From the original list of 19 species of Marmosini from Colombia (Solari et al. 2013; Voss and Giarla 2021), we gathered information on 16 species, 13 from the genus Marmosa and 3 from Monodelphis. A total of 648 records were gathered and visually verified. The number of species records varied from 9 (Marmosa phaea and M. jansae) to 199 (M. robinsoni) after cleaning the data. From the 16 species included, 13 species had records within Colombia, while for the species M. regina, Monodelphis brevicaudata, and Monodelphis palliolata no records within the country were gathered, but were included because records fall within ecosystems that occur in the country.
The records of Marmosini, gathered from 19 different sources (Appendix 1 — Sheet 1), extended from 17.50º S and 86.1º W to 13.45º N and 51.763º W, and occurred within fourteen countries: Bolivia (n=17), Brazil (n=28), Colombia (n=101), Costa Rica (n=30), Ecuador (n=49), French Guyana (n=1), Grenada (n=2), Guyana (n=8), Nicaragua (n=8), Panama (n=71), Peru (n=88), Suriname (n=6), Trinidad and Tobago (n=16) and Venezuela (n=208).
Following are the analyses of the different aspects of models taken into account in this work, they represent the filtered results by selecting the upper quantile of the average test AUC metric, except otherwise stated. A complete account of the unfiltered results are available in Appendix 1 — Sheet 3.
Predictors scenarios varied when model performance metrics were analyzed. The two best scenarios based on train AUC and average test AUC were user-defined with and without vegetation (‘ud.all’ and ‘ud.noplants’), which performed similarly. While ‘onylwc’ and ‘uncorr’ scenarios had the worst performance, irrespective to the rm analyzed (Fig. 2a-b). In contrast, for the orMTP and at low rm values, all models performed similarly. As the rm value increased they gradually differentiated, with all models being similarly good except for ‘onlywc’ scenario which was consistently worst from a rm value of 2.5, thus with a higher rate of omitting presences (Fig. 2c). For the AICc metric and at the lowest rm value (0.5), models performed distinctively, but as the rm increased models seemed to become similar in their performance, but with both user-defined scenarios being slightly better based on this metric (Fig. 2d).
When the different modeling areas were compared regarding their overall performance with the metrics used here, M2 outperformed M1 in most of the cases with different predictor scenarios, cross-validation type and rm (Online Resource 2 — Fig. S2). For train and test AUC, M2 was consistently better than M1, except for jackknife cross-validation with which models from M1 and M2 seemed to performed comparably (Online Resource 2 — Fig. S2a-b). When evaluated with orMTP, both areas performed similarly with only small differences in the high values obtained for M2, jackknife and ‘onlywc’ case from rm 1 to 2.5, and in the block type cross-validation, where M2 was slightly better than M1 (Online Resource 2 — Fig. S2c). However, for AICc M1 was better than M2 in most of the comparisons (Online Resource 2 — Fig. S2d).
A total of 8,320 models were generated, 2,080 for each predictor scenarios, and between 120 and 280 models for each species, depending on numbers of records (Fig. 1c). A total of 64 models were visually and critically inspected to decide based on biological data, which best represented each species’ distribution. Most of the chosen models were from the predictors scenarios based on ‘ud.all’ (n = 7), followed by ‘onlywc’ (n = 4), ‘ud.noplants’ (n = 3), and finally ‘uncorr’ (n = 2). Regarding cross-validation, 9 out of 16 final models were from those using random k-fold method, and 7 out of 16 used block method. Most of the models were from regularization multipliers smaller than two and had a train AUC > 0.7. Other metrics and configurations of the final models are presented in Table 1 and Appendix 1 — Sheet 3.
