Introduction

Climate change is one of the greatest challenges natural resources will face over the next century. While species are generally predicted to shift their ranges in response to warming conditions, individual ecosystems will have unique responses depending on local environmental variables and species’ sensitivities to environmental conditions [1–3]. As species disappear from or arrive into ecosystems as a function of shifting climate, the ecosystem services provided will likely be altered and managers must develop adaptive strategies to meet management goals (e.g., optimal take for stocks, MPA placement for conservation; [1, 4–6]). In order for resource management to be timely and responsive to climate changes, effective methodologies must be available to track changes in ecosystem components, such as community structure or function.

Classifying species by their affinities for environmental conditions can improve our ability to understand how local ecological communities are changing in response to acute and chronic climatic events and to contextualize any changes relative to specific environmental drivers. Because physiological data on most species’ thermal tolerances is absent, the most commonly used classifications are based on species thermal tolerances as inferred from biogeography in situ [4, 7, 8]. However, some classification schemes use a variety of data metrics including life history traits [9], other (non-thermal) environmental sensitivities and even co-occurring human activities [10]. Classification schemes create using readily available datasets, such as abundance or distribution monitoring data, can facilitate rapid information generation and quick management decision-making. Species classification schemes that are easily understood will best facilitate management communications with stakeholders and ultimately management action [11].

The most useful classification scheme for management uptake will not only be intuitive but also reliable. Reliability is dependent on the classification accuracy and sensitivity to detect changes that resource management would be concerned about. Classification schemes with high levels of erroneous classification could lead to overreaction or inaction. The risk of error and bias is especially high for schemes that use just a single data source to classify species or those that relay information in a single, composite indicator, such as the commonly employed Community Thermal Indicators (CTI; [7, 8, 12]). To alleviate spatial and temporal shortcomings of individual, often site-based or single time point data sources, researchers should utilize multiple data sources to build thermal classification schemes, including the expert opinion of local researchers, traditional resource users and naturalists.

In order to help regional resource managers understand climate impacts, we developed a robust thermal classification for a kelp forest fish community off the coast of California, USA. We created a composite classification using multiple sources of information and investigated agreement between the results of each individual source. Fishes were sorted into warm-affinity and cool-affinity (hereafter ‘warm’ and ‘cool’) species classifications using information from four different data sources (i.e., Expert Opinion, Range Limits, Museum Collections, and In Situ Density, see methods for details). Species were classified using each data source separately, and then a final aggregate classification was given to each species based on the majority agreement of classifications across the four data sources. An agreement score between the classification based on the four different sources was quantified to assess potential uncertainty in the classification scheme. The classification scheme was then used to evaluate changes in fish communities in the Santa Barbara channel before and after a marine heatwave event.

Materials and methods

Results

In the composite classification of all species (n = 152), 59 species (38.8%) were considered warm and 74 species (48.7%) were considered cool (S1 Table). Across each classification data source, the number of species falling in each group varied greatly: 44–59 for warm species, 32–83 for cool species and 1–29 for eurythermal species. Each of the data sources had a number of unclassified species; however, the ‘diver surveys’ was the data source with the highest number of unclassified species (n = 61).

We ranked the level of agreement of thermal affinity classifications (warm, cold, eurythermal) for each species across the four different data sources (i.e., Expert Opinion, Range Limits, Museum Collections, and Diver Surveys). While the majority of species had High classification agreement across all four data sources (n = 78, 51%), some species had Moderate (1 data source is in disagreement from the others, n = 27, 18%) or Low agreement (2 sources disagree from the others, n = 47, 31%).

All data sources captured the increase in warm species and the decrease in cool species after the onset of the marine heatwave except for range limits, which did not show a decrease in cool species (Fig 1). Warm species tended to show larger percent changes compared to cool species. Dive Surveys was also the data source with the smallest increase in density of warm water species (127.07% increase in density) but also showed the largest decrease in cool water species (47.34% decrease in density). Museum Collections showed the largest increase in warm water species (134.38% increase in density). The composite score was in the middle of the recorded changes for the four data sources both warm and cool species. In the ten species that had the largest differences pre and post heatwave (Fig 2), six were of high agreement, two were moderate agreement and two were low agreement (Fig 2). These strongest species responses were split pretty evenly between warm (n = 6) and cool species (n = 6). The summed changes in fish density responses to the heatwave were heavily driven by these top 10 species which was 6 times total observed change in density for the remaining species (Fig 2).

