Introduction
The effects of climate change in the Arctic are particularly prominent, as temperatures have risen nearly three to four times as fast as in the rest of the planet (Field & Barros, 2014; Rantanen et al., 2022; Zhu et al., 2016). Shifts in climatic patterns enable the expansion of temperature-limited shrub vegetation at a global scale to higher latitudes and elevations (Asmus et al., 2018; Myers-Smith et al., 2020; Sweet et al., 2015). This phenomenon is not only particular to shrubs but also to treelines that occur at the edge of the forest-tundra habitat, where temperature-limited trees are barely capable of growing. Treelines shadow climate patterns by expanding or shrinking their distribution across large spatial scales to ensure optimal abiotic and biotic conditions to develop (Dial et al., 2022; MacDonald et al., 1993; Reich et al., 2022). Warmer temperatures release the nutrients trapped in cold soils, allowing optimal establishment and development of trees (Hobbie & Chapin, 1998; Sullivan et al., 2015). Treelines are expanding in range across the planet due to warmer ambient temperatures (Holtmeier & Broll, 2005; Kruse et al., 2019; Pearson et al., 2013; Rees et al., 2020; Rupp et al., 2001). The response of treelines to climate change is tree species-specific, since the set of traits particular to each tree species is what determines its ability to establish, develop, reproduce, and disperse to novel sites. Studies in Fennoscandia however have forecasted the distribution of broadleaf trees and overseen the distribution of conifer trees (Rees et al., 2020).
The treelines in Fennoscandia consist mainly of birch (Betula pubescens) and in less proportion of pine (Pinus sylvestris) or spruce (Picea abies) and occasionally mix-species treelines are found (Kullman, 2002, 2007). The evergreen pine trees, the target species of this study, are generally known to have a unique set of traits that allows them to adapt to unique conditions such as saving energy during winter by not shedding their needles (Ottander et al., 1995), high plasticity in root architecture to secure water and nutrients in situations of drought and depleted soils (Moser et al., 2015) and possessing recalcitrant needles that are barely palatable for herbivores (Maes et al., 2019; Ramirez, Jansen, den Ouden, Moktan, et al., 2021). Pine trees on the other hand, cannot grow in shaded areas because they are light-demanding (Niinemets, 2010) and they have low tolerance to tissue damage by herbivores due to their strong apex shoot dominance (Aarssen, 1995). These set of unique traits allow pine trees to grow at their minimum temperature range while forming treelines in the Fennoscandian Arctic (Kullman, 2007). The extent to which pine trees will shift their spatial distribution as a response to future environmental conditions due to climate change remains largely unknown.
The purpose of this study is to predict the distribution of pine trees in the Fennoscandian Arctic by drawing from three datasets that implemented distinctive methods during data collection. This is done by employing three independent species distribution models (SDM), one for each of the datasets. The first dataset includes observations of the presence and absence of pine made only by me (a researcher) across an elevation gradient in four distinct locations in Sweden and Norway. The second dataset belongs to the Swedish National Forest Inventory and includes a wide network of permanent plots across Sweden to estimate forest metrics and track these metrics over time. The third dataset belongs to the Artdatabanken which is an online platform that gathers observations reported by citizens, nature officers, and researchers from across all regions in Sweden. Understanding the implications of how different data collection methods have on the predictions yielded by SDM provides researchers with the necessary input to choose the best experimental design for their study.
To address this research objective, I raised two questions: (i) How will the distribution of pine trees respond to climate change in the next 50 years? (ii) Which method used to collect data is better at predicting the distribution of pine trees as a response to climate change? (I) I predict that the pine distribution will only expand to higher elevations and latitudes as climate change releases the temperature limitation in pine performance. (ii) Citizen science data will be best suited to predict the distribution of pine trees because this method yields a greater number of observations that are distributed across an entire elevation gradient (Richardson & Whittaker, 2010). Although my dataset compiles pine observations along an elevation gradient, it falls short in the number of observations due to the limitation in manpower. The NFI compiles a great number of pine observations, but these do not cover the entire elevation gradient.
Methods
Study area
The study area covered the Fennoscandian peninsula, which includes Sweden, Norway, and Finland. The visual representation of the pine SDM includes other countries that need to be ignored since no data was collected in these territories.
Vegetation plot network
I walked along an elevation gradient at four sites in Sweden to establish a network of 5x5-meter vegetation plots to record the presence and absence of pine trees. Abisko, Jäkkvik, Nordli, and Ånn were chosen because they are among the few locations in Fennoscandia that harbor a treeline composed of birch and pine trees. Between the summer of 2020 and 2021, I carried out data collection by annotating the presence and absence of pine, and for each observation, I annotated the geographical coordinates. A minimum distance of 150 meters was used between observations to ensure an accurate representation of the true distribution of pine. This fieldwork campaign yielded a total of 195 data points unevenly distributed between the four locations: Abisko (n=77), Jäkkvik (n=37), Nordli (n=43), Ånn (n=38). This dataset is now referred to as PLOT.
