Environmental factors determining the occurrence of the canids
We only used the records of the three canid species occurring within Brazil to analyze which environmental variables and landscape structures influence each species’ occurrence. We limited it to Brazilian points using the Landsat classified image database from MapBiomas. v5.0 collection available just for this country (https//:mapbiomas.org). In the GeoTiff scenario, this collection informs the annual mapping of land cover and land use in Brazil between the historical period of 1985 to 2019 (MapBiomas, 2021). We listed the coordinates of each occurrence of A. microtis , with information on scenarios in the Amazon Biome, from 2002 to 2019, a period that includes all records of occurrences of this species. We did the same thing for C. thous , but using data from 1988 to 2016, including the Amazon, Caatinga , Cerrado , Atlantic Forest,Pantanal , and Pampas biomes. For S. venaticus , we used data from 1900 to 2017, including biomes from the Amazon,Cerrado , Caatinga , Atlantic Forest, and Pantanal(Figure 3).
For each coordinate of occurrence of each species, we delimited a buffer with a radius of 1 km, from which we extracted information on land use. We obtained the pixel values ​​and land-use class codes in each buffer (available at: https://mapbiomas.org/downloads_codigos) and transformed these values ​​into categories of percentages. If the buffer areas overlap by more than 20%, the rightmost occurrence point was excluded for this analysis. After eliminating buffers with overlapping criteria, we selected 277 independent records with a minimum distance of 2 km between them. We set 64 records of A. microtis , 63 of C. thous and 100 of S. venaticus (Supplemental Material-Table S2).
About 16 land-use classes were recorded in the occurrence scenarios of the buffers. Some classes with similar landscape effects were grouped into new more significant categories, and others with lower registration frequency than three occurrence points were discarded (Supplemental Material-Table S2). In the end, we considered five significant land use categories to assess the environmental factors that determine the occurrence of species. They are: (Forest) Percentage of forest cover, which includes all native forest formations of the Amazon; (Water) Percentage of water bodies, including streams, rivers, lakes and non-forest natural wetlands; (Nat_open_areas) Percentage of natural open areas, including savannas, grasslands and rocky outcrops; (Urban_areas) Percentage of urban infrastructure; and (Anthr_open_areas) Percentage of anthropogenic open areas, including pasture, perennial annual crop, agriculture and pasture mosaic (Appendix-Figure 2, Supplemental Material-Table S2).
In addition to the land-use variables, we also extracted the altitude (Alt) and the Enhanced Vegetation Index (EVI). The altitude was extracted from each point of occurrence of the species from the satellite images of Google Earth Pro (version 7.3.1). The Enhanced Vegetation Index (EVI) is an index related to plant complexity and heterogeneity, since it is based on plant cover and the fraction of absorbed photosynthetically active radiation (Gurung et al., 2009). The EVI was extracted from the Earth Explorer platform (available at: https://earthexplorer.usgs.gov/). For each year of occurrence of the species, we extract images in the rasters format and calculate the statistical average of the index using the Quantum Gis software(v2.8.4).
To avoid data multicollinearity, we used Principal Component Analysis (PCA) to select correlated variables from the first two axes for further explanation of the results (Kassambara, 2017) (Appendix-Figure 2). We used the Kaiser criterion to select the variables by eigenvalues, not including those that were correlated on the same axis (Jackson, 1993). We selected the following predictor variables to be used in Generalized Linear Models (GLM’s) analyses: Forest, Water, Nat_open_areas, Urban_ areas, and Anthr_open_areas (Figure 3, Supplemental Material-Table S2). We constructed GLM’s with binomial distribution and logit linkage function to verify the influence of environmental predictors on the occurrence of the three canids. We initially ran all models with binomial distribution and evaluated the dispersion assumption. In the case of A. microtis , which had a dispersion greater than 1, we used quasibinomial correction. To analyze the possible effects of predictor variables on the occurrence of the three species, we built six models considering different combinations of the predictor variables (Appendix-Table 2). We use the glm function (Hardin et al. 2007) from the stats package (Marschner and Donoghoe, 2018). We evaluated the dispersion of residues to confirm the absence of overdispersion. We then ran the analysis and selected the best model using the table ANOVA criterion, including the null model, and we used the data generated in the GLM’s analysis to interpret the size measure of the effects of the predictor variables on the occurrence of species (Figure 3). We considered values for p < 0.05 as level of significance.