The study area extends over the Amazonian rain forest biome and contains the drainage basins of the Amazon and some other large rivers (e.g., Orinoco) but excludes areas that are mostly covered by non-forest vegetation or are at elevations higher than 600 m above sea level (Fig. 1). The exact study area was delimited by the explanatory raster layer with the smallest extent, namely the soil base cation concentration map of Zuquim et al. (2023). See below (section “Environmental predictors”) for more information on the explanatory variables.

Long-term satellite data indicate that most of the area has a tropical rainforest or monsoon climate, with mean annual rainfall exceeding 2,400 mm and mean annual temperature exceeding 25 °C (Malhi et al. 2009; Carvalho et al. 2020). The wettest areas are around the Equator near the Andean foothills in the west, and the climate becomes progressively drier as one moves further away from the core rainforest area towards the south and east.

We compiled species occurrence data both from quantitative field surveys that we had conducted over 30+ years (Tuomisto et al. 2024) and from occurrence records available in the Global Biodiversity Information Facility (GBIF) database (www.gbif.org). GBIF data for all Adiantum species were downloaded using a polygon that covered all of Amazonia with a wide margin (https://doi.org/10.15468/dl.cf8b38). Only occurrences within the extent of the environmental layers used in modelling were retained for analysis. To ensure that the GBIF data used the same species delimitations as the field data from Tuomisto et al. (2024), the identities of all GBIF occurrences were verified (and updated if necessary) by HT with the help of photographs of the herbarium specimens and field observations that were available through the GBIF portal. Specimens were excluded if their identity could not be verified, or if they represented species that were not present in our field data.

No specimens collected before 1980 were used, because their georeferencing was considered unreliable (their labels generally had no coordinates, as accurate maps and GPS devices did not exist at the time of collection). Records with known georeferencing problems and those whose coordinates were given with a precision of less than two decimal degrees in GBIF were excluded. Additional coordinate cleaning was done using the CoordinateCleaner package in the R software (Zizka et al. 2019). Duplicate records were removed before the selection of the final occurrence data, and a spatial thinning step using the “spthin” package (Aiello-Lammens et al. 2015) was performed to maintain a spatial distance of at least 450 m between the occurrences to reduce spatial sampling biases. We took care of data cleaning, positional uncertainty, and sampling bias following the recommendations of (Moudrý et al. 2024). Ideally the analyses would have at least 10 occurrences per variable, but this is not a strict rule (Benavides Rios et al. 2024). The final dataset contained 3570 occurrence points for a total of 18 Adiantum species (Fig. 1, Suppl. material 1: fig. S1): 2611 occurrences were from our field data, and 959 were obtained from GBIF. The source of taxonomic nomenclature for the 18 Adiantum species is World Ferns database (https://www.worldplants.de/world-ferns/ferns-and-lycophytes-list; Hassler 2026). Species that retained less than 20 occurrence records were not modelled (Benavides Rios et al. 2024). Final models were based on between 27 and 488 occurrence points per species (Suppl. material 1: table S3).

We obtained several raster datasets with predictors for the species distribution models (SDMs): climatic (Worldclim version 2.1; Fick and Hijmans 2017), atmospheric and vegetation (Myneni et al. 2021; Karger et al. 2023), microclimatic (Maeda and Ismaeel 2023), soil properties and soil types (SoilGrids250m v2.0; Poggio et al. 2021; Zuquim et al. 2023), and elevation (Shuttle Radar Topography Mission SRTM; Farr et al. 2007). We selected both continuous soil properties (6 predictors) and individual World Reference Base (WRB) soil group probability maps (21 predictors) as separate raster layers in our models (Suppl. material 1: table S1). All modeling was confined to within the extent of the soil base cation concentration raster, which had applied a forest/non-forest classification layer and only provided data for forested pixels at less than 600 m elevation.

A two-step procedure was used to select a subset of the predictors. Firstly, a total of 40 predictors were screened and grouped into six categories (atmospheric-vegetation, bioclimatic, microclimatic, soil types, soil properties, and topographical) (Suppl. material 1: table S1). The screening was based on their wide use in earlier studies (Figueiredo et al. 2018; Dambros et al. 2020; Silva et al. 2024), data quality, and correlation among predictors. These predictors were resampled to a common resolution of 450 m (corresponding to the resolution of the soil base cation concentration map) using the nearest neighbor method. Correlation was investigated using the variance inflation factor (VIF) function from the “usdm” package in R software. A threshold value of > 7 was used in VIF as an indicator of multicollinearity among predictors. A list of the predictors and their sources is shown in Suppl. material 1: table S1. Secondly, a pre-assessment of species-specific predictor importance was made within those categories that contained multiple predictors using a boosted regression tree model (Irving et al. 2020). For each species, the most important 30% of the predictors in each category were then selected for modelling. When SDMs are applied to multispecies having varying sample sizes, the selection and number of predictors should be carefully performed to avoid potential overfitting, unstable variable importance, and poor estimation of response curves (Merow et al. 2014; Moudrý et al. 2024). A list of final predictors used for each species is shown in Suppl. material 1: table S2.

