Fig 1 We also computed FPExt andrelated traditional flood pulse metrics in eachwindow which does not require a baseline forcomparisonThe end result of the windowed bootstrap is a

pair of distributions for eachmetric of hydrologicvariability one for the historical record and

another for the current record both referencedto the average historical baseline Visual inspec-tion of overlap of distributions allows for heuristicassessment of statistical change In addition tothis comparisonof distributionswe plotted trendsof historical and current records chronologicallyto provide a tool for visualizing breakpoints indirection For several key parameters (FPExt NAAIFI HSAM and LSAM) we conducted a formalbreakpoint analysis Specifically to test for sig-nificant breakpoints we used a divisive hierar-chical estimation algorithm for multiple changepoint analysis (49 50) This nonparametric ap-proach is robust to violations of key assumptions


Fig 6 The potential for designed flows and trade-offs between mainstem hydropower lowerin the basin and the possibility of implementing design-based adaptive management Topgraph Example of designed flow versus long-term historical signal of the Mekong River at theStung Treng gauge in Cambodia Note that variation rather than magnitude of the flood pulse ismodified Middle graph Total discharge volume needed to deliver designed flows (gray bar) andvolume available from storage facilities in the mainstem China hydropower cascade (dark blue)mid-river tributaries (royal blue) and planned reservoirs in the Sesan Srepok and Sekong river basins(light blue) Bottom graph Proportion of river length (km) disconnected from Tonle Sap Lake by closureof mainstem dams from the mid-reach to the lowest proposed mainstem dam (Sambor) Numberscorrespond to Gongguoqiao (1) Xiaowan (2) Manwan (3) Dachaoshan (4) Nuozhadu (5) Jinghong(6) Ganlanba (7) Mengsong (8) Pak Beng (9) Luang Prabang (10) Xaybouri (11) Pak Lay (12)Sanakham (13) Pak Chom (14) Ban Koum (15) Lat Sua (16) Don Sahong (17) Stung Treng (18) andSambor (19) Colors and numbers in middle and bottom graphs correspond to those in the map


ugust 20 2020


Dow

nloaded from

of regression and provides a tool foridentifying multiple transitions inthe time series

Step 2 Estimating theinfluence of hydrologicvariation on fishery catches

The multivariate autoregressivestate-space (MARSS)modeling approach

The core of the hydro-fisheriesmodel(Fig 7) is a MARSS model that con-nects measures of hydrologic var-iation to CPUE We first extractedhydrologic covariates from the stagedata set at the Stung Treng gageThis was done for the period 1993ndash2012 to be consistent and overlap-ping with the Tonle Sap Dai catchdata set (1996ndash2012) Analysis of this20-year time series yielded annualestimates of FPExt NAA IFI IDItransition timeHSAMLSAMHSAFLSAF and timing of annual HSAMand LSAM for a total of 11 possiblehydrologic covariates Then wemod-eled total catch at all 14 Dais usingstate-space versions of multivariateautoregressive (MAR) models orMARSS models MARmodels havebeen traditionally used in econome-trics and more recently in someecological applications to quantifytemporal population and commu-nity trends and their drivers [re-viewed in (36)] MARmodels rely ontheory about the patterns of tempo-ral correlation that emerge fromenvi-ronmental drivers and species interactions (42)and are advantageous because they donot requireecophysiological parameters Moreover MARSSmodels allow estimation of non-process or obser-vation error in the data which is important be-cause ignoring observation error can change ourinference about the underlying process (44) Ourfish count data certainly contained observationerror that could potentially bias the measuredinfluence of abiotic drivers (hydrology) on fisherycatch Therefore MARSS allowed us to separatethe variation in the fish count data attributableto observation error from the variation attribut-able to true population fluctuations We used theldquoMARSSrdquoR-package (39) which provides supportfor fitting MARSS models with covariates (ieannualmeasures of hydrologic variation) tomulti-variate time-series data via maximum likelihoodusing an expectation-maximization algorithm Inthe matrix form a MARSS model takes the formfollowing Eq 1 and 2 above (Eq 1 is the stateprocess Eq 2 is the observation process)Data enter the model as y (in Eq 2) and as c

