2.1 The CLIMEX model

In this study, we used CLIMEX (version 4.0; Hearne Software) to evaluate the climate-based potential distribution of the WCSB in Europe. The use of CLIMEX requires the estimation of parameters based on the biological characteristics of a target species and climate data pertaining to the region for which species distribution is to be predicted (Kriticos et al., 2015). There are two primary indices in the CLIMEX model; Growth Index (GI) and Stress Index (SI) for growth and inhibition potential under the given climatic condition, respectively. The GI is calculated by multiplying various sub-indices, such as Temperature Index (TI), and Moisture Index (MI), while SI is comprised of Cold Stress (CS), Heat Stress (HS), Dry Stress (DS), and Wet Stress (WS). The TI requires four parameters, which are lower temperature threshold (DV0), lower optimal temperature (DV1), upper optimal temperature (DV2), and upper temperature threshold (DV3). Similarly, MI involves four parameters, including lower soil moisture threshold (SM0), lower optimal soil moisture (SM1), upper optimal soil moisture (SM2), and upper soil moisture threshold (SM3).

In this study, we estimated parameter values from DV0 to DV3 and population degree-days (PDD) (minimum degree-days above DV0 necessary to complete one generation) based on lower developmental temperatures and accumulative temperature (degree-days) of WCSB reported in a previous study (Barta, 2016). By contrast, other parameters were set to have the best fit to the actual distribution of WCSB in North America because there was no available data for estimating those parameters. Given that most of the distribution coordinates for WCSB are located in North America and Europe, the parameters were calibrated for North America, and the set parameters were applied to Europe. However, using PDD, we were unable to project a potential distribution that adequately covered the actual distribution of WCSB. Consequently, we performed CLIMEX modeling both with and without consideration of PDD. The approach for setting up the parameter values has been described elsewhere (Jung et al., 2016; Kriticos et al., 2015), and all parameter values used are listed in Table 1.

On the basis of the estimated parameter values and regional climate data (described below), CLIMEX calculates the Ecoclimatic Index (EI), which is a quantitative measure of the climatic suitability of a specific area for a given target species by multiplying GI and SI. EI values are scaled from 0 to 100, with values close to 0 indicating that the target species may not survive in an area, and higher values being considered indicative of favorable climatic conditions for species establishment (Jung et al., 2016). In general, climatic suitability is categorized as follows: EI = 0 (unsuitable); 0 < EI < 10 (marginal); 10 < EI < 30 (suitable); and EI > 30 (optimal) (Byeon et al., 2018; Jung et al., 2016; Kriticos et al., 2015).

2.2 Distribution of the western conifer seed bug in Europe

First observed in northern Italy in 1999, the WCSB is now distributed in approximately 30 European countries (Barta, 2016; CABI, 2019). Recently, large numbers of WCSB occurrences have been reported in the United Kingdom (UK), Belgium, and the Netherlands (GBIF.org, 2020), and there has been a continuous expansion in the species' northern limit of distribution, with reports of these bugs in several Northern European countries, including Sweden and Norway (Mjøs et al., 2010). The occurrence records of WCSB were obtained from GBIF and crosschecked with other sources including the Centre for Agriculture and Bioscience International (CABI), and previous studies to minimize the uncertainty and bias caused by human accessibility (Barta, 2016; CABI, 2019; Mjøs et al., 2010). A total of 15,241 coordinates recorded from 1985 to 2020 in Europe were collected, while 1982 coordinates were additionally obtained to fit the parameters of CLIMEX model in North America (GBIF.org, 2020; Figure 1). Given that uneven sampling needs to be adjusted to improve the reliability of model results and a sufficient number of occurrence coordinates were confirmed, the obtained data were spatially filtered with 10 km radius buffer to additionally reduce the sampling bias as well as to avoid spatial autocorrelation (Beck et al., 2014; Boria et al., 2014; Brown et al., 2017; Kramer-Schadt et al., 2013). As a consequence, we obtained a total of 1939 data points and used it as a presence data (GBIF data). With respect to pseudo-absence data, we used ArcGIS (version 10.4.1; ESRI) to extract 1939 points from grid cell at which WCSB was not recorded using under-sampling in all Europe. We set pseudo-absence points equal to the number of presence points to avoid biased learning caused by one set of data being excessively larger than the other. Subsequently, both the presence and pseudo-absence data were integrated into the total occurrence data in Europe to validate the CLIMEX model and to operate the RF algorithm.