Variable contribution and permutation importance varied greatly between species (Fig. 3). Precipitation related variables were the most important (with values above 50%) in the models generated for 11 species, and temperature related variables were the most important in 4 (M. isthmica, M. phaea, M. rutteri, and M. xerophila) (Fig. 3 and Appendix 1 — Sheet 4). In M. regina, temperature related variables, terrain roughness (tri) and MSAVI were the most important variables (Fig. 3 and Appendix 1 — Sheet 4). In seven species precipitation related variables represented the most important contribution and permutation values, while only three species had contribution and permutation values represented by temperature variables exclusively (Appendix 1 — Sheet 4). Six species had a combination of precipitation and temperature (M. alstoni, M. lepida and M. waterhousei), or temperature and precipitation (M. xerophila, M. zeledoni, and M. brevicaudata), as the highest contribution and permutation variables, respectively. The first three species had Mean Temperature of the Coldest Quarter (bio11) as the variable with the highest information not present in the others (permutation importance) but different variables with the highest contribution to their models (i.e., bio15 for M. waterhousei, bio16 for M. alstoni, and bio17 for M. lepida) (Appendix 1 — Sheet 4). The variable “topographic wetness” was markedly high for Monodelphis species, especially for M. adusta and M. palliolata. The best model configuration for M. phaea resulted in a single variable contributing to the model, Mean Temperature of the Warmest Quarter, and the model for M. germana with only two variables (Precipitation Seasonality (Coefficient of Variation) and Precipitation of the Driest Quarter) (Appendix 1 — Sheet 4).
After visual inspection of final models, geographical barriers were taken into account to modify and generate final range maps. A complete account of the geographical barriers proposed for delimiting each species and final range maps are in Online Resource 3. Different spatial patterns were found regarding their distribution in the country. To describe them, we followed the national categorization of continental biogeographic regions for local relevance and clarity, published by the Instituto de Investigación de Recursos Biológicos Alexander von Humboldt of Colombia (maps available at https://www.redalyc.org/articulo.oa?id=49150103 and reproduced in inset map of Fig 4a).
Marmosa germana, M. jansae, M. regina, M. rubra, M. rutteri, and M. brevicaudata showed a distribution restricted to the Amazonian region, with several species showing small distribution areas in the Amazonian-Andean transition, east of Nudo de Los Pastos. Marmosa lepida and M. phaea were distributed in the Andean, Orinoquia and Amazonian regions, with the former being mainly distributed in the Amazonian and Orinoquia regions, and a distribution area at mid-elevations of the Eastern Andean region, while M. phaea was mainly distributed in the Andean region and few areas in the Amazonian and Orinoquia region, especially east of Nudo de los Pastos and Sierra de La Macarena. Marmosa waterhousei and M. adusta were distributed mainly in the Andean, Pacific, and Amazonian regions, with the former being widely distributed in the Andean and Amazonian regions and partly in the Pacific and Orinoquia regions, with few areas in the Caribbean region, while the latter is mainly distributed in the Pacific and Andean region, including the Sierra Nevada de Santa Marta at the Caribbean region, and a few spots predicted in the Amazonian region. Marmosa alstoni and M. zeledoni showed an Andean-Pacific distribution, with the former being more widely distributed in the Andean region compared to the latter, and with some areas in the Serranía de San Lucas (Caribbean). Marmosa robinsoni and M. isthmica were distributed mainly in the Caribbean region and the inter-Andean valleys, especially the Magdalena river valley, with the former having more predicted areas in the Andean reagion, and the latter with more predicted areas in the Pacific region. The distribution of Monodelphis palliolata was mainly concentrated in the Andean and Caribbean regions, limited to mid-low elevations at the northern end of the Eastern Andean region, Catatumbo, and low elevations of the Sierra Nevada de Santa Marta. Finally, M. xerophila with a restricted distribution, concentrated in the Guajira, north of the Caribbean region.
The areas of the final ranges varied from 681,717 km2 in M. waterhousei to 8,526 km2 in M. rubra, with a median area of 135,517 km2. Taken together, the combined distribution of all Marmosini species’ distribution covers 1,038,318 km2 of continental Colombia, about 91.4% of the continental area of the country, being absent only in parts of the Orinoquia region, North to the Meta river and East of the Eastern Andes near the border with Venezuela, and two small portions one in the northwestern end of the Western Andes (Chocó department), and the other in the southwestern coast near Tumaco (Nariño department).
