Impact of data-partitioning strategies on the predictive performance and geographic patterns of machine learning-based species distribution modeling

Global occurrence records for four species—Lycorma delicatula, Metcalfa pruinosa, Spodoptera litura, and Linepithema humile—were obtained from the Global Biodiversity Information Facility (GBIF) to estimate their potential geographic distributions. To minimize sampling bias and enhance model accuracy, spatial filtering was performed using ArcGIS. To evaluate the models with independent datasets, the occurrence data were divided using four partitioning schemes: random, temporal, spatial, and spatiotemporal. For the spatial and spatiotemporal partitions, two strategies were implemented: region-based partitioning, in which test datasets were selected from designated regions (e.g., native ranges or main distribution areas), and proportion-based partitioning, where a fixed proportion of records from the target regions was used as test data. Climate data were obtained from WorldClim (https://www.worldclim.org) at a spatial resolution of 10 minutes for the period 1992–2021, including mean maximum temperature, mean minimum temperature, and precipitation. These variables were subsequently used to generate 19 bioclimatic variables through the biovars function in the dismo package within R. In addition, elevation data with the same 10-minute spatial resolution were also obtained from WorldClim. All bioclimatic variables and the elevation layer were converted into raster (.asc) format to facilitate their application in modeling using ArcGIS Pro. We used ENMeval in the R package to determine the model features and regularization multiplier (RM). Files and variables File: Occurrence data.zip: Occurrence data for MaxEnt model File: enmeval_code and results.zip: Example code for selecting the optimal model structure of MaxEnt and the corresponding results File: Model_variables.zip: Selected environmental variables for models under different data-partitioning strategies File: MaxEnt_results.zip: Results of species distribution models under different data-partitioning strategies