To model future species distributions, we replaced the bioclimatic variables used to train the SDMs with future climate scenarios downloaded from Worldclim version 2.1 (Fick and Hijmans 2017). We selected two shared socio-economic pathways (SSPs) that have been commonly used in climate change studies, namely SSP2-4.5 (medium level of CO₂ emissions) and SSP5-8.5 (high level of CO₂ emissions) for the periods 2021–2040 (here called ‘2030s’), 2041–2060 (‘2050s’), and 2061–2080 (‘2070s’) (Monteiro et al. 2023; Abebe et al. 2024). For each SSP, we downloaded the raster layers corresponding to five Coupled Model Intercomparison Project Phase 6 (CMIP6) (Eyring et al. 2016) Global Climate Models (GCMs): ACCESS-CM2 (Bi et al. 2020), BCC-CSM2-MR (Wu et al. 2019), CMCC-ESM2 (Lovato et al. 2022), EC-EARTH3-Veg (Wyser et al. 2020), and MIROC6 (Tatebe et al. 2019). These GCMs take into account recent land use changes, emission scenarios, and improved parameterization of physical and atmospheric processes (Eyring et al. 2016; Shiru et al. 2022; HamadAmin and Khwarahm 2023; Monteiro et al. 2023; Amindin et al. 2024). To reduce uncertainty in the SDM predictions, we derived a consensus bioclimatic variable for each SSP by averaging the values from the corresponding five GCMs (Song et al. 2024). Since soils and topography can be expected not to change much within the 55-year modelling period, the same explanatory variables related to these were used in the SDMs of the present and the future. We used datasets of the highest available spatial resolution (30-arc seconds, ~ 1 km) and resampled them to 450-m resolution using the nearest neighbor method.

We used the package biomod2 v.4.2-5 available in the R software (Thuiller et al. 2009; Guéguen et al. 2026) to model distributions of 18 Adiantum species under current and future climates. Biomod2 carries out ensemble forecasting of species distributions, provides information about species responses, and tests the importance of the included predictors (Zhao et al. 2021; Huang et al. 2023; Cao et al. 2024). Biomod2 can also be parallelized to reduce computational time when modelling many species (Guéguen et al. 2026).

Our modelling flowchart is shown in Fig. 2. Although our field observations provided presence-absence data, the GBIF data are presence-only. Therefore, we generated the background occurrences (or pseudo-absences) that are needed in SDM for model training and testing by using the surface range envelope (SRE) method (Thuiller et al. 2009). This creates an environmental envelope around the species occurrences, and pseudo-absences are selected outside of the envelope (Barbet-Massin et al. 2012). For each species, we created six times as many pseudoabsences as there were presences for the species. Labelled data were created by extracting pixel values of the species-specific predictors for the occurrence and pseudo-absence points (latitude and longitude). 70% of the labelled data were used for training and 30% for testing.

We produced SDMs for each species using three different methods: Generalized Linear Models (GLM), Gradient Boosting Models (GBM), and Random Forest models (RF). All models were tuned with a default set of parameters using the “bigboss” strategy available in biomod2. A 3-fold cross-validation strategy (where all analyses were repeated three times) was used to assess model performance. Two accuracy metrics, True Skill Statistic (TSS) and Receiver-Operating Characteristic curve (ROC), were used to evaluate the models. TSS is a threshold-dependent accuracy metric that ranges from −1 to +1. The value +1 indicates perfect agreement between prediction and reference data; values > 0.8 are considered excellent, 0.4–0.8 moderate, and < 0.4 poor (Zhang et al. 2016). ROC is a threshold-independent metric that ranges from 0 to 1, with values > 0.9 considered good to excellent, 0.7–0.9 moderate, and < 0.7 poor (Imani wa Rusaati and Won Kang 2024).

To obtain a final distribution map for each species, we ensembled the results of the three models (GLM, GBM, and RF) using the Ensemble weighted mean (EMwmean) algorithm. This allows building ensemble SDM models by weighted averaging, which gives more weight to models with better performance. Only models with TSS ≥ 0.4 were used when building the Ensemble models (EM) (Zhang et al. 2015). We set the parameter “EMwmean.decay” in EMwmean algorithm to assign weights to each model in proportion to their TSS scores. The “var.import” parameter (the number of permutations when calculating the importance of each predictor) was set to a moderate value of 3 to achieve reliable values of predictors importance at minimum computation time (Guéguen et al. 2026).