(in Eq 1) where yt is fish abundance (here totalcatch of all species combined) at each Dai andct is a suite of metrics of discharge variation usedas possible covariates (see below) Catch data aremodeled as a linear function of the ldquounobservablerdquo

or true catch (xt) and vt a vector of non-processor observation errors where observation errorsat time t are multivariate normal with mean 0and covariance matrix R (Eq 2) In the stateprocess (Eq 1)B is an interactionmatrix and canmodel the effect of abundances on each other(here we set it to ldquoidentityrdquo ie ones in the diago-nal and zeros in the off-diagonal) C is the matrixwhose elements describe the effect of each co-variate on abundance at each Dai (hereafter dis-charge anomaly effects) and wt is a vector ofprocess errors with process errors at time t beingmultivariate normal withmean 0 and covariancematrix QThe model used here assumes a single process

error variance (ie a single diagonal value in Qfor all Dai) as the biology of the modeled fishesshould not differ across Dais In contrast weassumed Dai-specific observation error (ie 14different diagonal values in R) as site-specificconditions can potentially affect capturabilitymdashand hence observation errormdashin differentwaysin each Dai The off-diagonal structure also dif-fered between Q and R Whereas we estimatedcovariance inQ (as ldquotruerdquo fish abundances shouldbe to some extent coordinated over time acrossDais eg due to stochastic events not accountedfor in the hydrologic covariates that could affect

all Dais simultaneously) we assumedno covariance in R by fixing zeroesin the off-diagonals (as it is unlikelythat measurement error varies in acoordinated manner across Dais)For all models we used the natural

logndashtransformed and Z-scored fishbiomass time series as variates andthe Z-scored hydrologic time seriesas covariates All parameters includ-ing the ones of interest (dischargeanomaly effects in C) were assessedvia 95CIs (1000 bootstrap samples)After fitting the models autocor-relation in the residuals was inspect-ed via the autocorrelation functionTheMARSSmodeling involved threesubsteps1) Covariate selection We first

tested for multicollinearity amongthe 11 covariates As expected NAAwas correlated with almost all NAA-component drivers (ie IFI HSAMetc) Hence we implemented twoseparate MARSS-DFFT models onewith FPExt and NAA as HL driv-ers and the other with 11 NAA-component drivers (see descriptionin Fig 7) For each of these modelswe subsequently tested for collinear-ity FPExt and NAA were stronglypositively correlated (R2 = 085) butcollinearity was not strong enoughto force exclusionof one variable fromthemodel (VIFlt5) By contrastNAA-component drivers were collinear(VIF gt 10) so we culled this set ofdrivers to a group of eight that pro-duced acceptable levels of collin-

earity (VIF lt 5)2) MARSS model fitting and selection We

then fitted the HL MARSS model (ie a modelusing both FPExt andNAA as covariates) and theNAA-component MARSS models For the NAA-component MARSS models we compared allpossible unique combinations among the eightNAA-component (noncollinear) drivers We com-pared model fits (using AICc) as well as co-efficients for all nonzero coefficients in eachmodelWe then used these outputs to calculate variableweights (importance of each driver) andweightedcoefficient estimates (effect sizes because catchand driver data are Z-scored) Effect sizes werecompared betweenMARSSmodel families (ie HLversus NAA-component)3) Inclusion of density dependence To test the

robustness of our findings under different assump-tions we compared the effect sizes obtained in thepreviously described HL model to those obtainedin an HLmodel that incorporates density depen-dence (that is an aggregate measure of carryingcapacity of the fisheries) A deterministic modelfrequently used to describe density-dependentpopulation growth is a discrete-time Gompertzmodel (64)

nt frac14 nt1 exputhorn ethb 1THORN lnethnt1THORN

eth5THORN


Fig 7Workflow diagram of the ldquohydro-fisheriesrdquo model implementedThe modeling workflow included three steps (i) analysis of historicalhydrology (ii) analysis of historical flow-fish relationships and (iii) projectionof fish harvest using designed and historical flows


ugust 20 2020


Dow

nloaded from

where nt is population abundance u is theintrinsic population growth rate and b governsthe strength of density dependence When b = 1there is no density dependence When b getscloser to 0 the strength of density dependenceincreases that is the stronger is the pullbackto the mean (or mean-reverting effect) On a logscale we can substitute ln(nt) for xt

xt frac14 uthorn bxt1 eth6THORN

which is the origin of MARSS Eq 1 (39) BecauseMARSS is multivariate B is a matrix and it canmodel the effect of a Dairsquos catch on the catch atthe next time step (diagonal values) Here wecompared discharge anomaly effects under HLmodels withB diagonal values being estimated(density dependence) versus fixed at 1 (densityindependence)

Step 3 Projecting future fisherycatches under current and designedflow scenarios

Generating designed flows usingFourier series

We compared projected fishery yields achievedby two hypothetical designed flows (ldquoGoodrdquo andldquoBadrdquo) to simulated historical flows in a stochasticframework (Fig 4) To simulate stochastic de-signed flows we created a Fourier series withdesign principles from MARSS-DFFT and cor-rupted this deterministic profile with red noiseto simulate atmospheric forcing Design princi-ples identified byMARSS included a strong floodpulse (high FPExt) and negative NAA with shapecharacteristics that include a long IFI punctu-ated by a large-magnitude flood with multipleimpulses (high HSAM and HSAF) To capturethis shape we created an irregular rectangularpulse train described by Eq 3Large-magnitude negative NAA and long IFI

