Impacts of climate change on high priority fruit fly species in Australia

and GFDL-ESM2M (Global Coupled Climate Carbon Earth System Model Part-1). CanESM2 projects a hot future with drying across central regions of Australia and higher precipitation in the north-east. ACCESS1.0 projects a hot and dry future across most areas of Australia, while MIROC5 projects moderate warming, with drying in the north-east and south-west but higher precipitation in central Australia. NorESM1-M projects moderate warming. HadGEM2-CC and GFDL-ESM2M project a hot future with greater warming typically in central regions. We downloaded the 19 bioclimatic variables from these six models from CCAF, for 20-year periods centred on 2030, 2050 and 2070, for the Representative Concentration Pathway 8.5 (RCP8.5) [43]. With a radiative forcing exceeding 8.5 Wm-2 by 2100, this is the highest of the four RCPs presented in the Intergovernmental Panel on Climate Change’s Fifth Assessment Report [43]. It is also the RCP that emissions are currently tracking most closely [44]. After downloading, we reprojected data to a spatial resolution of 1 × 1 km (Australian Albers Equal Area, EPSG: 3577) via bilinear interpolation, using the gdalwarp function provided by the R package gdalUtils [45] in R version 3.1.2 [46]. Species distribution models We used the machine learning approach, Maxent (v3.3.3k [27]), to assess climate suitability for species under baseline and future climate scenarios. Maxent accommodates presence-only data and has performed well in multimodel assessments [47]. It produces a continuous proba-

bility surface, which can be interpreted as an index of climatic suitability given the predictor variables included in model calibration. Detailed descriptions of Maxent are given elsewhere [48, 49]. We optimized models by assessing the effects of different combinations of feature types, of competing predictor sets deemed ecologically sensible a priori, and of the extent of regularization on model performance. We found that Maxent performed best when product (first-order interactions), linear and quadratic features were used, with a regularization multi- plier of 1 (the default), and used this configuration to calibrate our final models. Maxent requires background data, to which it compares the environmental characteristics of presence locations. There is flexibility for users to specify which points to use as background, as well as the number of records and the spatial extent from which they are chosen [48]. Fol- lowing Ihlow et al [50], we generated background points by randomly selecting up to 100,000 cells from terrestrial areas within 200 km of occurrence records of the target species. Our choice of background achieves a balance between fine-scale discrimination of suitable and unsuitable sites along environmental gradients, and generalization of model predictions. To assess model performance, we used five-fold cross-validation to reduce model errors that may occur from the random splitting of data into test and training subsets. The perfor- mance of each model was evaluated using the area under the receiver operating characteristic

curve (AUC), which describes the consistency with which a model ranks randomly chosen presence sites as more suitable than randomly chosen background sites. AUC ranges from 0 to 1, with a value of 0.50 indicating discrimination ability no better than random, while values greater than 0.75 indicates that the model has a discriminative ability that is better than “fair” [51]. Cross-validated AUC scores were presumed to reflect the performance of a single final model for each species, which used all available data. Following previous studies of pest species [26], continuous suitability scores projected by Maxent models were converted to binary layers (0 = unsuitable, 1 = suitable) using the 10th percentile training presence threshold (i.e. the value that corresponds to 10% training omis- sion). We note that the selection of a threshold value may vary depending upon the goals of the study [52], thus we also provide continuous output for baseline climate as supplemental data (S1 File). For each species, the six binary suitability grids (i.e., one for each GCM, with Impacts of climate change on high priority fruit fly species in Australia PLOS ONE | https://doi.org/10.1371/journal.pone.0213820 February 13, 2020 5 / 19