(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .
The great urban shift: Climate change is predicted to drive mass species turnover in cities [1]
['Alessandro Filazzola', 'Centre For Urban Environments', 'University Of Toronto Mississauga', 'Mississauga', 'Ontario', 'Apex Resource Management Solutions', 'Ottawa', 'Marc T. J. Johnson', 'Department Of Biology', 'Kimberly Barrett']
Date: 2024-04
Human experiences with nature are important for our culture, economy, and health. Anthropogenically-driven climate change is causing widespread shifts in biodiversity and resident urban wildlife are no exception. We modelled over 2,000 animal species to predict how climate change will impact terrestrial wildlife within 60 Canadian and American cities. We found evidence of an impending great urban shift where thousands of species will disappear across the selected cities, being replaced by new species, or not replaced at all. Effects were largely species-specific, with the most negatively impacted taxa being amphibians, canines, and loons. These predicted shifts were consistent across scenarios of greenhouse gas emissions, but our results show that the severity of change will be defined by our action or inaction to mitigate climate change. An impending massive shift in urban wildlife will impact the cultural experiences of human residents, the delivery of ecosystem services, and our relationship with nature.
Data Availability: The data that was used during this study was already publicly available. The occurrence records for species were obtained from Global Biodiversity Information Facility ( www.gbif.org ) and a list of the data citations used can be found in S2 Table . The climate data was acquired from ClimateNA ( www.climatena.ca/ ). All code will be made publicly available upon manuscript acceptance at
https://github.com/afilazzola/GreatUrbanShift . Code used for analyses and data visualization can be found at
https://afilazzola.github.io/GreatUrbanShift/ .
Here, we provide a synthesis of the extent that climate change is anticipated to have on biodiversity within cities. We hypothesized that climate change will drive a significant turnover in the composition of urban species in Canadian and American cities causing a great urban shift by the end of the century as species ranges track shifting temperature and precipitation patterns. We modelled the historic and future species distributions for 2,019 terrestrial animal species found in 60 cities in Canada and the United States. These 60 cities represent highly developed urban areas each with a population over 400,000 in the core municipal area ( S1 Table ). We selected species based on the frequency of verified observations per city (i.e., n > 10 individuals per city) by researchers and community scientists. Future climate models included an ensemble of six global circulation models (GCMs) and under three shared socio-economic pathways (SSPs) predicted until the end of the century (2081–2100). We compared the change in predicted occurrence of species based on climate suitability between historical and future climates to determine the species and cities that are expected to be most affected. Although it was not the original motivation for our study, our analyses allowed us to compare the differences in species native status (i.e., native vs. exotic) and IUCN Red List status (
https://www.iucnredlist.org/ ), since these species have important conservation implications.
Anthropogenically-driven climate change is threatening species globally [ 14 , 15 ], and cities are no exception. There has been repeated evidence that climate change will cause widespread shifts in a range of species and from all types of taxa [ 16 – 19 ]. While climate change is moving species across the continents (e.g., poleward and into higher elevations) [ 18 – 21 ], city boundaries are relatively fixed in space and are therefore likely to undergo climate driven changes in biodiversity patterns. For instance, common migratory songbirds in backyards have begun moving poleward in response to warming winter temperatures in North American cities [ 22 ]. Certain bioregions will also have greater vulnerability to climate change, including areas of North America where many major cities are located—such as temperate mixed forests and boreal coniferous forests [ 23 ]. Within the coming decades, we may observe significant species turnover (i.e., changes in the abundances and occurrence of species) in some areas as rapid climate change affects community assembly and species dispersal [ 10 , 24 ]. As a result, an individual who lives a lifetime within the same city will likely observe changes in the species that occur around them. Some research has already projected significant changes in the composition of urban plants and bird species for European cities in the next 60 years [ 25 , 26 ]. However, an examination of the potential shifts in community composition from climate change for all animal taxa in cities has not been comprehensively conducted in North America.