(ie the Good design) can be achieved in thisFourier series by setting the duty cycle (d = kT )low (d = 041 k = 150-day flood pulse) and ampli-tude high (A=07) Here the period (T= 365 days)is much longer than the length of the pulse (k)such that the pulse train has a lengthy period ofzero amplitude punctuated by strong pulsesWe also set the mean annual flow (MAF) to 95of the value from the observed record used inMARSS-DFFT By assuming lower than histori-calMAF we ensure that our Good designs do notrequire more water than is typically available inthe LMB and are attainable even under heavierwater resource withdrawals than might be re-quired to support irrigated agriculture or munic-ipal use in the future In addition to this Gooddesign we created a Bad design with shorter IFIand a dampened flood peak characteristics ofmanymanaged river systems Here the duty cyclewas set higher (d = 046 k = 170-day flood pulse)and amplitude low (A = 05) and MAF was at100 of historical levelsWe note here that the transfer function re-

presented by Eq 3 produces amuchmore diverseset of designs including square triangle and

rectified waves (also bn = 0) and sawtooth waves(bn = Anp) as well as combinations of each ofthese shapes

Designed flows in a stochastic framework

We created 1000 realizations of a single Goodand a single Bad design using red noise via Eq 4above Red noise creates more realistic storm se-quences than random (white) noise can achieveTo simulate a stochastic set of historical (pre-dam)hydrographs we resampled HL and NAA driversfrom DFFT analysis of daily stage data at StungTreng over the time period 20 years before theonset of construction of the first dam (1941ndash1960)Specifically we reconstructed 1000 unique 20-yearsequences of annual drivers by drawing years(1941ndash1960) from a uniform random distributionand annealing sets of drivers to create sequencesfor each driver that preserved the annual correla-tion between HL and NAA drivers

MARSS simulation framework

Preserving the same model structure specifiedfor the historical record we projected the fisheryforward using the estimated coefficients for dis-charge anomaly effects (in C) and process errorvariance and covariance (in Q) Because in thiscase we were forecasting ldquotruerdquo as opposed toldquoobservedrdquo abundances there was no need for anobservation process (Eq 2) and thus noRmatrixwas used Trajectories of fishery catch were con-strained by the HL metrics described above Weran 100 realizations in each of the 1000 realiza-tions per each scenario (for a total of 100000realizations per scenario) using total catch ateachDai in 2012 as initial valuesWe then pooledabundances across Dai by realization and calcu-lated 25 25 50 75 and 975 percentile catch levelsat each time step (2013ndash2020) across realizationsThus probabilistic distributions of future catchrepresent pooled catches acrossDai but arise fromDai-specific forecasts

REFERENCES AND NOTES

1 E A K Kalitsi Problems and prospects for hydropowerdevelopment in Africa In United Nations Workshop for AfricanEnergy Experts on Operationalizing the NEPAD Energy Initiative(United Nations 2003) pp 1ndash25

2 A Bartle Hydropower potential and development activitiesEnergy Policy 30 1231ndash1239 (2002) doi 101016S0301-4215(02)00084-8

3 X L Chang X H Liu W Zhou Hydropower in China atpresent and its further development Energy 35 4400ndash4406(2010) doi 101016jenergy200906051

4 R E Grumbine J Dore J C Xu Mekong hydropower Driversof change and governance challenges Front Ecol Environ10 91ndash98 (2012) doi 101890110146

5 US Agency for International Development Cambodia CountryDevelopment Cooperation Strategy 2014ndash2018 (2014)

6 US Energy Information Administration International EnergyOutlook 2016 (US Department of Energy 2016)

7 C J Voumlroumlsmarty et al Global threats to human water securityand river biodiversity Nature 467 555ndash561 (2010)doi 101038nature09440 pmid 20882010

8 W L Graf Dam nation A geographic census of American damsand their large-scale hydrologic impacts Water Resour Res35 1305ndash1311 (1999) doi 1010291999WR900016

9 C Nilsson C A Reidy M Dynesius C RevengaFragmentation and flow regulation of the worldrsquos large riversystems Science 308 405ndash408 (2005) doi 101126science1107887 pmid 15831757

10 N L Poff J D Olden D M Merritt D M PepinHomogenization of regional river dynamics by dams and global

biodiversity implications Proc Natl Acad Sci USA 1045732ndash5737 (2007) doi 101073pnas0609812104pmid 17360379

11 J L Sabo et al Reclaiming freshwater sustainability in theCadillac Desert Proc Natl Acad Sci USA 107 21263ndash21270(2010) doi 101073pnas1009734108 pmid 21149727

