Ecological Forecasts: An merging Imperative James S. Clark,'* Steven R. Carpenter: Mary Barber: Scott Collins: Andy Dobs~n,~Jonathan A. F~ley,~ David M. L~dge,~Mercedes Pascual: Roger Pielke Jr.,' William Pizer,1° Cathy Pringle," Walter V. Reid," Kenneth A. Rose,13 Osvaldo Sala,14 William H. ~chlesinger,'~Diana H. David wear"

Planning and decision-making can be improved by access to reliable for which uncertainty can be reduced to the forecasts of ecosystem state, ecosystem services, and natural capital. point where a forecast reports a useful Availability of new data sets, together with progress in computation and amount of information. Information content statistics, will increase our ability to forecast ecosystem change. An is affected by all sources of stochasticity. agenda that would lead toward a capacity to produce, evaluate, and communicate forecasts of critical ecosystem services requires a process that engages scientists and decision-makers. Interdisciplinary linkages are 'Department of Biology, Duke University, Durham, necessary because of the climate and societal controis on ecosystems, the NC 27708 USA. 'University of Wisconsin Center for feedbacks involving social change, and the decision-making relevance of Limnology, Madison, WI 53706, USA. 3Ecological So- ciety of America, 1707 H Street NW, Suite 400, forecasts. Washington, DC 20006, USA. 'Division of Biology, Kansas State University. Manhattan, KS 66506, USA. Scientists and policy-makers can agree that ecosystem services, and natural capital, with SDepartment of Ecology and Evolutionary Biology, success in dealing with environmental change fully specified uncertainties, and is contin- Princeton University, Princeton, N] 08544, USA. Ten- ter for Sustainability and the Global Environment, rests with a capacity,to anticipate. Rapid gent on explicit scenarios for climate, land University of Wisconsin, Madison, Wl 53706, USA. change in climate and chemical cycles, de- use, human population, technologies, and 'Department of Biological Sciences, University of pletion of the natural resources that support economic activity. The spatial extent ranges Notre Dame, Notre Dame, IN 46556, USA BDepart- regional economies, proliferation of exotic from small plots to regions to continents to ment of Ecology and Evolutionary Biology, University of Michigan, Ann Arbor, MI 48109, USA. 'Environmen- species, spread of disease, and deterioration the globe. The time horizon can extend up to tal and Societal Impacts GrouplNational Center for of air, waters, and soils pose unprecedented 50 years. The information content of a fore- Atmospheric Research. 3250 Mitchell Lane. Boulder, threats to human civilization. Continued cast is inversely proportional to forecast un- CO 80301. USA. loResources for the Future, 1616 P food, fiber, and freshwater supplies and the certainty (2). A 'wide confidence envelope Street NW, Washington, DC 20036, USA. "Depart- maintenance of human health depend on our indicates low information content. A scenario ment of Ecology, University of Georgia, Athens, GA 30602. USA. '' Millennium Ecosystem Assessment. ability to anticipate and prepare for the un- assumes changes in "possible future bound- 731 North 79th Street, Seattle, WA 98103, USA. certain future (1). Anticipating many of the ary conditions (e.g., emissions scenarios). . . . 13Coastal Fisheries Institute and Department of environmental challenges of coming decades For the decision maker, scenarios provide an Oceanography and Coastal Sciences, Louisiana State requires improved scientific understanding. indication of possibilities, but not definitive University, Baton Rouge, !A 70803, USA. "Depart- ment of Ecology, Faculty of Agronomy-IFEVA, Uni- An evolving science of ecological forecasting probabilities" (3). Scenarios can be the basis for versity of Buenos Aires-CONICET, Buenos Aires 1417, is beginning to emerge and could have an projections, which apply the tools of ecological Argentina. lSNicholas School of Environment and expanding role in policy and management. forecasting to specific scenarios. Earth Sciences, Duke University, Durham, NC 27708, An initiative in ecological forecasting USA. 16Natural Resource Ecology Laboratory, Colo- rado State University, Fort Collins, CO 80523, USA. must define the appropriate role of science in What Is Forecastable? l7United States Department of Agriculture Forest Ser- the decision-making process and the research Accurate estimation and communication of vice, Post Office Box 12254, Research Triangle Park, that is required to develop the capability. information content will determine the suc- NC 27709. USA. Ecological forecasting is defined here as the cess of an ecological forecasting initiative. *To whom correspondence should be addressed. E- process of predicting the state of ecosystems, "Forecastable" ecosystem attributes are ones mail: [email protected]

