We estimate the potential distribution area and habitat connectivity of the Antioquia Brushfinch (Atlapetes blancae), a critically endangered bird endemic to the Altiplano de Santa Rosa de Osos in Antioquia, Colombia. We generated two essential inputs to guide its conservation: I. a hypothesis about its potential distribution, and II. a landscape connectivity model. We identify three priority areas for its protection, considering its occurrence, as well as its ecological niche and habitat requirements (connectivity models). We used the maximum entropy algorithm (MaxEnt) and minimum volume ellipsoids (MVE) to estimate the niche and map its potential distribution. We developed a habitat model and used it in conjunction with the potential distribution to generate a map of landscape connectivity for the species. Our results support a potential distribution composed of two disconnected nuclei: an epicenter to the north of the Altiplano (513.26 km² according to MaxEnt and 300.12 km² according to MVE) and another to the south (133.1 km² according to MaxEnt and 101.53 km² according to MVE). Connectivity models indicate that suitable habitat for the species is highly fragmented (effective network size for MaxEnt and MVE, respectively, was 18.88 km² and 14.08 km²). Finally, we highlight two key zones for connectivity that should be prioritized in future exploration campaigns, since there is a high probability of finding new populations in these places, the first one is located to the north, between the limits of Santa Rosa de Osos and Angostura, and the second one to the south of the Altiplano, on the limits of Bello and San Pedro de los Milagros. Given the vulnerability of this species and its disproportionate importance to conservation, we expect our results will be instrumental in establishing protected areas and enacting sustainable practices to enhance the conservation of Atlapetes blancae. All occurrence records used in the models are provided, in addition to consensus models for MaxEnt and Minimum Volume Ellipsoids (50-90%), habitat patches under both modeling strategies, and the areas prioritized for conservation. 

Ecological niche models and potential distribution

To estimate the environmental preferences (ecological niche) for A. blancae and project them into geographic space (potential distribution), we modeled the relationships between occurrence records and environmental variables, using two approaches: the maximum entropy algorithm, MaxEnt (Phillips et al. 2004), and minimum volume ellipsoids (MVE). For both MaxEnt and MVE, a calibration and model selection protocol based on statistical significance, predictive ability, and complexity was implemented.

Occurrence records

Occurrence records were obtained from field data previously collected by the “Proyecto Atlapetes” (https://arcg.is/1yz0eK), by the “Iniciativa de conservación del Montañerito Paisa” (iCMP, http://reporte.humboldt.org.co/biodiversidad/2020/cap3/302/), complemented with the eBird (https://ebird.org/home) and GBIF (https://www.gbif.org/) databases. All occurrence records were reviewed for quality assurance following protocols proposed by Cobos et al. (2018). We eliminated records less than 1 km away from other occurrences to minimize spatial autocorrelation and possible biases in the observations.

Accessibility area (M)

The genus Atlapetes includes 30 species, most of which are found in mountainous systems and have limited distributions (Donegan et al. 2009). In general, their dispersal capacity is low due to their short wings and is inefficient for long flight distances (Capllonch et al. 2014). For this reason, considering the ecology of these species and the geomorphology of the ASRO, we defined an area of accessibility above 1500 meters for A. blancae, using the digital elevation model of Robinson et al. (2014).

Predictor variables for ENM

The predictor variables were selected based on previous research on the species, and studies in species of the same genus, such as Atlapetes pallidiceps (Oppel et al. 2004) and Atlapetes flaviceps (Botero-Delgadillo et al. 2022; Chaparro-Herrera et al. 2020). For the construction of the ecological niche model we considered 53 variables from five different sources; 19 bioclimatic variables from Wordclim (Fick et al. 2017), 11 soil variables obtained from the SoilGrids database (Poggio et al. 2021), a forest canopy height layer from (Potapov et al. 2021), a forest canopy percentage layer (Contreras-Díaz et al. 2022; Tuanmu and Jetz, 2014), 11 soil-level bioclimatic variables (Lembrechts et al. 2021) and 10 variables related to geomorphometric characteristics from Geomorpho90m (Amatulli et al. 2020).

Predictor variables were processed in R software version 4.2.1 (R Core Team, 2022), using Terra version 1.7.46 (Hijmans, 2022a) and raster version 3.6.23 (Hijmans, 2022b) packages. The layers were aligned, reprojected in the same coordinate reference system (EPSG:4326 - WGS 84 - Geographic), and clipped according to the area of accessibility (M). Predictor variables (forest canopy height, soil properties, geomorphometric characteristics) were resampled to ~1km2 resolution by assigning the average value of the cells. We performed Spearman's bivariate correlation test (Spearman, 1904) between predictors, selecting variables with the highest ecological relevance and a correlation of less than or equal to 0.7.

ENM calibration and evaluation

To choose between niche models, we evaluated combinations of parameters of each algorithm (MaxEnt and MVE) to maximize a set of evaluation criteria. For the MaxEnt algorithm, we evaluated all combinations between thresholds (0.1, 0.5, 1, 2, 3) and relationship types (linear (l), quadratic (q), threshold (t), hinge (h)) and their respective combinations, for 20 different sets of environmental variables (Table S2). Each one of these sets included between five and seven variables. We evaluated model performance based on three criteria: statistical significance evaluated by partial ROC tests (with 500 iterations and 50% data for bootstrap; Peterson et al. 2008), predictive ability assessed through omission rate (less than 5%), and the Akaike criterion corrected for sample size (AICc) as an indicator of model complexity (Warren and Seifert, 2011), where a threshold of two AIC units was set to identify models with equal performance (Cobos et al. 2019). The omission rate threshold selected (<5%) rests on our high confidence in the occurrence points, and the AIC threshold follows general rules of thumb from model selection theory (Burnham and Anderson 2004).

For the MVE algorithm, we used as evaluation criteria the omission rates (< 5%), the significance of the partial ROC test (p ≤ 0.05), and the AUC value (> 0.9) (Osorio-Olvera et al. 2020b). Three, four, and five-dimensional ellipsoid models were constructed and evaluated using all possible combinations of the selected variables. To calculate the partial ROC test and the AUC, we used 4968 background points in relation to the proposed accessibility area.

Final ENM

From the final set of selected models, we generated a binary consensus layer where a value of 1 was assigned to those pixels with at least 80% congruence in presence prediction among all models that met our performance criteria. This process was developed for both modeling algorithms (MaxEnt and MVE).