12 P T J Johnson J D Olden M J Vander Zanden Daminvaders Impoundments facilitate biological invasions intofreshwaters Front Ecol Environ 6 357ndash365 (2008)doi 101890070156

13 D W Seegrist R Gard Effects of floods on trout in SagehenCreek California Trans Am Fisheries Soc 101 478ndash482(1972) doi 101126science27352811558

14 J T Wootton M S Parker M E Power Effects of disturbanceon river food webs Science 273 1558ndash1561 (1996)doi 101126science27352811558

15 N L Poff et al The natural flow regime BioScience 47769ndash784 (1997) doi 1023071313099

16 D A Lytle N L Poff Adaptation to natural flow regimesTrends Ecol Evol 19 94ndash100 (2004) doi 101016jtree200310002 pmid 16701235

17 N L Poff et al The ecological limits of hydrologic alteration(ELOHA) A new framework for developing regionalenvironmental flow standards Freshw Biol 55 147ndash170(2010) doi 101111j1365-2427200902204x

18 B D Richter J V Baumgartner J Powell D P Braun Amethod for assessing hydrologic alteration within ecosystemsConserv Biol 10 1163ndash1174 (1996) doi 101046j1523-1739199610041163x

19 P C D Milly et al Stationarity is dead Whither watermanagement Science 319 573ndash574 (2008) doi 101126science1151915 pmid 18239110

20 M Acreman et al Environmental flows for natural hybrid andnovel riverine ecosystems in a changing world Front EcolEnviron 12 466ndash473 (2014) doi 101890130134

21 N L Poff et al Sustainable water management under futureuncertainty with eco-engineering decision scaling Nat ClimChang 6 25ndash34 (2016) doi 101038nclimate2765

22 M Palmer et al Ecology for a crowded planet Science 3041251ndash1252 (2004) doi 101126science1095780pmid 15166349

23 A Horne et al Using optimization to develop a ldquodesignerrdquoenvironmental flow regime Environ Model Softw 88 188ndash199(2017) doi 101016jenvsoft201611020

24 A Horne et al Optimization tools for environmental waterdecisions A review of strengths weaknesses andopportunities to improve adoption Environ Model Softw84 e338 (2016) doi 101016jenvsoft201606028

25 Food and Agriculture Organization of the United NationsMekong Basin (2016) wwwfaoorgnrwateraquastatbasinsmekongindexstm

26 P J Dugan et al Fish migration dams and loss of ecosystemservices in the Mekong basin Ambio 39 344ndash348 (2010)doi 101007s13280-010-0036-1 pmid 20799685

27 K O Winemiller et al Balancing hydropower and biodiversityin the Amazon Congo and Mekong Science 351 128ndash129(2016) doi 101126scienceaac7082 pmid 26744397

28 R L Welcomme Fisheries Ecology of Floodplain Rivers(Longman 1979)

29 G De Graaf The flood pulse and growth of floodplain fish inBangladesh Fish Manag Ecol 10 241ndash247 (2003)doi 101046j1365-2400200300341x

30 H I Suzuki et al Inter-annual variations in the abundance ofyoung-of-the-year of migratory fishes in the Upper ParanaacuteRiver floodplain Relations with hydrographic attributes BrazJ Biol 69 (suppl) 649ndash660 (2009) doi 101590S1519-69842009000300019 pmid 19738971

31 CGIAR Mekong river facts (2017) httpswle-mekongcgiarorgmekong-river-facts

32 R Stone Mayhem on the Mekong Science 333 814ndash818(2011) doi 101126science3336044814 pmid 21835993

33 H Hori The Mekong Environment and Development (UnitedNations Univ Press 2000)

34 K G Hortle ldquoConsumption and the Yield of Fish and OtherAquatic Animals from the Lower Mekong Basinrdquo MRCTechnical Paper No 16 (Mekong River Commission VientianeLao PDR 2007)

35 A S Halls et al ldquoThe Stationary Trawl (Dai) Fishery of theTonle Sap-Great Lake Cambodiardquo MRC Technical Paper No 32(Mekong River Commission Phnom Penh Cambodia 2013)

36 S E Hampton et al Quantifying effects of abiotic and bioticdrivers on community dynamics with multivariateautoregressive (MAR) models Ecology 94 2663ndash2669 (2013)doi 10189013-09961 pmid 24597213



ugust 20 2020


Dow

nloaded from

37 J M Szemis G C Dandy H R Maier A multiobjective antcolony optimization approach for scheduling environmentalflow management alternatives with application to the RiverMurray Australia Water Resour Res 49 6393ndash6411 (2013)doi 101002wrcr20518