www.sciencernag.org SCIENCE VOL 293 27JULY 2001 657 ECOLOGY THROUGH TIME Low information content can result because plant composition and reduced carbon stor- predictable from fine-grained studies (23, drivers (and, thus, model structures) are un- age potential in tallgrass prairie soils (15). 24). The feedbacks from vegetation to cli- certain, parameters are uncertain, and un- Knowledge of fertilizer and irrigation effects mate become important only when the spa- known human responses to ecosystem change on carbon storage in agroecosystems can be tial extent of a study exceeds a critical (or to forecasts of ecosystem change) affect used to forecast how managed ecosystems threshold. Factorial, whole-ecosystem ex- outcomes. Many sources of stochasticity are will contribute to or stem the future rise of periments with CO,, temperature, moisture, typically ignored in ecological models. When CO, in Earth's atmosphere (16). and nutrients may be the only way to de- reported at all, prediction uncertainties are Analysis of projections can help anticipate termine forest responses to global change typically confined to estimation error (4, 5), change, even where forecasts are uninforma- (25). For example, free-air CO, enrichment which is reduced by sampling and is often tive. Although forecasts of population migra- (FACE) studies show that the water stress overwhelmed by other sources of uncertainty. tion rates will typically have low information expected from studies of individual plants Most daunting is the "inherent" uncertain- content, analysis shows that productive re- may not be realized in an intact stand (26). ty that results from strong nonlinearities and search will focus on factors affecting inva- Data networks can provide a baseline for stochasticity. For example, the inherent un- sion potential, such as the mechanisms of forecasting. Missing variables, low resolu- certainty involved in extinction risks leads long-distance dispersal and propagule pro- tion, inadequate duration, temporal and spa- ecologists to disagree on the value of predic- duction, as opposed to precise estimation of tial gaps, and declining coverage are perva- tions from population viability models (6). long-distance dispersal (8). Rates will remain sive limitations. Due to abandonment of pre- Extinction forecasts are highly sensitive to uncertain, but we may improve our ability to cipitation, stream-height, and discharge gaug- poorly constrained assumptions (7). Inherent predict introduced species that can success- es, the capacity to forecast droughts and uncertainty will always limit informative fully invade (17). floods was greater 30 years ago than it is forecasts of spread velocity for invasive The developing capacity for prediction today. Countries with the poorest hydrologi- plants with high reproductive rates. Even pre- requires careful model evaluation, which can cal networks (e.g., sub-Saharan Africa, arid cise knowledge of parameters that might be involve model selection, m+l averaging, or regions of the former Soviet Union) have the estimated, for example, through detailed both. Model selection methods are routinely most pressing water needs (27). The problem study of long-distance dispersal, would do used in ecological applications. Because the is not restricted to developing and transitional little to increase forecast information (8). models themselves are often uncertain, eco- economies. There is an average density of Large inherent uncertainty does not nec- logical forecasting may eventually rely more one stream gauge per 1024 km2 in the lower essarily neutralize efforts to anticipate heavily on model averaging. Techniques for 48 states of the United States (28). Since change. Forecasting will improve as