2.3 Constructing the RF algorithm

The RF algorithm is an ensemble algorithm used to perform regression or classification analysis based on the construction of a multitude of decision trees (Cutler et al., 2007). Compared with other machine learning algorithms, the use of the RF algorithm facilitates a more accurate classification of species distribution (Park et al., 2016). When using the RF algorithm, classification or regression is trained based on a subset of the total data, and an out-of-bag (OOB) element is calculated to estimate the accuracy and error rate (Cutler et al., 2007; Garzon et al., 2006; Vincenzi et al., 2011). The number of trees (ntree) was varied from 50 to 400 with increments of 50, while variables at each split (mtry) were set to 1, 2, 3, or 4. Then, we evaluated all 32 combinations of ntree and mtry, and the optimal structure was determined by the values minimizing the OBB error rate with no signs of over-fitting. For the purposes of the present study, we randomly divided the distribution data coding for the presence and pseudo-absence of the WCSB in Europe into training (2716) and test (1162) data at a ratio of 7:3. In addition, both training and test data contain the same ratio of the presence and pseudo-absence of the WCSB at 5:5. The values of each of the eight variables located in the presence or pseudo-absence points were extracted using ArcGIS. The ntree parameter was set to 100 trees, as this was found to minimize the OOB error rate, and mtry was set to 3, corresponding to the square root of the number of variables. mtry generally uses one-third or the square root of the number of variables for classification (Cutler et al., 2007). Thereafter, we visualized the results of classification (presence and pseudo-absence coded as 1 and 0, respectively) on a map of Europe.

To assess the performance of the model based on sensitivity and specificity, we used a confusion matrix recording true positives (TPs), false positives (FPs), false negatives (FNs), and true negatives (TNs). The TP (TP/[TP + FN]) and FP (FP/[FP + TN]) rates were calculated for sensitivity and specificity, respectively (Garzon et al., 2006; Table 2). Subsequently, a receiver-operating characteristic (ROC) curve was plotted to depict the relationship between sensitivity and specificity (sensitivity vs. 1 − specificity) on a two-dimensional plane. True Skill Statistic (TSS) was calculated by sensitivity + specificity − 1, and the area under the ROC curve (AUC) was measured (Allouche et al., 2006). In addition, we also examined the relative importance of the variables used in model, according to the degree of improvement in model accuracy attributable to these variables and their contribution to the classification impurity by means of mean decrease accuracy (MDA) and mean decrease in Gini (MDG), respectively (Cutler et al., 2007). The entire model construction process is illustrated in Figure 2.

2.4 Environmental factors required for model operation

The survival rate, growth rate, and reproduction of WCSB depend on climate factors such as temperature and precipitation (Barta, 2016; Lesieur et al., 2019). In addition, the distribution of WCSBs differs depending on geographical barriers and land use; thus, climate and environmental factors were selected to develop the models (Barta, 2016; Bernardinelli & Zandigiacomo, 2001; Fent & Kment, 2011; Lesieur et al., 2019; Tamburini et al., 2012).

To operate the CLIMEX model, long-term average climate data, such as the average maximum and minimum temperatures, precipitation, and relative humidity, are essential. Therefore, we obtained climate data for the period between 1960 and 1990 with a resolution of 10 min (corresponding to approximately 18 km × 18 km grids) from CliMond (https://www.climond.org) for fitting the potential distribution using records obtained for the WCSB in North America (Figure 3), whereas with respect to European climate data, we used the E-OBS dataset obtained from the EU-FP6 project UERRA (https://www.uerra.eu), which included the average temperature, minimum temperature, maximum temperature, precipitation, relative humidity and global radiation for the years from 2001 to 2020 (Cornes et al., 2018; Kriticos et al., 2012). CLIMEX model used monthly average temperature, minimum temperature, maximum temperature, relative humidity, and precipitation in Europe except global radiation obtained from the EU-FP6. Finally, the potential distribution of WCSB in Europe was projected with 18 × 18 km resolution using the five variables in CLIMEX.

To run the RF algorithm, we obtained variables of land cover, the DEM, and monthly climate records (minimum temperature, maximum temperature, average temperature, precipitation, relative humidity, and global radiation) obtained from the EU-FP6 project UERRA (https://www.uerra.eu). We performed correlation analysis to remove variables with Pearson correlation coefficients greater than .8. Finally, we selected the following variables: monthly maximum temperature in March and May; minimum temperature in August and December; total precipitation in April, May, August, September, and October; radiation in May and December; and relative humidity in April and December. Because the development period of WCSB was considered to be between mid-March and mid-July, the maximum temperatures of March and May were chosen (Barta, 2016). By contrast, the minimum temperature was determined by selecting August and December, which mark the beginning of the second generation and entry into winter for WCSB (Barta, 2016). To ensure the seasonality, variables of precipitation were chosen for each season. For radiation and relative humidity, April or May and December were chosen to represent the start of growth and diapause in spring and winter, respectively.

