Spatial distribution models
We gathered information on the distribution (= occurrences) of the Strigidae in Brazil from: (1) skin specimens deposited in several museums according to the Global Biodiversity Information Facility (GBIF.org, 2019); (2) more than 164 publications in peer-reviewed literature regarding taxonomic assessments, fauna inventories or owl biology; and (3) field records from the bioacoustics database www.xeno-canto.org. We provide the citations of these sources in Supplementary material A, Appendix 2. The quality of the geographic coordinates varied from GPS recordings until those of the nearest town listed on the specimens’ labels. We corroborated the localities through an ornithological gazetteer specific for Brazil (Paynter & Traylor, 1991) and online (www.geonames.org).
There are no records for the buff-fronted owl (Aegolius harrisii ) in Northern Brazil, but in the nearby Northern border at both Cerro de la Neblina (Willard et al., 1991) and Roraima Tepui (Braun et al., 2003). Similarly, most of the records for the foothill screech owl (Megascops roraimae ) come from outside Brazil in Cerro Urutaní (Dickerman & Phelps, 1982), Cerro de la Neblina (Willard et al., 1991), Acary Mountains (Robbins et al., 2007), and Roraima Tepui (Milensky et al., 2016). In both cases, we included these records in our analyzes by reassigning coordinates within their respective closest Brazilian territory. The Pernambuco pygmy-owl is known from two localities (J. M. C. da Silva et al., 2002), to which we added eight random points located within a polygon resulting from two merged circles, each centered in one of the known localities and radius equaling the distance between both, clipped by the neighbor coastline. We excluded a record of the short-eared owl (Asio flammeus ) in the Roraima State (wikiaves.com.br; consulted on April 10, 2021), likely belonging to the subspecies A. f. pallidicaudus from “Venezuela, Guyana and Suriname” (Gill et al., 2021).
The geographical and environmental clustering of field surveys, known as spatial autocorrelation (Araújo & Guisan, 2006; Loiselle et al., 2008), can negatively affect the performance of the SDMs (Veloz, 2009). Consequently, some authors remove those records under the same environmental conditions within an arbitrary distance (Delgado-Jaramillo et al., 2020). Thus, we created two datasets for each species, one including all the records and another excluding those closer than 25 km, and computed empirical entrograms for both using “elsa” (B. Naimi et al., 2019), comparing the entropy-based local indicators of spatial association for both categorical or continuous environmental covariates. Entrograms are variogram-like graphs quantifying the spatial association of geographical covariates based on information entropy concepts (B. Naimi, 2015).
We used “ENMeval” (Muscarella et al., 2014), a package based on Maxent (Phillips et al., 2006, 2017, 2004), that automatically splits data into training/test subsets, performs SDMs across a range of settings, and calculates diverse evaluation metrics. For each taxon, we ran 10 models, each one after partitioning occurrences in testing and training bins using a 10-fold cross-validation scheme (Fielding & Bell, 1997). For each run, we created 10 000 pseudoabsence points distributed randomly throughout Brazil and selected the model with the lowest Akaike information criterion corrected for small samples sizes (ΔAICc =0) as the best one, since it reflects both model goodness-of-fit and complexity (Burnham & Anderson, 2002; Warren & Seifert, 2011) and less overfitting (Muscarella et al., 2014).
Different habitat suitability thresholds may disagree in terms of suitable areas and omission errors (Bean et al., 2012; Liu et al., 2016; Nenzén & Araújo, 2011). Thus, for each taxon, we plotted the extension of the predicted area (in number of pixels) against the number of omissions and compared across taxa the performance of the different thresholds, keeping the one that consistently provided the lowest values for both measurements. The final binary models combined the best models (ΔAICc = 0) and the threshold with the lower number of omissions within the smallest predicted area. We stacked these binary distributions to create two maps of taxa richness for (A) the 21 species evaluated (hereafter sensu stricto map), (B) the 12 monotypic species (including polytypic ones represented by only one subspecies in Brazil) and 21 subspecies (henceforth sensu latomap). We overlapped the protected areas distributions corresponding to IUCN’s categories I to IV (according to Protected Planet 2021) on each richness map.