ecolo- model evaluation developed in econometrics, 197 1 there has been a 22% decline in gauging gists identify the "slow" variables that fore- finance, and meteorology make use of hind stations that record flow on small U.S. rivers. warn of consequences years in advance. casting (la), including the ability to identify Sustained monitoring is needed that can Whereas deterministic weather forecasts con- turning points and events (12). dovetail with forecasts in an adaptive feed- front an approximate 2-week limit, probabi- Failing to accommodate the important back design. listic climate prediction makes use of the sources of stochasticity makes for a forecast The ability to anticipate exotic invasions system memory represented by sea-surface that contains less information than it purports would benefit from historic records of species temperatures. The limitations imposed on a (confidence intervals are misleadingly nar- introductions and their vectors (e.g., ship traf- deterministic weather forecast by nonlineari- row). In the case of western North America's fic). Where eventual colonization seems in- ties may not defeat efforts to provide infor- Northern Spotted Owl (Strix occidentalis), evitable, forecasts may guide mitigative ac- mative climate forecasts (9). There are many confidence intervals on population growth tions. Disease forecasting can also require "slow variables" that constrain ecological rates became basis for policy (19). Ecological extensive spatial and temporal data, such as processes (10). For example, successional models typically ignore variability among in- those used to inform intervention for foot- change in forests is constrained by climate dividuals, which is large and has impact on and-mouth disease (29). Prediction of child- and soils. If these change slowly relative to population growth and decline. New compu- hood epidemics depends on long records of tree life-spans, successjon is predictable us- tational approaches represented by hierarchi- births and vaccinations (30). Cholera and ma- ing physiology and competitive interactions cal models accommodate multiple stochastic laria predictions require climate data, which among trees (5, 11). Land-use change is de- elements (20) and can be used to estimate the determine growth andlor spread of pathogens termined by individual decisions that are in- uncertainty in growth of populations having and vectors (31). fluenced by a variety of uncertain needs and variability among individuals (21). New ap- Developing technologies do not fully goals. Yet decade-scale land-cover change plications of these recent techniques are used compensate for sparse data, but they promise can be predictable based on overriding con- in weather and climate models (22), but they to facilitate forecasting. Hydrologic forecast- trols imposed by topography and distance to are not exploited by ecologists. .1nevitabie ing and remote sensing, together with geo- market centers (12). failures that result from forecast uncertainties physical tomography, can provide high-reso- Agricultural practices result from com- that are unrealistic would eventually erode lution coverage of precipitation and the ef- plex decisions, but slow variables can be the confidence (9). fects of dams and irrigation (32). Biogeo- basis for useful projections. Projections of chemical cycles, hydrology, and biodiversity subsidies to global food production (irriga- Data from Experiments and forecasts require land inventory and census tion, fertilizers, and transport and storage of Monitoring data (33) in combination with satellite-based crops) (13) can inform forecasts of down- Technical construction of forecasts requires data (34). Satellites could be used to monitor stream eutrophication in coastal fisheries and initiatives to develop new or augment ex- habitat loss, a predictor of extinction risk. increases in atmospheric greenhouse gases isting data networks and to support exper- Satellite data could be used to develop (CH,, CO,, and N,O) (14). Ecologists can imental research. Experimental and obser- global scenarios for disease spread in re- forecast how environmental change affects vational data that extend to landscapes or sponse to environmental degradation and cli- carbon storage in agriculture, by production regions are a foundation for forecasting mate change (35). Prevalence of hantavirus forestry, and in natural ecosystems. Nitrogen capability. Large experiments are critical, pulmonary syndrome @IPS), a viral disease deposition leads to predictable changes in because landscape processes are often un- characterized by acute respiratory distress

