Drivers of urban biodiversity in Mexico and joint risks from future urban expansion, climate change, and urban heat island effect

extension of the Kolmogorov normality test and a correlation test for the residuals obtained from the model-fitted values. We built a multimodel-averaged SAR curve by multiplying the predicted species richness of each model by its Akaike weight and then summing the resulting values [62]. Random forest models We used a supervised machine learning Random Forest (RF) algorithm to assess how environ- mental and socioeconomic variables drive urban biodiversity across the Mexican municipali- ties (Table 1). RF is a machine learning technique that allows us to evaluate different variables simultaneously by maximizing its predictive prediction power [63]. Furthermore, RF allows us to treat complex multidimensional datasets with complex structures. We trained RF models using total species richness for the four taxonomic groups and 19 variables: total urban area (i.e., log10 urban area), the total size of urban protected areas (log10), number of ecoregions overlapping urban areas, climate (bio1; bio5; bio6; bio12; bio15; cloud cover), topographic complexity, urban tree height, habitat heterogeneity, remnant vegetation, human density, demographic pressure index, and socioeconomic factors (GDP; marginalization; percentage of the population working in the primary sector). A full description of these variables and the hypothesized causality relationships with species richness is in Table 1. We explored various parameters to select the best configuration that minimizes the prediction error using the root-

mean-square error (RMSE). A cross-validation strategy was implemented, performing 100 random partitions of each dataset into a calibration (80% of the data) and validation (20%) dataset to avoid RF overfitting and increase model accuracy. Similarly, we quantified variable effects (i.e., effect sizes) for each variable by separating using partial dependence averaging over the effects of all variables with a non-parametric bootstrapping approach [64]. We imple- mented this approach using the rfUtilities R package [65]. This strategy allows us to interpret more directly how each predictor affects the urban species richness. We analyzed the same effect sizes but controlled urban areas’ effects on species richness by using residuals from a log- log linear regression between species richness and urban area. Multivariate future risk index We estimated the exposure to future urbanization across Mexican municipalities using land- use/cover-change (LUCC) scenarios recently developed for Mexico (see [27] for further details). The number of accumulated urban pixels projected toward 2070 within each munic- ipality was calculated using the assumptions of two scenarios, business-as-usual (BAU) and worst scenario (WRT) (Fig D in S1 Appendix). The annual mean temperatures for urban areas in each municipality were calculated using the climatological database from the period 1960–1990 from the Worldclim repository [56]. Future temperature projections for 2070 were obtained from an ensemble of 11 General Circulation Models and two Representative

Concentration Pathways (RCP; -RCP2.6 and RCP8.5-) from the CMIP5 database available in the Worldclim repository. The RCP2.6 is an emission scenario that represents the goals of the Paris Agreement and, therefore, limits the increase in global mean temperature to 2˚C [66]. In contrast, the RCP8.5 is a high-emission scenario with a combination of high popula- tion and limited technological change to improve energy consumption [67]. We calculated the temperature anomalies between present (1960–1991) and future (2070) for both RCP emission scenarios. This anomaly represents the projected shift in urban temperatures for 2070 in each municipality. Similarly, we calculate the difference between the historical urban scenario from INEGI for 2011 and the future projections for business-as-usual (BAU) and PLOS ONE Urban biodiversity and anthropogenic pressures PLOS ONE | https://doi.org/10.1371/journal.pone.0308522 October 4, 2024 6 / 19