Swedish National Forest Inventory
Sweden for the last 100 years has been systematically collecting and storing data to create an updated inventory of its forests. An extensive network of permanent plots with a size of 154 m
2 is used for the forest inventory (Fridman et al., 2014). For this project, I drew a subset of the official data by downloading the readily available file from the webpage of the Swedish University of Agricultural Sciences (
https://www.slu.se/nfi) and filtering in data from the 2022 field survey (Swedish NFI, 2023). The presence and absence of pine trees were extracted from these data along with the geographical coordinates of the plot. This data extraction yielded a total of 2,356 data points. This dataset is referred to as NFI.
SLU Artdatabanken
The Swedish University of Agricultural Sciences administers the species the Swedish Species Observation System (Artportalen) to register and map the species diversity of Sweden. This is an online portal where citizens, conservationists, and researchers can report the species they have observed (
https://artportalen.se/). I used a subset of this databank by filtering in data on pine and only for observations made between June 2
nd of 2020 and June 2
nd of 2023. The presence of pine trees was extracted from this dataset along with the geographical coordinates of the observations. This process yielded a total of 9359 data points. Since these data do not include the absence of pine trees, I merged the absence data from the PLOT network into this dataset (n=71). From now on, this dataset is referred to as ART.
Environmental data
Current and forecast global environmental data in the form of minimum temperature (K) and annual precipitation (kg m
-2 s
-1) was sourced from Coupled Model Intercomparison Project phase 5 (i.e., the CMIP5 multi-model ensemble data,
https://esgf-node.llnl.gov/projects/esgf-llnl/). This project developed a scientific framework that yielded likely estimates of environmental data linked to GPS coordinates (Meehl et al., 2009; Taylor et al., 2012). The data used for this project belong to the 50-year forecast data library and these data are extensively used for the Intergovernmental Panel on Climate Change Assessment Reports (Appendix 1, Fig. S1.1). CMIP5 was favored over CMIP6 because its estimates are in line the observations, while CMIP6 robustness remains to be tested due to its recent release (Carvalho et al., 2022).
Data analysis
Generalized Linear Models (GLM) were used for understanding how environmental variables can be used to predict the current distribution of pine trees in western Fennoscandia. The spatial extent of the model includes Norway, Sweden and north Finland, while the temporal extend is the year 2023 for the current predictions and 2073 for the forecast. Three independent GLMs with a binomial and logit link function were specifically used to test the current pine distribution (presence vs. absence) set as a response to temperature (min) and precipitation as fixed factors sourced from the CMIP5 data. Each of the GLMs employed three distinct datasets (PLOT, NFI, ART). An interaction between temperature and precipitation was also tested, but it was omitted from this study because it did not improve the fit of the model. The fit of the model was evaluated by comparing the AIC values. The model diagnosis was done by assessing the spread of the residual and by plotting a Receiver Operator Characteristic Curve (ROC), both tools widely used in SDM (Miller, 2010; Zurell et al., 2020). The area under the ROC curve (AUC) was also extracted, which indicates the efficiency of the model to predict. The value ranges between 0 and 100, with 100 being a highly accurate model. A p-value < 0.05 was used as the threshold for a significant relationship between response and predictors. The forecast of the pine distribution was made by extrapolating to a map of Fennoscandia the three GLM responses linked to the future environmental data sourced from CMIP5. The AUC and the coefficient of determination was used to rank the models with distinct data collection methods. All statistical analysis was conducted in “R 4.0.2” (R Core Team, 2013). Maps and figures were done in “R” with the packages “ggplot2” version 3.3.5 and “ggmap” version 3.0.0.903 (Kahle & Wickham, 2013; Wickham & Winston, 2016).
Results
The probability of current pine presence responded positively with temperature (b=0.046, p=0.005, Table 1) and negatively with precipitation (b=-0.051, p<0.001) when employing the dataset belonging to the PLOT network. The probability of current pine presence responded negatively to temperature (b=-0.007, p<0.001) and precipitation (b=-0.020, p=0.009) when employing the NFI data. Finally, the probability of current pine presence responded positively to temperature (b=0.199, p<0.001) and negatively to precipitation (b=-0.192, p<0.001). See Table 1 for AUC and R2 values and Appendix 1, Fig. S1.2 for ROC curves. When spatially forecasting with the PLOT and ART datasets, the pine distribution will likely expand in range to higher elevations and remain the same in lowland Fennoscandia (Fig. 1).
Table 1. Generalized Linear Model for the likelihood of having pine trees as a response to current temperature (min) and precipitation. The results for three identical models, one for each dataset type, are presented. The models are accompanied by the coefficient of determination (R2), intercept, coefficients, and p-values.
Data | Response | AUC | R2 | Intercept | p value | Temp. | p value | Precip. | p-value |
Plot | Pine | 0.657 | 0.09 | 3.070 | <0.001 | 0.046 | 0.005 | -0.051 | <0.001 |
NFI | Pine | 0.561 | 0.02 | 1.306 | <0.001 | -0.007 | <0.001 | -0.020 | 0.009 |
ART | Pine | 0.992 | 0.63 | 10.854 | <0.001 | 0.199 | <0.001 | -0.192 | <0.001 |