2.5 Influence of landscape on gene flow and genetic distance
To preliminarily predict the influence of the different landscape types on the genetic differentiation of M. alternatus , the distance models IBD and LCP were used at the fine scale (with Shunchang as an example). In the LCP model, four groups of resistance values were set for the ten landscape types base on the opinion method (Supporting information Table S5): i ) host landscapes and roads were given the lowest resistance value (the resistance value of P. massoniana , P. elliottii , mixed-forest with hosts, and roads was 1), and nonhost landscapes (except roads) were divided into medium and high resistance (the medium resistance value of broad-leaved forest,C. lanceolata , water, and nudation was 8; the high resistance value of farmland was 64 and for urban was 512); ii ) host landscapes and roads were given the medium resistance value of 8, whereas nonhost landscapes (except roads) were given the lowest resistance value of 1; iii ) host landscapes and roads were given the medium resistance value of 64, whereas nonhost landscapes (except roads) were given the lowest resistance value of 1; iiii ) host landscapes and roads were given the high resistance value of 512, whereas nonhost landscapes (except roads) were given the lowest resistance value of 1. The LCP on the four resistance surfaces was calculated using the R package gdistance (Vanetten, 2014). Finally, Mantel tests between the two distances and genetic distances (FST(1-FST)) were performed in the R package ecodist (Goslee & Urban, 2007).
In addition, to maintain objectivity, the least-cost transect analysis (LCTA) developed by Strien et al. (2012) was used to study the influence of landscape types on dispersal behavior and gene flow of M. alternatus at the fine scale. This analysis provided an effective solution to objectively define the resistance values of different landscape types based on transects. In the analysis, each landscape type in a linear path, with different transect widths, between two sampling points is quantified, and the correlation between those values and the genetic distance is analyzed. The analysis can also determine the probable migration habitats and the landscape types that either inhibit or facilitate the gene flow. However, it is unlikely that species will migrate directly between two points following a straight line. Therefore, the LCTA uses the least-cost path to replace the straight line between two points and then calculates the percentage of each landscape type in the transect. For each transect, the proportions of each of the landscape types and the transect length were used as explanatory variables.
In this study, the first step in the LCTA was to create a series of resistance surfaces, in which each landscape type (P. massoniana ,C. lanceolata , P. elliottii , mixed forest (including some hosts), broad-leaved forest, urban, farmland, road, water, and nudation) was successively considered as the optimal dispersal habitat. The optimal dispersal habitat was given the lowest resistance value of 1, whereas the other landscape types were given gradually increasing resistance values (23, 26, 29) (Supporting information Table S6). A variety of transect widths (200 m, 300 m, 600 m) were evaluated. In each transect, Geospatial Modelling Environment v0.7.4.0 (Beyer, 2015) was used to calculate the proportions of each of the landscape types and the transect length. A total of 90 LCP data sets were obtained (10 land cover types × 3 resistance values × 3 transect widths). The correlations between those variables (proportions of each of the landscape types and the transect length) and the FST were tested through maximum-likelihood population effects (MLPE) (Clarke et al., 2002). In each MLPE, the model was estimated with restricted estimation maximum likelihood (REML), and the individual deme was used as a random effect. To explain the relationships between the different landscape types and the gene flow, Akaike’s Information Criteria (AIC) (Gurka, 2006), which is considered to be effective in choosing an REML hybrid model, was used to select the optimal model from the 90 data sets.