38 J L Sabo D M Post Quantifying periodic stochastic andcatastrophic environmental variation Ecol Monogr 78 19ndash40(2008) doi 10189006-13401

39 E R Holmes E Ward K Wills Package lsquoMARSSrsquo MultivariateAutoregressive State-Space modeling Version 39 (2014)httpcranr-projectorgwebpackageMARSS

40 A Ruhiacute J D Olden J L Sabo Declining streamflow inducescollapse and replacement of native fishes in the AmericanSouthwest Front Ecol Environ 14 465ndash472 (2016)doi 101002fee1424

41 See supplementary materials42 A R Ives B Dennis K L Cottingham S R Carpenter

Estimating community stability and ecological interactionsfrom time-series data Ecol Monogr 73 301ndash330 (2003)doi 1018900012-9615(2003)073[0301ECSAEI]20CO2

43 A Ruhiacute E E Holmes J N Rinne J L Sabo Anomalousdroughts not invasion decrease persistence of native fishes ina desert river Glob Change Biol 21 1482ndash1496 (2015)doi 101111gcb12780 pmid 25430731

44 J Knape P de Valpine Are patterns of density dependencein the Global Population Dynamics Database driven byuncertainty about population abundance Ecol Lett 15 17ndash23(2012) doi 101111j1461-0248201101702x pmid 22017744

45 K P Burnham D R Anderson Model Selection and MultimodelInference A Practical Information-Theoretic Approach (Springer2002)

46 M E Arias T Piman H Lauri T A Cochrane M KummuDams on Mekong tributaries as significant contributors ofhydrological alterations to the Tonle Sap Floodplain inCambodia Hydrol Earth Syst Sci 18 5303ndash5315 (2014)doi 105194hess-18-5303-2014

47 M Kummu J Sarkkula Impact of the Mekong River flowalteration on the Tonle Sap flood pulse Ambio 37 185ndash192(2008) doi 1015790044-7447(2008)37[185IOTMRF]20CO2 pmid 18595273

48 T A Raumlsaumlnen et al Observed river discharge changes due tohydropower operations in the Upper Mekong Basin J Hydrol545 28ndash41 (2017) doi 101016jjhydrol201612023

49 D S Matteson N A James A nonparametric approach formultiple change point analysis of multivariate data J Am StatAssoc 109 334ndash345 (2014) doi 101080016214592013849605

50 N A James D S Matteson ecp An R package fornonparametric multiple change point analysis of multivariatedata J Stat Softw 62 1ndash25 (2014) doi 1018637jssv062i07

51 E R Cook et al Asian monsoon failure and megadroughtduring the last millennium Science 328 486ndash489 (2010)doi 101126science1185188 pmid 20413498

52 T A Raumlsaumlnen C Lehr I Mellin P J Ward M KummuPalaeoclimatological perspective on river basin hydrometeorologyCase of the Mekong Basin Hydrol Earth Syst Sci 17 2069ndash2081(2013) doi 105194hess-17-2069-2013

53 T A Cochrane M E Arias T Piman Historical impact ofwater infrastructure on water levels of the Mekong River andthe Tonle Sap system Hydrol Earth Syst Sci 18 4529ndash4541(2014) doi 105194hess-18-4529-2014

54 G Ziv E Baran S Nam I Rodriacuteguez-Iturbe S A LevinTrading-off fish biodiversity food security and hydropower inthe Mekong River Basin Proc Natl Acad Sci USA 1095609ndash5614 (2012) doi 101073pnas1201423109pmid 22393001

55 D Lamberts J Koponen Flood pulse alterations andproductivity of the Tonle Sap ecosystem A model for impactassessment Ambio 37 178ndash184 (2008) doi 1015790044-7447(2008)37[178FPAAPO]20CO2 pmid 18595272

56 T K Harms N B Grimm Influence of the hydrologic regimeon resource availability in a semi-arid stream-ripariancorridor Ecohydrology 3 349ndash359 (2010) doi 101002eco119

57 J R Welter S G Fisher N B Grimm Nitrogen transport andretention in an arid land watershed Influence of stormcharacteristics on terrestrial-aquatic linkages Biogeochemistry76 421ndash440 (2005) doi 101007s10533-005-6997-7

58 T Meixner et al Influence of shifting flow paths on nitrogenconcentrations during monsoon floods San Pedro RiverArizona J Geophys Res 112 G03S03 (2007) doi 1010292006JG000266

59 C S Holling Adaptive Environmental Assessment andManagement (Wiley 1978)

60 W J Rainboth Fishes of the Cambodian Mekong (Food andAgriculture Organization of the United Nations 1996)

61 H I Jager B T Smith Sustainable reservoir operation Can wegenerate hydropower and preserve ecosystem values RiverRes Appl 24 340ndash352 (2008) doi 101002rra1069