Fig 1

Changes in average annual fish densities.

Percent changes in average annual fish densities for warm (red bars) and cool (blue bars) species are shown for each classification data source for the years prior to (2012–2013) and after (2014–2015) the onset of the marine heatwave.

Fig 2

Variations in species response by classification metric.

Scatterplots of individual species responses to the marine heatwave are shown with pre-heatwave (2012–2013 summed density) on the x axis and post-heatwave (2014–2015 summed density) on the y axis. Points that fall below the black line are species that responded with density declines while points above the line experience density increases in response to the heatwave. Points are colored according to their classification (red for warm, green for eurythermal, blue for cool and purple for unclassified) for the A) composite classification B) expert opinion C) range limits D) museum Collections and E) diver surveys. In the composite classification, point size enlarge as agreement among classification data rises.

Discussion

Species classifications based on traits that are responsive to climate change (e.g., thermal and otherwise) may be prone to biases depending on the classification method employed and these biases can lead to inconsistencies among studies [8, 11, 17, 18]. Species responses to temperature shifts will vary both on temperature tolerance width, maximum survivable temperature, preferred temperature and the plasticity around these traits [4, 6]. In our system, thermal driven change to fish community composition appears to be heavily driven by the ten most responsive species. Fortunately, the composite scores for the majority of these top ten species were in high agreement and we can be confident in the utility of the thermal classification scheme as a tool for accurately tracking community-level response. As research and management aims to rapidly detect and respond to climate-driven ecological changes, users of classification schemes must have confidence that each response group is constructed such that all members respond similarly to a climate driver, but is inclusive of the range of intensity of responses species can display. Further work describing ecosystem response to the marine heatwave in this region can be found in Freedman et al. 2020 [19]. In this study, the number of kelp forest fish species classified as having warm and cool affinities differed by the data source used for classification. Differences in temporal and spatial resolution for each data source likely contributed to the number of species in each classification group and the number left unclassified due to data gaps.

The number of species left unclassified (e.g., present in the ecosystem but not captured in the classification data source) was highest (n = 61) for the subtidal diver survey data source. Recall that we used only two years (2009–2010) of the full SCUBA survey data for the classification in part to achieve complete spatial coverage of California, and also to avoid circularity in using the same data for classification as we are using to test the classification. Thus, 61 species were observed in the complete dataset that were not observed in 2009 or 2010. Researchers wanting to use in situ, field surveys in their composite classifications might have some issues because long running continuous surveys are uncommon across the world and are rarely conducted across large spatial domains [20]. The other two spatial datasets, museum collections and range limits, contained data from longer time periods and had fewer species left unclassified.

Museum collection records have spatial and temporal biases as well that may limit their effectiveness as a data source to determine thermal affinity when used in isolation. While Museum collections may encompass very long time periods, effort by collectors likely are infrequent, inconsistent and may not span the entire range of an individual species. This means collection events are likely affected by sampling bias and may lead to incorrect classifications alone. However, incorporating sampling events over a long time period allows classifications to incorporate historic variability in ecosystem structure.

Data sources can be biased in other ways that users of classification schemes should consider when implementing classification schemes on their data. Caution should be used when community indices are based on a single source of information as they are potentially error prone. For example, in the California Current Large Marine Ecosystem, El Niño events cause pulses of species recruitment northward of their typical ranges [21, 22] where they might establish but in very low densities; thus, biasing range limits as a data metric. Used alone, this could cause warm- affiliated species to be mis-classified as eurythermal or cool. In fact, in this study, range limits had the highest proportion of species classified as cool. Misclassification of some warm species may also explain why the range limits-based classification method was the only method that had cool water species increasing in abundance in response to the marine heat wave. Marine heatwaves, which have occurred locally [23, 24], also could drive warm species range expansion and alter classification findings [25]. Using older data prior to the prevalence of these phenomena will help establish community change from a relevant baseline in the Anthropocene.

