Reservoir fish habitats have been in decline because of reservoir aging. Projected changes in climate are likely to accelerate this decline in regions expected to be impacted by shifts in temperature and precipitation. Change in climate may result in irrecoverable loss of fish habitat and shifts in fish assemblage composition in some reservoirs. However, depending on various reservoir attributes (e.g., depth, catchment size and composition, reservoir age) and their influence, reservoirs may differ greatly in their vulnerability to climate change.
Vulnerability may be defined here as the extent to which a reservoir is susceptible to, or unable to cope with, adverse effects of climate change, including climate shifts, variability, and extremes. Vulnerability is a function of the character, magnitude, and rate of climate change to which a reservoir is exposed and the reservoir’s sensitivity and adaptive capacity to such changes. The complexity of vulnerability encompasses a plethora of attributes and may be abstracted with an index. An index of vulnerability is a quantitative indicator of the relative vulnerability of a system and may often be easier for managers to interpret and may be more useful in identifying reservoirs at risk.
This research project will provide scientific information to guide future reservoir fish habitat projects and decision-making for the Partnership/Friends of Reservoirs.
Because time, funding, and personnel are limited, it is critical for managers to direct resources towards reservoirs where the investment has the greatest likelihood of maintaining desired outputs at the least cost. Determining which reservoirs are most vulnerable enables managers to set priorities for management. Highly vulnerable habitats are likely to experience greater impacts whereas habitats with low vulnerability will be less affected or may even benefit from climate change. Distinctions between reservoirs can be made based on a variety of factors estimated to be impacted by climate change or based on characteristics of reservoirs that influence their resilience to climate change.
A reservoir vulnerability assessment is a tool that can provide managers answers to questions such as what types of reservoirs are most likely to be affected by climate change. A vulnerability index can reveal which reservoirs may be at risk of habitat decline, which are high priorities for management to sustain desired fish assemblages, and which may serve as climate change refugia. Understanding what in-lake or off-lake habitat traits make certain reservoirs more vulnerable than others provides a basis for developing appropriate adaptation measures for specific reservoirs and regions.
Nevertheless, vulnerability of reservoirs to climate change has not been established. Moreover, methodology for such assessments remains undefined. There are various databases that describe key physical characteristics of reservoirs nationwide, and there is wide availability of spatial and temporal climate models that permit reservoir-specific projections on likely changes in temperatures and precipitation at relevant scales. Reservoir characteristics (e.g., latitude, depth, catchment area, catchment land use) in combination with recent temperature and precipitation averages can be linked to existing reservoir habitat issues and vulnerabilities. These associations, in turn, can be used to predict reservoir habitat issues and vulnerabilities under expected future climate scenarios.
We propose using existing information to develop such models, quantitatively through correlative approaches and qualitatively through expert opinion, with the intent to provide nationwide rankings of the relative vulnerability of reservoirs, and to distinguish the specific factors that pose the greatest threats to reservoir fish habitats.
- construct a methodological framework for evaluating potential effects of climate change on reservoirs as fish habitats
- apply methodology to assess vulnerability of large U.S. reservoir
- explore patterns of vulnerability scores and vulnerability aspects
We propose this study as a series of one-year projects over a 4-year period to reach the overall objectives. Each year will have a measurable objective(s) and deliverables with the overall project objectives being reached at the end of year 4.
Objective 1– construct a methodological framework
There has not been a standard approach to assessment of fish habitat vulnerability to climate change. Because of this limitation, we intend to apply two methods to double our chances of successfully developing an effective index of reservoir habitat vulnerability. Method 1 will take a correlative approach in which quantitative correlative models are built to relate observed habitat characteristics to current climate. Once adequate models have been built, they will be applied to climate projections to infer potential climate-induced habitat changes. Method 2 will be based on expert opinion. This method will involve expert elicitation to estimate the general vulnerability of reservoirs relying on key habitat characteristics. This approach will fill the need for broad and relatively quick evaluation of the vulnerability of multiple reservoirs and habitats. There are many approaches to conducting expert elicitation. As part of this research we will investigate different approaches to select one or more most likely to provide the information needed.