658 27 JULY 2001 VOL 293 SCIENCE www.sciencernag.org ECOLOGY THROUGH TIME . that has a high death rate, depends on the ban of DDT and the Kyoto discussions on tation of fuels and construction materials. infection rates of its host, the common deer greenhouse gasses. Policy-makers can re- Ecologists could have foreseen that the mouse (Peromyscus maniculatus) (36). The spond to research that is motivated by man- floods of Hurricane Floyd would release 1993 HPS outbreak in the Southwest United agement or conservation interests. For exam- hog waste into North Carolina rivers and States was attributed to unusual weather of ple, population studies, together with 30-year sounds. Ecological forecasting may target 1991-1992 that was quantified from Landsat discharge records, were used by the Puerto the vulnerabilities that decision-makers Thematic Mapper satellite imagery. A model Rican Aqueduct and Sewage Authority to must consider, if not the events themselves. developed for the 1993 outbreak, which fol- develop a system for water withdrawal from lowed an El Niiio year, provided accurate streams to meet human demands while min- Next Steps predictions for the 1995 non-El Niiio year. imizing' the loss of migrating freshwater Linking science with decision-making will Likewise, surveillance networks could im- shrimp (45). depend on scientific accuracy and effective prove understanding of climate constraints on Ecologists should increasingly consider communication. Sources of uncertainty, malaria and its vectors (37,38) and of climate their own role in the decision-making pro- their potential impacts on forecast informa- events that forewarn of cholera risk (31). cess. "Bet-hedging" uncertainty may in- tion, and the identification of overriding volve choosing policies that are relatively controls that change slowly must be con- Forecasts in Decision-Making insensitive to uncertainty, that increase the sidered when deciding where efforts can be A 1981 report (39) predicting that the Eur- ability of ecosystems to provide services of most value. Two broad classes of rec- asian zebra mussel (Dreissena polymorpha) even if a surprise occurs, or both. Ecolo- ommendations address these goals. First is would become established in North America gists can help develop options. For exam- a definition of forecasting priorities gained the attention of neither policy-makers ple, maintaining local species diversity and through dialogue involving scientists, man- nor the general public. Zebra mussels were heterogeneity of land cover may stabilize agers, and policy-makers. Priorities are discovered 5 years later and soon spread regional primary production despite uncer- based on potential benefits balanced throughout the upper Midwest. In the Great tain changes in climate. Limnologists have against costs of business as usual. They Lakes alone, annual mitigation costs to indus- shown that optimal nutrient loadings to should meet user needs and be scientifical- try of $20 to $100 million (38) will continue lakes decrease if thtvinformation content of ly feasible. into the foreseeable future. Unquantified, ecological forecasts is taken into account The second recommendation involves noncommercial costs include losses of biodi- (46). Ecologists have found correctives for definition of a science agenda that includes versity, such as the extirpation of native eutrophication that offer managers a num- (i) identifying data and research needs and clams (40), and shifts in ecosystem energy ber of options. (ii) setting priorities for estimation, propaga- flows and productivity (41). No regulatory In situations where uncertainties are large tion, and communication of uncertainty. Fo- actions can be traced to the 1981 prediction. and impossible to quantifjr, information con- cus should be on the problems for which The invasion itself prompted a flurry of reac- tent is necessarily low and decisions can be forecasts are now possible and those that are tive legislation, culminating in the