62 E J Barbour et al Optimisation as a process forunderstanding and managing river ecosystems Environ ModelSoftw 83 167ndash178 (2016) doi 101016jenvsoft201604029

63 M J Kuby W F Fagan C S ReVelle W L Graf Amultiobjective optimization model for dam removal Anexample trading off salmon passage with hydropower andwater storage in the Willamette basin Adv Water Resour 28845ndash855 (2005) doi 101016jadvwatres200412015

64 B Dennis M L Taper Density dependence in time seriesobservations of natural populations Estimation and testingEcol Monogr 64 205ndash224 (1994) doi 1023072937041

ACKNOWLEDGMENTS

Supported by MacArthur Foundation award 100717 and NSF grantDEB-1541694 (GWH) NSF grant CBET-1204478 (JLS) NSFgrant EAR-1740042 (JLS and GWH) the National Socio-Environmental Synthesis Center (SESYNC) with funding from NSFgrant DBI-1052875 (AR) MAAndashJa Vesitekniikan Tuki (TAR) andMacArthur Foundation awards 100718 and 15-108455-000-CSD toConservation International All authors thank the Mekong RiverCommission and the Inland Fisheries Research and DevelopmentInstitute (IFReDI) of Cambodia for access to data and C PhenT Cochrane M Cooperman C Costello T Farrell L HannahL Kauffman K McCann E Moody T Pool N Rooney and S Limfor helpful comments on the design of the research or previousversions of the manuscript Illustrations were created by S Deviche(Fig 1) and M Northrop (Fig 6) Arizona State University Dataused in this analysis are presented in the supplementary materialsand are available in raw form by request from the Mekong RiverCommission by writing to SN (sonammrcmekongorg)

SUPPLEMENTARY MATERIALS

wwwsciencemagorgcontent3586368eaao1053supplDC1Supplementary TextFigs S1 to S10Tables S1 to S3Data S1 to S11

17 June 2017 accepted 27 October 2017101126scienceaao1053



ugust 20 2020


Dow

nloaded from

Designing river flows to improve food security futures in the Lower Mekong BasinJ L Sabo A Ruhi G W Holtgrieve V Elliott M E Arias Peng Bun Ngor T A Raumlsaumlnen and So Nam

DOI 101126scienceaao1053 (6368) eaao1053358Science

this issue p eaao1053 see also p 1252Sciencethat could help to deliver food security in other river systemsSuch data-driven approaches can be used to link hydrology to ecology and food production and specify design principles Poff and Olden) Fish futures can be maximized within a managed hydrologic system with careful prescription of flowsfeatures of the flow regime that optimize the fishery that is crucial to food security in Cambodia (see the Perspective by

used a data-based time series modeling approach to estimate theet alfood security For the Mekong River Sabo of the waters downstream Where fisheries in large tropical river systems are affected there can be knock-on effects on

Hydropower dams radically alter river flow regimes often with consequences for the functioning and productivityOptimizing flow in dammed rivers

ARTICLE TOOLS httpsciencesciencemagorgcontent3586368eaao1053

MATERIALSSUPPLEMENTARY httpsciencesciencemagorgcontentsuppl201712063586368eaao1053DC1

CONTENTRELATED

httpsciencesciencemagorgcontentsci3646444eaav9887fullhttpsciencesciencemagorgcontentsci3646444eaav8755fullhttpsciencesciencemagorgcontentsci3616398eaat1477fullhttpsciencesciencemagorgcontentsci3616398eaat1989fullhttpsciencesciencemagorgcontentsci3616398eaat1225fullhttpsciencesciencemagorgcontentsci35863681252full

REFERENCES

httpsciencesciencemagorgcontent3586368eaao1053BIBLThis article cites 50 articles 10 of which you can access for free

PERMISSIONS httpwwwsciencemagorghelpreprints-and-permissions

Terms of ServiceUse of this article is subject to the

is a registered trademark of AAASScienceScience 1200 New York Avenue NW Washington DC 20005 The title (print ISSN 0036-8075 online ISSN 1095-9203) is published by the American Association for the Advancement ofScience

Science No claim to original US Government WorksCopyright copy 2017 The Authors some rights reserved exclusive licensee American Association for the Advancement of

on August 20 2020

httpsciencesciencem

agorgD

ownloaded from

RESEARCH ARTICLE

SUSTAINABILITY

Designing river flows to improvefood security futures in theLower Mekong BasinJ L Sabo123 A Ruhi24 G W Holtgrieve5 V Elliott6 M E Arias7