Nature is an integral element of cities globally. Over half the world’s population live in cities and the wildlife that people observe within their respective urban realm represents the species with which they have the most direct familiarity [ 1 , 2 ]. We value these urban species because they provide a benefit in terms of delivering ecosystem services, such as supporting mental well-being, providing pollination or pest removal, and recreation [ 3 – 6 ]. Iconic species can also be emblematic of the community within cities [ 7 ], such as the animal species used as mascots for sports teams or represented on governmental flags. However, anthropogenic impacts such as climate change can threaten the presence of species in cities [ 8 ], making iconic and familiar species at risk of extirpation from the communities they represent. Just like the California grizzly bear is extinct from where it is displayed prominently on the state flag, with climate change, the floodgates are open and many other emblematic species are at risk of extirpation from the communities they represent [ 9 , 10 ]. In other instances, gradual changes in species composition can go unnoticed between generations of human residents because of changing expectations of what constitutes the natural environment, i.e., the shifting-baseline syndrome [ 11 – 13 ]. Thus, future generations of urban dwellers may be unaware that the wildlife they experience in their home cities is different than what exists today. Alternatively, the shift of urban species may be so substantial and within a single generation that it will be clearly noticeable among residents.
Methods
City and species selection We chose the 60 most populated cities in Canada and the United States, which all have populations over 400,000 people (S1 Table). In each of these 60 cities, we created a 20 x 20 km quadrat around the centroid of the municipal boundary. For consistency, we picked this quadrat size for all cities regardless of the municipal boundaries to capture the core urban areas of selected cities. The size of this quadrat also minimized placement outside of the city boundaries or in large waterbodies. Using the Global Biodiversity Information Facility (GBIF;
https://www.gbif.org/), we downloaded all species records for terrestrial animals found within that quadrat. All records of species occurrences used and their associated databases can be found at S2 Table. The term “terrestrial” here is meant to represent animals that do not spend their entire life cycle in water (e.g., fish, cetaceans) and thus would include semi-aquatic organisms (e.g., amphibians, dragonflies) and flying organisms (e.g., bats, birds). Species records were filtered to include all animal species that have at least ten records within the last ten years for any of the 60 cities, indicating the species has been observed enough times that it was not incidental. Many target taxa were observed in multiple cities, such as hawks (Accipiter spp., Accipitridae), dabbling ducks (Anas spp., Anatidae), and bumble bees (Bombus spp., Apidae) but some species were found unique to only one city, such as the bark anole lizard (Anolis distichus) in Miami or Strand’s carpenter bee (Xylocopa strandi) in Houston. There was a bias in the species list towards taxa that are larger and more identifiable, as is typically found in community science, but also in traditional science [27]. In total, we found 2,259 unique species that matched our criteria. For each of these species, we used GBIF to download all occurrences between 2000 and 2020 for all North America. We selected this area, larger than Canada and the USA where our selected cities are present, to capture the total climatic niche and range of conditions that each selected species can occupy. In total, we downloaded over 18.4 million occurrence records from GBIF with a median of 1,059 records per species (minimum 10 records, maximum 138,746 records). Although there were large differences in records per species, our modelling approach was robust to infrequently surveyed species [28, 29] such that similar confidence could be treated among model results. There have been reported issues with the reliability of GBIF data concerning the accuracy of records in time, space, and species identification [30, 31]. While no one approach can be applied to solve all issues associated with GBIF records [30], steps can be taken to minimize the impact and increase confidence [32]. We recognize that the size of our dataset makes verification of every individual record impractical, and thus despite our efforts, some amount of inaccuracy will remain. For all records, we restricted occurrence to North America, which removes common errors associated with coordinates labelled as zero or mistakenly entered records (e.g., latitude and longitude swapped). Our analysis was not reliant on time, therefore temporal issues, such as mismatches in months or days, would not be impactful on our results. We removed all records in the oceans and removed duplicates. Removing duplicates will also mitigate issues such as when records are reported as the centroid or capital of a country since, if inaccurate, would only represent one out of potentially thousands of records. Similarly, inaccuracies in species identification may remain within the dataset, but we expect that the occurrence of relatively few incorrect methods would have a small impact on our large dataset distributed across Canada and the US.