Moreover, the distance from coniferous forest and a digital elevation model (DEM) was used in conjunction with the climatic variables. Land cover data, comprising 44 land types, were obtained in the form of polygons compatible for use with ArcGIS (CLMS, 2019), and were extracted for the conifer land type (mixed and coniferous forest). In addition, we calculated the distance of conifers from each cell and used this as a variable for RF analysis. Data used for the DEM were obtained from the European Environment Agency and coded with altitudes ranging from −226 to 5111 m above sea level (a.s.l.) in the raster format (30 × 30 m cells) (EEA, 2019). Since all variables have different resolutions, the layers of each variable were overlapped through ArcGIS, and the values of variables were extracted using a lattice point. The potential distribution of WCSB in Europe was projected with 1 × 1 km resolution in RF.

3.1 Potential geographical distribution of the western conifer seed bug in response to climate

We found that the potential distribution of the WCSB predicted using climate data based on CLIMEX modeling did not fit the actual range of WCSB distribution (Figure 4a). In particular, according to the model, the effective accumulated temperature required for the completion of a single WCSB generation was insufficient in the UK, Belgium, and the Netherlands. Even though we used the lowest value reported by Barta (2016) as the lower developmental threshold (LDT) and effective accumulated temperature, we were unable to simulate the potential distribution of WCSB, such that it fitted the actual distribution. Agreement between the projection and the actual distribution could, nevertheless, be achieved by reducing the 13.4°C LDT reported by Barta (2016) by 3°C. However, the potential distribution without the inclusion of PDD could project more than 99% of the occurrence records (Figure 4b). In addition, the CLIMEX model typically predicted occurrences at the border between the United States and Canada. The potential distribution in Europe, particularly that in Northern Europe, was found to be influenced to a large extent by using PDD. Consequently, the occurrence pattern of the WCSB would differ depending on the accumulation of a requisite number of degree-days necessary for the completion of a single generation.

Compared with other European regions, the climate in Southern Europe was found to be more suitable for the establishment of the WCSB. In contrast to Southern Europe, Northern Europe was deemed to be climatically unsuitable for WCSB, owing to high cold stress leading to high winter mortality. Moreover, the climatic suitability of Southern Europe for WCSB was shown to be 2.4-fold higher than that of Western Europe, where the maximum occurrence of the WCSB was recorded. Whereas the climate in most areas of Italy, France, Germany, Poland, Hungary, and the Czech Republic is considered favorable for WCSB expansion (EI > 20), that in some parts of Switzerland and Austria, notably the high-altitude alpine regions, was found to be unfavorable owing to the low temperatures (Figure 4a). Furthermore, a latitude of approximately 54° N was predicted to be the northern limit of WCSB distribution, which implies that if the effective accumulative temperature is not satisfied, these bugs should not be able to survive in Denmark, Sweden, or Norway.

3.2 Analysis of occurrence (presence and pseudo-absence) maps and identification of important variables affecting western conifer seed bug distribution using the RF algorithm

Random forest modeling using test data predicted a higher WCSB occurrence in Northern Europe (the southern part of England, Belgium, and the Netherlands) than in Southern Europe, with 77.1% accuracy and AUC of 0.82 for the presence of WCSB (Figure 5). The TSS was 0.54, which was relatively low, but acceptable (Tobeña et al., 2016). In addition, the reliability of the RF model was evaluated using the OOB error rate for classifying either the presence or pseudo-absence of the WCSB, which indicated a reliability of 21%.

Given that higher values of MDA and MDG for each variable are taken to be indicative of greater importance, we established that the maximum temperature in March was the most important variable for classifying the presence or pseudo-absence of the WCSB (Figure 6), with the MDA and MDG values for this variable being 1.5–4 and 1.5–3 times higher, respectively, than those of the other assessed variables. It should be noted that these outcomes are reliable when the model is correctly developed, and this is why this study used previously accepted methodology for developing the model with reliable data and their sources. The effect of distance from coniferous forest and DEM on model predictions was identified based on the construction of density of the observed values and the independent variables in Europe. Although the MDA and MDG of distance from coniferous forest were very low, WCSB occurrence was recorded in the vicinity of coniferous forests, with approximately 78% being distributed within 10 km of a forested area. In addition, distance from coniferous forest was not used as an important variable in the model because coniferous forests existed in most of Europe, and this result means that WCSB is easy to find hosts anywhere. Meanwhile, we found that the occurrence of the WCSB was high when the distance from coniferous forest was short and the DEM was low. Moreover, approximately 80% of WCSB occurrences were predicted for regions with altitudes below 100 m a.s.l., such as southern England, Belgium, the Netherlands, Denmark, northern Germany, southern Sweden, and Poland, with the WCSB occurring at a high percentage (approximately 64%) in areas with an altitude of less than 50 m a.s.l. For example, the low-lying areas in northern Belgium had more records of WCSB occurrence than the higher altitude areas in southern Belgium. Comparatively, in South Korea, wherein WCSB generally occurs at altitudes of less than 200 m, altitude has been found to explain approximately 18% of WCSB occurrence (Kim et al., 2020).