Given that biases exist are likely to exist in any source of data typically used to classify thermal affinity, using a composite classification scheme is one option to mitigate bias. Using only species densities is often limited by sampling design and the potential for “double dipping”–that is, using the same dataset to classify species as to track them. Although we included density datasets in our classification and went on to use this classification to assess trends through time, using additional classification methodology in equal weights and a temporally limited sample allows us more confidence in testing the scheme on density trends. Including Expert Opinions into the composite score also allows for potential data biases to be crosschecked with the understanding of experts in the field. In this study, most species had high agreement (n = 78) between datasets, however a large number of species (n = 47) had low agreement between techniques. This means that these 47 species had a high likelihood of being classified differently depending on which dataset was used. High levels of species misclassifications would limit the scheme’s utility and could lead to incorrect conclusions about true community responses to climate change [20, 26, 27]. By using the majority classification across all datasets, a composite score limits the potential a species is misclassified by spreading misclassification risk across multiple sources of data.

In our study, the composite classification method resulted in density responses of warm and cool species that were less extreme compared to single data source techniques and as expected, did not have any of the maximum or minimum community responses to the marine heatwave. For most individual data sources, warm water species increased and cool water species decreased in response to the marine heatwave [25, 28]; however the magnitude of change was variable and using only Range Limits as a data source, we found the opposite pattern, that cool water species increased in abundance. False or inflated extreme responses to acute climate drivers could cause overcompensation with conservation measures [29–31]. By using the composite classification, there is a risk reduction of lost resource use opportunities, potential undesired ecological outcomes or potentially not intervening quickly enough to limit environmental impacts.

While we used a composite score based on agreement between four data sources, other means of averaging could be used as well as differential weighting schemes (we weighted all four data sources equally). For example, if one data source was deemed of higher ‘quality’ than another, it could be weighted more highly in the creation of a composite. Here we provide a simple example of a composite thermal indicator that uses data from multiple sources and that captured changes in fish communities over a thermal event in our region.

Conclusion

Classification schemes offer a quick, easy to interpret, and functional way to quantify community responses to climate change. Commonly used single trend measures, such as Community Thermal Index (CTI), can mask ecological complexity and make it difficult to make alternate management decisions for groups that respond differently to climate change [4, 8]. Because CTI is a single number metric, further analysis will always be warranted, to understand the actual structural changes driving CTI shifts. By splitting species into groups that likely will have differential responses to climate drivers, conservation practioners can create finer resolution conservation measures and attempt to mitigate direct pressures on each group independently (Fig 3). In contrast, traditional ecological community analysis metrics (e.g., PCA, CCA, regression trees) may be too complex for managers to make actionable conservation decisions as they can create too many groups that may confound responses to drivers of interest or not be amenable to policy or management actions [11, 17].

Fig 3

Comparisons between CTI and our multi-data source, thermal classification scheme.

The single trend line for Community Thermal Index (CTI) obscures the responses of specific groups of species to environmental drivers challenging management decision-making. The Y-axis on each of these graphs would represent some indicator of ecosystem change such as density or biomass. Because CTI is a single trend line, it masks the potentially different contributions of species or species groups to the trend. For example, the three scenarios on the right could all result from the single CTI trend shown on the left. The additional resolution offered by thermal classification gives managers finer scale information to enact specific actions that either benefit or boost populations at risk (e.g., decreased fisheries take, restoration action, marine protected) or increase resource use on populations experiencing increases (e.g., increased fisheries take, species culling).

There needs to be flexibility to increase, maintain and relax management strategies where appropriate and single number ecological metrics are not able to provide adequate information to inform complex management strategies [32–34]. Managers could mix restoration efforts [35, 36], species removal [37–39], fisheries management [34], and other management methods to mitigate changes in ecosystem structure. However, careful thought is required in the creation and use of classification schemes to ensure findings are consistent with known ecological phenomena [8, 17]. Combining sources of data to minimize temporal and spatial limitations is one way to address potential biases that could potentially misinform important management decision and lead to unintended outcomes. As managers struggle to keep up with rapidly increasing temperatures and more frequent marine heat waves, rapid assessment of local ecology with accurate classification strategies will be valuable to maintain critical ecosystem function and design effective conservation measures.