Peng Bun Ngor8dagger T A Raumlsaumlnen9 So Nam8Dagger

Rivers provide unrivaled opportunity for clean energy via hydropower but little is knownabout the potential impact of dam-building on the food security these rivers provide Intropical rivers rainfall drives a periodic flood pulse fueling fish production and deliveringnutrition to more than 150 million people worldwide Hydropower will modulate this floodpulse thereby threatening food security We identified variance components of theMekong River flood pulse that predict yield in one of the largest freshwater fisheries inthe world We used these variance components to design an algorithm for a managedhydrograph to explore future yields This algorithm mimics attributes of dischargevariance that drive fishery yield prolonged low flows followed by a short flood pulseDesigned flows increased yield by a factor of 37 relative to historical hydrologyManaging desired components of discharge variance will lead to greater efficiency in theLower Mekong Basin food system

Hydropower is currently the dominant sourceof renewable energy worldwide (1ndash4) andis poised to power some of the poorest pre-dominantly rural populations in the world(5 6) Storage of streamflow in reservoirs

and dam operations often dampen peak flowsdeliver higher base flows and change the rangeand frequency of discharge variability in riversworldwide (7ndash11) This hydrologic alterationmdashfrom dams and other stressorsmdashpromotes theinvasion of non-native aquatic species (12 13)alters food web structure (14) and reduces bio-diversity by homogenizing regional freshwaterfaunas (10) Traditional river conservation underthe natural-flow regime paradigm (15) posits thatthe restoration of pre-dam flow variability max-

imizes ecological outcomes (15ndash18) Althoughmany empirical studies support this hypothesisremoval of large dams and restoration of pre-dam conditions is rarely possible especially forsystems with large hydropower dams or storagereservoirs that have long lifespans and are centralto socio-environmental water systems (19)In tropical basins construction of mainstem

and tributary dams is imminent and requires fore-sight about how these new facilities are imple-mented In these cases strategic operation ofexisting and future infrastructure to deliver theright dose of variation via smart control (20ndash22)may be a more effective and immediately de-ployable strategy to deliver human and ecologicaloutcomes This recent paradigm shift is fuelingmuch research on algorithm development to pre-scribe ecological objective functions [eg (23)]and optimize these with other competing objec-tive functions to harness water for power genera-tion irrigation and other human uses [reviewedin (24)] The framework for optimization is richbut the algorithms for prescribing an ecologicalobjective function are generally less well de-veloped Here we address this limitation usinga data-driven time-series approach relating hydro-logic variation to fishery yields and we providetangible alternatives in the context of hydropowerdevelopment and inland fishery production inthe Lower Mekong Basin (LMB)The Mekong is the eighth largest river by dis-

charge and hosts one of the largest inland fish-eries in the world (25) The proposed scope ofhydropower development in the Mekong andother tropical rivers in Asia Africa and SouthAmerica poses pronounced trade-offs for bio-

diversity and the production of freshwater fishfor food (26 27) In the LMBmonsoon rains drivea flood pulse that inundates floodplain habitatsthroughout the basin In flood-pulse fisheriesthe duration andmagnitude of the flood pulsemdashwhich controls the spatial extent of inundation(28)mdashis a key driver of production (28) but thereis some support for the notion that flood timingespecially the relative sequence of weak andstrong flood pulses is important (29 30) Thissuggests that regulation of the timing and mag-nitude of the high and low flows could be har-nessed to improve the fisheries However theexact features of a hydrograph that could pro-mote fishery yields in the LMB and how thesemight be translated into actionable design prin-ciples remain unknownWe tested the historic role of multiple var-

iance components of the LMB hydrograph indriving harvest of fishes (total annual harvestin kilograms standardized by effort) from thebag net or ldquoDairdquo fishery on the Tonle Sap RiverThe river connects the Mekong to the Tonle SapLake the largest lake and wetland in SoutheastAsia and a nursery habitat for as many as 300species of fishes (25 31) many of which provide akey source of animal protein and vitamin A forrural subsistence fishing and agricultural com-munities (32ndash34) The Dai fishery is one of themost valuable and productive commercial fish-eries in the LMB (35 ) and has the longest andmost complete record of fishery catches in theregion with biomass and species compositiondata spanning 17 years (1996ndash2012) from 64 Daisat 14 locations along the Tonle Sap RiverOur analysis expands on previous research in

at least three ways (i) Our data-driven ratherthan mechanistic modeling approach allows usto connect key aspects of hydrology to fisherydynamics while minimizing assumptions arisingfrom specified eco-physiological parameters (36)Briefly we leverage observed relationships be-tween variance components of recent hydrologyand harvest to identify shape features in thehydrograph that should maximize desired out-comes for the future fishery Hence our analysisprovides a path for scientifically informed pro-active management of a flood-pulse fishery insystems in which a reasonable understanding ofthe biological mechanisms at play is not withinreach (ii) We focus on design of future hydrologyvia ecologically informed dam operations Thisfocus is timely in the Mekong given the im-minence of extensive hydropower in the LMBand the potential to undertake this developmentin ways that minimize the impact on the fishery(and thus on regional food security) The methodfor design that we propose leverages a highlyflexible and well-understood spectral toolboxthat would have direct transferable applicabilityto systems engineers who operate dams In shortthe design provides a potential tool for manag-ing mainstem hydrology for fisheries as newdams come on line (iii) We articulate a set ofecological design principles that achieve socio-ecological goals by modifying the variance withinan annual hydrograph without increasing mean