Nonindig- complex. Rarely can policies direct an out- not presently forecastable but could become enous Aquatic Nuisance Prevention and Con- come. Instead, they are often designed to forecastable within a decade. / trol Act of 1990. affect outcomes by influencing choices made The zebra mussel experience highlights by vast numbers of people. The effects can References and Notes issues concerning the state of environmen- extend beyond their intended targets and even 1. Supplemental text is available at Science Online at www.sciencemag.org/cgilcontentlfu11/293/5530/ tal science and its place in planning for have countervailing impacts. For example, 657lDC1. global change. A developing capacity for restrictions on tree harvest in one region can 2. Adapted from the notion of (generalized) Fisher in- prediction has not yet been integrated as lead to intensified harvesting elsewhere, as formation (as opposed to the Information Theory definition as the log-Likelihood ratio). part of a comprehensive prediction process trade offsets local scarcity. Thus, environ- 3. M. MacCracken, www.~ig.uar.edulsocaspIrine/26/ (9, 42). Missed opportunities to engage mental restrictions can lead to export of en- gueahtml. ecological understanding have become a vironmental hazard from one jurisdiction to 4. R. Lande, Am. Nat 130, 624 (1997). 5. 5. W. Pacala et dl.. Ecol. Monogr.66, 1 (1996). source for growing _concern. The zebra another. 6. 8. W. Brook et dl., Nature 404, 385 (2000). mussel experience illuhtes that the $138 When reaction to anticipated change is pos- 7. D. Ludwig, Ecology 80, 298 (1999). billion spent annually on control of nonin- sible, it is appropriate to explore scenarios that 8. 1. 5. Clark, M. Lewis. L Hofvath. Am. Nat. 157, 537 digenous species @IS) (43) can be blamed, are as consistent as possible with current scien- (2001). 9. Nat~onalResearch Council, Making Climate Forecasts in part, on failure to communicate. Fore- tific understanding but are not predictions (47, Matter (National Academy Press. Washington. DC. casts based solely on scientific objectives 48). Scenarios can embrace ambiguous and un- 1999). have little impact on policy (44) because controllable drivers, such as climate or global- 10. S. R. Carpenter, M. G. Turner, Ecosystems 3. 495 (2W). there is no stakeholder (9). Climate change ization of markets, and nonlinear and unpredict- 11. H. H. Shugart. A Theory of Forest Dynamics: The forecasts developed under the Intergovern- able dynamics, such as the reflexive responses Ecological lmplications of Forest Succession ModeLs mental Panel on Climate Change have been of people. Scenarios provide insight into drivers (Springer-Veriag, New York, 1984). 12. D. N. Wear, P. Bolstad, Ecosystems 1. 575 (1998). influential, in part, because they respond to of change, implications of current trajectories, 13. D. Tilman et dl., Science 292, 281 (2001). a request from governments. Priorities for and options for action. Alternative policies can 14. C. P. Robertson, E. A Paul. R R Harwood. Science ecological forecasting must come from di- be considered in light of contrasting scenarios 289. 1922 (2000). 15. D. Wedin. D. Tibnan. Science 274. 1720 (1996). alogue that ensures active participation by and to compare their robustness to possible 16. W. H. Schlesinger, Science 284, 2095 (1999). policy-makers, managers, and the general futures. 17. C. 5. Kolar, D. M. Lodge, Trends Ecol. Evol. 16, 199 public. Ecologists may provide decision-makers (2001). 18. C. C. judge, W. E. Criffiths, R. C. Hill, W. E. Lutkepolh, Some experience suggests that a proactive with information as part of an integrated T. C. Lee. The Thwry and Practice of Econometrics approach holds promise. Chlorofluorocarbon perspective of vulnerability to extreme (Wiley and Sons, New York. 1982). use has declined, in part, due to the Montreal events and their potential consequences. 19. 5. P. Hamson, A. Stahl D. Doak. Consew. Biol. 7.950 Protocol, which was drafted in response to For example, the tragic human toll of Hur- (1993). 20. 8. P. Carlin. T. A. Louis. Bayes and Empirical Bayes scenarios for ozone-depleting chemicals in ricane Mitch in Central America was exac- Methods for Data Analysis (Chapman & Hall Boca the atmosphere. Scenarios helped propel the erbated by degradation due to overexploi- Raton, FL. 2000).