Method 1.– We will investigate a correlative approach to developing vulnerability indices. Correlative models will be applied to relate observed habitat characteristics of reservoirs to current (1970-2000) climate. This model will attempt to link climate to habitat impact in reservoirs as they have existed over the last 2-3 decades. The resulting model will then be applied to climate projections to predict where and how climate in the future is likely to produce changes in reservoir habitats. In turn these can be used to assess the habitat quality for specific fish assemblages (e.g., warmwater, coolwater). Habitat distribution can be presence-only data, presence/absence, semi-quantitative habitat ratings based on Likert-scale surveys, or quantitative observations based either on empirical field measurements or existing GIS layers. Correlative models have been applied to species distribution predictions and have been used to explore the vulnerability of plants, invertebrates, birds, mammals, amphibians, and fishes. To our knowledge they have not been applied to score vulnerability of aquatic habitats.
Correlative models have the advantage of being spatially explicit and could be applicable to a wide range of reservoirs and fish habitats at various spatial scales. However, they have various uncertainties and potential errors originating from climatic, algorithmic, and habitat suitability ambiguities. Climatic uncertainties apply to all types of predictions as diverse models, parameters, and variables are used to predict future climate. Thus, depending on the climate prediction applied, estimated reservoir vulnerabilities may differ. Also, climate models often project future climate conditions at a coarser scale of resolution than that of reservoir habitat descriptors used to develop a model, and the estimated vulnerabilities may not be sufficiently fine scaled to properly score local reservoir habitats. Algorithmic uncertainties can arise from the differences in models used to predict upcoming habitat distributions and characteristics based on climatic variables, and from the selection of model variables (e.g., mean annual precipitation versus mean seasonal precipitation). This range of uncertainties can be addressed by applying multiple statistical methods and model structures, and then summarizing predictions across all models to generate a portfolio of forecasts, rather than one (i.e., ensemble modeling). Habitat suitability uncertainties may also arise if our assumptions made about habitat needs by fish are inappropriate. For example, the value assigned to a habitat is often linked to its ecological significance in aspects such as providing suitable water quality or food and shelter for the juveniles of a certain fish species. Our knowledge about habitat suitability is incomplete and extent of knowledge varies greatly among species. Lastly, correlative models require relatively large sample size and in many cases sample sizes may be small when considering a large geographical area such as the U.S.
We propose to develop a correlative model by using existing databases descriptive of reservoir fish habitat condition and characteristics across the U.S. and connecting these habitat databases to a database of predictions of future climate change. The habitat condition database consists of twelve habitat constructs developed from >50 habitat variables. The constructs reflect habitat ailments such as point source pollution, nonpoint source pollution, excessive nutrients, algae blooms, siltation, limited nutrients, mudflats and shallowness, limited connectivity to adjacent habitats, limited littoral structure, nuisance species, anomalous water regimes, and large water level fluctuations. The reservoir characteristics database consists of 62 variables that describe and quantify each reservoir (e.g., year constructed, average depth, catchment land use). The climate database consists of temperature and precipitation averages from recent history (1970–2000) and projections under several representative concentration pathway (RCP) emission scenarios in the 21st century.
We will combine data from each of these databases using a fuzzy join algorithm based on geocoordinates. This method uses spatial location of observations (rows) to join attributes (columns) from multiple sources. Once variables from the databases are joined, we will use the random forest algorithm to construct predictive models which relate climate and reservoir characteristics to habitat constructs. The habitat constructs are on a Likert scale from 0 to 5, with 0 representing little to no habitat impairment and 5 representing severe habitat impairment. We will use random forest models to predict the habitat impairment score of each habitat construct for reservoirs across the U.S. Random forests are an ensemble method which combine the results of many classification trees (i.e., CART) to make robust predictions. Classification trees partition data along predictor axes into homogenous classes (i.e., habitat construct scores of Krogman and Miranda 2016). Data is first partitioned using predictors that most clearly separate classes, forming a hierarchy or tree with the best predictors at the top. Random forests use subsets of data and predictors to optimize bias and precision of predictions. Because random forests use subsets of data, prediction accuracy and performance can be assessed integrally.