Climate variables We used a series of future climate models to capture the range of potential outcomes for the end of the century (2081–2100) under different greenhouse gas emission scenarios. All data climate models, data management, and statistical analyses were conducted in R Version 4.1.0 [33]. We downloaded 24 bioclimatically relevant variables from ClimateNA [34, 35] that represent down-scaled climate variables in 4.6 km grid cells. In addition to the current climate conditions (1990–2020), we also downloaded an eight-model ensemble of future climate condition [34]. These models were all selected under the Coupled Model Intercomparison Project Phase 6 (CMIP6) and include the global circulation models (GCM) that are more representative of the North American climate [34]. Using an ensemble model provides a more conservative estimate of climate change effects on species distributions because it reduces model-specific anomalies [36]. We downloaded the future climate conditions for 2081–2100 under three shared socioeconomic pathways (SSP 1–26, SSP 3–70, SSP 5–85). We selected the three SSP scenarios to represent a range of outcomes based on action to reduce greenhouse gas emissions including sustainable development (SSP 1–26), barriers to mitigating climate emissions and a lack of regional cooperation (SSP 3–70), and continued development of fossil fuels and land (SSP 5–85) [37]. These SSPs represent the latest framework for future climate projections that considers uncertainty in both the climate outcomes from greenhouse gas emissions (i.e., Representative Concentration Pathways; RCPs) [38] and socioeconomic development in the absence of policies to mitigate climate change [37]. In North America, SSP 1–26 and SSP 5–85 both project increased urbanization although for different reasons with the former under high density development and the latter under increased urban sprawl [39]. The SSP 3–70 projects a relatively little land cover change to urban [39].
Species distribution modelling We conducted species distribution modelling for each species to determine the historic climatic niche and use these models to predict their future range. For each species, we conducted corrections for survey bias, minimized spatial autocorrelation, and automated model tuning to quantify the relationship with climate. We used Maximum Entropy (MaxEnt) [40] because our data represents presence-only data and thus requires the generation of pseudo-absences [41]. MaxEnt is a machine learning algorithm that predicts the suitable conditions for a species by modelling the relationship of occurrence records to a set of environmental variables [40]. The GBIF occurrence records are collated from a series of community science sources (e.g., iNaturalist, eBird) and museum specimens. These records typically have unequal sampling efforts favouring areas with greater accessibility such as along roads and in parks, as well as under sampling in difficult-to-access areas such as mountains [42, 43] and private property. To account for unequal sampling, we conducted two methods for bias correction: spatial thinning and restricting background points. Spatial thinning is one of the most effective methods for accounting for sampling bias in MaxEnt [44] and involves removing multiple observations within a certain distance to approximate a systematic sampling of the target species. We spatially thinned our dataset by overlaying a 25 x 25 km raster (i.e., 5 factor larger) and by removing multiple occurrences within the same cell. We also restricted the background records (i.e., pseudo-absences) which has been observed to improve MaxEnt performance when the occurrences occupy an area smaller than the total study area [45]. Using the randomly generated background points, spatially filtered occurrence records, and climate variables without collinearity, we conducted MaxEnt modelling for each species. Since MaxEnt is a presence-only analysis, background points (i.e., pseudo absences) need to be generate in a manner that accurately captures climate conditions with the geographic study area. These background points serve to quantify the available climate conditions to be used as a comparative distribution against the climate conditions specific to the presence records. Spatial autocorrelation, the lack of independence between occurrence records, is a frequent problem when working with spatial environmental datasets [46] including species distribution models [47–49]. Without compensating for spatial autocorrelation, species distribution models tend to overestimate the accuracy of the model and suggest the results that are more reliable than is true [49]. For details on our methods in calculating background points, conducting spatial filtering, and removing collinear variables, see S1 File. We used an automated tuning and evaluation process for MaxEnt function (ENMevaluate, package ENMeval) [50]. MaxEnt was automated to assess best model using eight feature classes (L, Q, P, LQ, HQ, QPH, QPHT, and LQHP) and six regularization parameters (0.5, 1.0, 1.5, 2.0, 2.5, 3.0). The acronyms in the feature classes relate to relationship between the predictor variables and the predicted occurrence of the target species including linear (L), quadratic (Q), product (P), hinge (H), and threshold (T) [40, 50]. The regularization parameters control for overfitting by downweighing co-efficients, but must be balanced against preventing model tuning. Tuning was accomplished by using spatial block cross-validation, which splits the target area into a number of grids and then resamples data within each respective grid for training and testing to improve model metrics [50, 51]. Model statistics were then averaged across all spatial subsets. Each species was run with a different combination of feature classes and regularization parameters (48 different models per species) and the best model was selected using the highest average Boyce Continuous index (BCI) value [52, 53]. BCI is ideal for presence-only models because it measures model accuracy based on how the occurrence records differ from a random distribution, with values +1 being accurate, values of 0 suggesting the model is completely random, and values -1 indicating high predictions away from occurrence records. Models were conducted in parallel for efficiency in runtime using GNU parallel [54] on the Compute Canada super computer cluster (www.computecanada.ca/). From the best model determined for each species, we extracted the average training area under the curve (AUC), average BCI, percent contribution of each environmental variable, the optimal feature classes and regularization parameters, and the average difference between training and testing AUC values. We also determined the threshold to cut-off model predictions based on the lowest trade-off between sensitivity and specificity (function threshold, package dismo). For a visual workflow of the analyses conducted for species distribution modelling, see S1 Fig. We removed species from further analysis that failed to provide satisfactory model results. For example, a species was not included in the final analyses if there were insufficient records from GBIF to confidently model the distribution (n < 10), if the model failed to produce a best model, or the AUC value was less than 0.70 (240 species removed). All remaining analyses included 2,019 species that met these criteria. For a list of all meta-data associated with modelling for each species including AUC/CBI scores, parameters, and MaxEnt settings, see [55].
Predicted occurrence based on climate suitability The output predictions from MaxEnt were fitted to a logistic distribution and represent the predicted occurrence based on climate suitability for the target species to inhabit, and range between 0 (completely unsuitable, low species prevalence) and 1 (ideal climate, high species prevalence). These values can function as a probability that a species may be observed in a city (i.e., 0 = never, 0.5 = occasionally, 1 = often) when considering climate alone. However, we note that this value does not translate to a true probability of occurrence because many non-climate factors could restrict or increase the potential of the species observed (e.g., dispersal, species interactions, resource availability). Additionally, there is some discussion that the logistic output from MaxEnt represents an estimate of the probability of presence, rather than true probability, as the output values are based on user inputs [see 56]. While these considerations of estimating occurrence are especially relevant for determining a species-specific distribution (especially between studies), our study is exclusively examining the relative difference between historic and future estimates of probability within the same species using the same model to predict for both time frames. We estimated the predicted occurrence of each species for every city under each climate scenario. Within the 20 km quadrat in each city, we created a stratified grid of 100 points that we extracted the historic climate and future climate in each SSP and both timeframes. Using the best MaxEnt model, the predicted occurrence for each of the species was estimated using the extracted climates of the 100 points in each city. If the average predicted occurrence was above the identified threshold from the MaxEnt modelling, we considered that species to occur within the city. Our research question was interested in the relative change in predicted occurrence between future and historical timeframes. Therefore, for all analyses we calculated 1) the number of new, extirpated, and unchanged cities for each species, and 2) the number of gained, lost, and unaffected species for each city (S3 Table).
[END]
---
[1] Url:
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0299217
Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.
via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/