www.sciencemag.org SCIENCE VOL 293 27 JULY 2001 659 ECOLOGY THROUGH TIME

21. Hierarchical modeling allows for simplification of 30. J. D. Eam, P. Rohani. B. M. Bolker. B. T. Crenfell 40. D. L Strayer. J. North Am. Bentholog. Soc. 18, 74 models that have multiple sources of stochasticity by Science 287, 667 (2000). (1999). factoring to produce a set of conditional distribu- 31. M. Pascual, X. Rodo, 5. P. Ellner, R. Colwell M. J. 41. T. F. Nalepa, C. L Fahnenstiel, I. Great Lakes Res. 21, tions. For instance, in population dynamics, the n Bouma. Science 289. 1766 (2000). 41 1 (1995). members of a population can together define an 32. Grand Challenges in Environmental Sciences: Special 42. D. Sarewitz, R. A. Pielke Jr.. R. Byerly Jr.. Prediction in (intractable) n dimensional distribution of demo- Report of the National Academy of Sciences (National Science and Policy (Island Press, Washington, DC, graphic rates. A hierarchical structure renders the Academy of Sciences. Washington. DC, 2000). Avail- 2000). problem analyrable by factoring it into n conditional able at: www.nap.edu/openbooW0309072549/html. 43. D. Pimeritel L Lach, R. Zuniga. D. Morrison. Bio- distributions that submit to Markov Chain Monte 33. 1. F. Richards, in The Earth as Transformedby Human Science 50, 53 (2000). Carlo integration (79). J. S. Clark, in preparation. Action, B. L. Tumw et al.. Eds. (Cambridge Univ. Press, 44. D. Sarewitz, R A. Pidke Jr., in (42) pp. 11-22. 22. Hierarchical models developed to estimate the many New York. 1990), pp. 163-178. 45. J. P. Benstead. J. C. March. C. M. Pringle. F. N. Scatena, parameters needed for dimate modeling can be Ecol. Appl. 9, 656 (1999). adapted for estimation of invarion speed [C. Wikle, 34. N. Ramankutty, j. Foley. Global Biogeochem. Cycles R F. Milliff, D. Nychka, L M. Berliner, I. Am. Stat. 13, 997 (1999). 46. 5. R. Carpenter, D. Ludwig, W. k Brock, Ecol. Appl. 9, hsoc. 96, 382 (ZOO?)]. 35. Under the Weather: Climate, Ecosystems, and lnfec- 751 (1999). 23. S. R. Carpenter, Ecology 77. 677 (1996). tious Disease (National Academy of Sciences. Wash- 47. P. Raskin. C. Gallopin, P. Cutman, A Hammond, R 24. J. S. Clark et al., Am. 1. Bat. 86, 1 (1999). ington, DC. 2001). Swart. Bending the Curve: Toward Global Sustainabil- 25. P. B. Reich et at.. Nature 410, 809 (2001). 36. C. E. Glass et aL. Emerging infect. Dis. 6. 238 (2000). ity, Pole Star Report no. 8 (Stockholm Environment 26. D. S. Ellsworth, Tree Physiol. 20, 435 (2000). 37. C. Dye, P. Reiter, Science 289, 1697 (2000). Institute, Stockholm. 2000). 27. E. Stokstad. Science 285. 1199 (1999). 38. D. J. Rogers. 5. E. Randolph, Science 289.1763 (2000). 48. N. Nakicenovic, Emissions Scenarios (Cambridge Univ. 28. T. Brabets. Evaluation of the Streamflow-Caging Net- 39. Bio-Environmental Services, The Presence and Impli- Press, Cambridge, 2000). work of Alaska in Providing Regional Streamflow In- cation of Foreign Organisms in Ship Ballast Waters 49. Supported by the Ecological Sw'ety of America, the formation, U.S. Geological Survey, Water Resources Discharged into the Great Lakes, vok. 1 and 2, pre- Aldo Lwpold Leadership Program, the National Sci- Investigations Report 96-4001 (1996). pared for the Water Pollution Control Directorate, ence Foundation, the National Center for Ecological 29. N. Ferguson, K. Donelly, R. M. Anderson, Science 292, Environmental Protection Service (Environment Can- Analysis and Synthesis, and the Center for Global 1155 (2001). ada. Ottawa, Ontario, 1981). Change at Duke University.

Mind the gap. NEW! Science Online's Content Alert Service With Science's Content Alert Service, European subscribers (and those around the world) can eliminate the information gap between when Science publishes and when it arrives in the post. This free enhancement to your Science Online subscription delivers e-mail summaries of the latest news and research articles published each Friday in Science - instantly. To sign up for the Content Alert service, go to Science Online and eliminate the gap.

660 27 JULY 2001 VOL 293 SCIENCE www.sciencemag.org