RESEARCH


1Future H2O Knowledge Enterprise Development ArizonaState University Tempe AZ 85287 USA 2Julie Ann WrigleyGlobal Institute of Sustainability Arizona State UniversityTempe AZ 85287 USA 3Faculty of Ecology Evolution andEnvironmental Science School of Life Sciences Arizona StateUniversity Tempe AZ 85287 USA 4National SocialEcological Synthesis Center (SESYNC) University ofMaryland Annapolis MD 21401 USA 5School of Aquatic andFishery Sciences University of Washington Seattle WA98105 USA 6Betty and Gordon Moore Center for ScienceConservation International Arlington VA 22202 USA7Department of Civil and Environmental EngineeringUniversity of South Florida Tampa FL 33620 USA 8MekongRiver Commission Secretariat Chak Angre Krom MeancheyPhnom Penh Kingdom of Cambodia 9Department of BuiltEnvironment Aalto University 00076 Aalto FinlandCorresponding author Email johnlsaboasuedudaggerPresent address CNRS Universiteacute Toulouse III Paul Sabatier ENFAUMR5174 EDB (Laboratoire Eacutevolution amp Diversiteacute Biologique)F-31062 Toulouse France DaggerPresent address Mekong RiverCommission Secretariat Ban Sithane Neua Sikhottabong DistrictVientiane 01000 Lao PDR

on August 20 2020

httpsciencesciencem

agorgD

ownloaded from




Discharge variance













ugust 20 2020


Dow

nloaded from










Density dependence






ugust 20 2020


Dow

nloaded from
















HSAM 0522 0322 0119 0556 0088

HSAF 0478 0187 0070 0196 0177

LSAM 0151 0066 0085 0080 0051




ugust 20 2020


Dow

nloaded from



Breakpoint analysis










nfrac141

an cos2p ftn Xinfin

nfrac141

bn sin2p ftn








ugust 20 2020


Dow

nloaded from
























ugust 20 2020


Dow

nloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from




Discharge variance













ugust 20 2020


Dow

nloaded from










Density dependence






ugust 20 2020


Dow

nloaded from
















HSAM 0522 0322 0119 0556 0088

HSAF 0478 0187 0070 0196 0177

LSAM 0151 0066 0085 0080 0051




ugust 20 2020


Dow

nloaded from



Breakpoint analysis










nfrac141

an cos2p ftn Xinfin

nfrac141

bn sin2p ftn








ugust 20 2020


Dow

nloaded from
























ugust 20 2020


Dow

nloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from










Density dependence






ugust 20 2020


Dow

nloaded from
















HSAM 0522 0322 0119 0556 0088

HSAF 0478 0187 0070 0196 0177

LSAM 0151 0066 0085 0080 0051




ugust 20 2020


Dow

nloaded from



Breakpoint analysis










nfrac141

an cos2p ftn Xinfin

nfrac141

bn sin2p ftn








ugust 20 2020


Dow

nloaded from
























ugust 20 2020


Dow

nloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from
















HSAM 0522 0322 0119 0556 0088

HSAF 0478 0187 0070 0196 0177

LSAM 0151 0066 0085 0080 0051




ugust 20 2020


Dow

nloaded from



Breakpoint analysis










nfrac141

an cos2p ftn Xinfin

nfrac141

bn sin2p ftn








ugust 20 2020


Dow

nloaded from
























ugust 20 2020


Dow

nloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from



Breakpoint analysis










nfrac141

an cos2p ftn Xinfin

nfrac141

bn sin2p ftn








ugust 20 2020


Dow

nloaded from
























ugust 20 2020


Dow

nloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from
























ugust 20 2020


Dow

nloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from





Methods












ugust 20 2020


Dow

nloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from













ugust 20 2020


Dow

nloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from















eth5THORN




ugust 20 2020


Dow

nloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from






















































ugust 20 2020


Dow

nloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from





























ACKNOWLEDGMENTS







ugust 20 2020


Dow

nloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from








CONTENTRELATED


REFERENCES






on August 20 2020

httpsciencesciencem

agorgD

ownloaded from

sustainability designing river flows to improve food...

Documents