- Richness and conservation metrics
Richness of Marmosini in Colombia varied from a maximum of 10 species to a minimum of 1 species per 25 km2 (Fig. 4). We found a clear pattern of maximum and sub-maximum richness concentrated at mid-elevation slopes of the Andes, with an approximate upper elevation limit of richness at 2000 m, especially for the Central and Eastern Andes (Fig. 4b). The highest richness (10-8 species) was found in the Amazon-Andes transition, east of Nudo de los Pastos, and southeast of the Colombian massif (Fig. 4c). The mid-high richness (7-5 species) was found mainly at the eastern slope of the Eastern Andes, Sierra de La Macarena, Catatumbo (northwest of the Táchira depression at the limit between Colombia and Venezuela in the Eastern Andes), Serranía de San Lucas northwest of the Central Andes, in the northern end of the Western Andes, the northwestern coast of the Pacific region and southern Amazonas in the region contained between the Putumayo and Caquetá rivers (Fig. 4a-b). The mid-low richness (4-2 species) was found mainly at the Amazon (north of Caquetá river), the transition zone between Amazonas and Orinoquia regions, inter-Andean valleys, Central Andes, high elevations of the Andes including Eastern Andes and most of its western slope, Pacific region (excluding areas in the northwest with mid-high richness), Sierra Nevada de Santa Marta, and much of the Caribbean region (Fig. 4a). Interestingly but with a lower richness compared to other areas, a local upper limit of high richness occurs at ~2000 m in Sierra Nevada de Santa Marta (Fig. 4a). The lowest richness (1 species) was found mainly in lowlands of the Orinoquia region, mid/low Magdalena river valley, Guajira, some areas of the Pacific region, and in the highest elevations of the Andes and Sierra Nevada de Santa Marta (Fig. 4b).
Of the total modeled areas, conservation within each species was highly variable, ranging from 29.9% in M. rutteri to 5.3% in M. xerophila, and a median of 15% (Table 2). This shows that roughly 85% of all distribution areas for Marmosini lack effective protection. Of the total preserved areas for each species, areas with strict conservation preserved between 28.9% in M. rutteri, and 2.3% in M. xerophila, with a median of 10.4% (Table 2), with large unprotected areas for all species. The areas preserved under managed resources ranged from 6.9% in M. zeledoni to 0% in M. brevicaudata, with a median value of 2.7% (Table 2). Human pressure within each species area ranged from 76.2% to 0% (M. robinsoni and M. brevicaudata, respectively) and a median of 27.9% for high pressure, and from 94.3% to 23% (M. germana and M. robinsoni) and a median of 71.1% for low pressure (Table 2). No data values regarding pressure varied from a maximum of 7.6% in M. brevicaudata to 0.3% in M. phaea (Table 2).
Our conservation-only analysis based on IUCN criteria showed that most of the species had more preserved area under strict conservation (median of 11.98%), than area under managed-resources (median of 2.24%). Conservation based on governance showed that most protection comes from national governed areas (median of 11.66%), followed by sub-national governed areas (median of 3.83%), with private areas representing very little of the protected areas (median <0.1%) (Table 3). Among species, M. rutteri had the highest percentage of its area strictly protected while M. xerophila had the lowest, which were mostly under national governed areas. The percentage of sub-national governed areas was higher than national governed areas in M. isthmica, M. robinsoni and M. xerophila (the last two with values slightly above the national governed areas; Table 3). No protected areas of Indigenous territories and local communities were found throughout the distribution of Marmosini in Colombia.
Conservation-pressure analysis showed that regions within species ranges that are under strict-conservation have a lower area under high pressure (median of 0.54%) compared to areas under managed-resources (median of 38.9%) (Table 4). Within governance-types, the lower median values were for areas under national governance (1.06%), followed by sub-national areas (33.72%) and private areas (62.09%). This pattern is consistently found throughout all species of Marmosini analyzed (Table 4). These results show that Marmosini in Colombia are more exposed to higher human pressure in managed, sub-national and private areas, while strict reserves and national governance areas have smaller percentages of areas under high pressure throughout the species’ ranges.