Method 2.– As previously suggested, correlative approaches when applied to the construction of a vulnerability index have climatic, algorithmic, and habitat suitability ambiguities that could influence the utility of the resulting index. To offset this hazard, we propose to develop an alternative tool that relies on an online questionnaire administered to local managers to generate vulnerability scores. With detailed instructions for scoring, the manager will be able to score reservoir attributes related to potential vulnerability or resilience associated with projected changes in climate and related phenomena. The attributes will be associated with the 12 constructs attributed to Krogman and Miranda (2016) and listed under Method 1. The manager will also be asked to score the relevance of each of the 12 attributes to each reservoir. Vulnerability to climate change could be scored as the cumulative value of a set of habitat attributes, weighted by their importance.
Response of reservoir habitats to climate change is influenced by characteristics associated with exposure and sensitivity. Exposure refers to the extent to which a reservoir will experience changing climate, whereas sensitivity is the extent to which a reservoir is affected. We can rely on basic ecological principles and published studies linking habitat changes (e.g., vertical stratification, algae blooms) to climate and related phenomena (e.g., average seasonal temperature, length of growing season, water residence time) to identify characteristics indicative of exposure and sensitivity. We then can integrate these characteristics to develop criteria predictive of climate change response into a scoring system that is regional and reservoir specific. Thus, scores can include both the predicted climate-related change (e.g., change in precipitation timing) and the predicted response of the reservoir (e.g., change in hydrograph). Before scoring the respondent will be provided with information on exposure extracted from climate projections for the region.
Repeatability is always a concern in surveys, especially those that involve expert opinion. To increase repeatability, each question in the questionnaire will assesses projected negative, positive, or no impact, but not the extent of impact. We expect that such an approach is likely to reduce error and increase repeatability.
It is difficult to make predictions about the potential effects of climate change when the possible effects may be unknown, or only partially predictable. The complexity of interactions in abiotic and biotic characteristics is also likely to lead to unpredictable changes. The correlative analyses described under Method 1 can provide variances applicable to produce confidence limits around vulnerability scores. With Method 2 we can generate equivalent measures of confidence by scoring uncertainty. To rate level of uncertainty respondents may be asked to (1) give a range of scores in the 0-5 scale, in addition to a single score, or (2) score uncertainty of response on a scale of 1 to 10. These estimates of uncertainty could help identify specific aspects where scoring was complicated, and to provide an indication of the certitude of the predictions. Uncertainty scores can also play an important role in identifying research needs or aspects that will require a more complex assessment of vulnerability.
Objective 2 — apply the method to assess habitat vulnerability of large U.S. reservoir
We will use models with satisfactory performance developed with methods 1 and 2 to estimate, under future climate predictions, scores for each of the 12 habitat impairment constructs and for each reservoir in our database. Vulnerability will be assessed by combining (e.g., by subtraction, as a ratio, or other) current reservoir habitat impairment scores and predicted change to scores (i.e., increase or decrease in impairment for each of the 12 constructs) under future climate. A composite vulnerability index can then be derived from the combination (e.g., summation) of individual habitat construct vulnerability scores. If appropriate, constructs may be assigned different weights depending on the strength of the correlation with climate variables or based on expert opinion.
The vulnerability index can be used to assess individual or multiple reservoirs. The index score for a single reservoir provides a quantitative measure of the vulnerability of a reservoir relative to other reservoirs or relative to the span of the index scale. More than one index score may also be calculated for the same reservoir across different climate projections, to assess a range of vulnerability relative to different climate models. We will also investigate if a vulnerability index may be estimated for distinct locations (e.g., embayments, main channel) of the same reservoir.
Objective 3– assess patterns of reservoir habitat vulnerability
Patterns of vulnerability likely exist over multiple reservoirs. Managerially relevant classes may be assembled relative to reservoir geography, altitude, purpose (e.g., flood control, hydropower, irrigation), size, fish assemblages (e.g., warmwater, coolwater, coldwater), basin, and others. Overall scores, for example, can be used to rank vulnerability for a class of reservoirs, or across geographical regions. Such group scoring can aid in prioritization, selecting targets for management actions, and broadscale planning for climate change.
Furthermore, the separable scores associated with each of the 12 constructs can highlight the relative contribution of the various pieces of the vulnerability index. By highlighting specific aspects of concern, construct scores suggest effective themes to target for management actions either by directing management towards an expected source of vulnerability or by identifying common areas of vulnerability for a reservoir class or geographic region.