Ensemble maps provide continuous data on predicted species occurrence probability. We converted these to binary distribution maps by using species-specific threshold values. These were chosen as the smallest occurrence probability that predicted presence in all the grid cells that had an actual occurrence point for the species in question. All grid cells with at least this probability value were coded as suitable for the species, and grid cells with lower probability were coded as unsuitable. This was done by applying the “bm_BinaryTransformation” function in biomod2 to the current and future scenarios.

To assess which predictors were most important for each species, we calculated relative importance scores for the predictors and produced response curves to illustrate how the probability of species occurrence varies in response to the value of the predictor. For each species, we modelled the current distributions and six different future scenarios. The differences between these were used to make inferences about species responses to climate change in terms of loss or gain of suitable habitat.

Each of the 18 Adiantum species had between 27–488 occurrence points in the dataset. The final species distribution models (SDMs) based on the ensemble weighted mean (EWM) modelling produced binary maps where the extent of suitable habitat varied between 0.12–4.45 million km² under current climatic conditions (Fig. 3, Suppl. material 1: table S3).

The six species with the smallest modelled ranges were restricted to either the Guianas (one species), the Andean forelands (four species), or the southern margin of the rain forest biome (one species; Fig. 3). The other 12 species had wide ranges in Amazonia. Three species had a doughnut-shaped (circum-Amazonian) distribution, as they were missing from the central parts of Amazonia but were found in most peripheral areas. Four species had a bimodal distribution as they were found both in the Andean countries and in the Guianan region but were missing from the central and southeastern parts of Amazonia. Finally, five species were found distributed across almost all Amazonia, although they were missing from the southeast (Fig. 3).

In general, Generalized Boosted Models (GBM) and Random Forest (RF) performed better than Generalized Linear Models (GLM), but all three produced at least moderately good models for all species (TSS 0.46–0.81, ROC 0.75–0.99). Ensembling improved performance further, and the final ensemble model (EM) had high performance for all species (TSS > 0.78 and ROC > 0.95, see Suppl. material 1: fig. S2). We used TSS > 0.4 (instead of the sometimes applied 0.6: Wang et al. 2016; Dogan and Yilmaz 2026) as a threshold for including a model in ensembling in order to produce a model for all species (Suppl. material 1: fig. S2). For species with fewer than 40 occurrences, the evaluation metrics (TSS and ROC) for the ensemble model showed very high values (~0.99), which may indicate overfitting.

Soil base cation concentration (SBCC) was the single most impactful predictor, as it was included in the models of all species and was the predictor with the highest importance for many of them (Fig. 4). Understory temperature (UT) and elevation were also included in all models, but their importance was clearly smaller. Cation exchange capacity (CEC) was important for half of the species, but its importance was generally smaller than that of base cation concentration. Probabilities of several of the soil types were also found to be important, especially those of Ferralsols, Podzols, and Acrisols.

Most species responded to either Annual mean temperature (bio1) or Annual mean precipitation (bio12), but the climatic variables that reflect diurnal or seasonal variation in temperature or rainfall emerged as more important in the species distribution models (Fig. 4). For many of the widespread species, the most important climatic variable was either Mean diurnal temperature range (bio2) or Temperature seasonality (bio4), with Precipitation seasonality (bio15) and Isothermality (bio3) providing additional predictive power. Net primary productivity (NPP) and Potential evapotranspiration (PET) were important for more than half of the Adiantum species but mean monthly near-surface relative humidity (HURS) provided better predictive power for some species.

The species with the most restricted range sizes (A. adiantoides, A. platyphyllum, A. poeppigianum, A. scalare, and A. macrocladum) and several of the intermediately widespread ones (A. argutum, A. petiolatum, and A. pulverulentum) showed an increasing probability of occurrence with increasing SBCC values, while the most widespread species were most likely to occur on cation-poor soils (A. tomentosum, A. terminatum, A. paraense, and A. cajennense) (Fig. 5a).

Species for which the mean diurnal temperature range (bio2) was important had generally higher probability of occurrence in areas with low variability; bio2 values > 10 were associated with clearly decreasing probability of occurrence for all of them except one (Fig. 5b). Even the highest predicted probabilities were relatively low for the species with very narrow distributional ranges and few occurrence points. Given that the vast majority of Amazonia is below 300 m elevation, it was to be expected that widespread species have peak probability of occurrence at low elevations. Notable exceptions in our analyses were the species mostly found close to the Andes, whose probability of occurrence increased with elevation, and A. tuomistoanum, whose probability of occurrence was very similar across the entire observed elevation range (Fig. 5c). Response curves for other predictors are shown in Suppl. material 1: figs S3–S7.

How species distributions were predicted to react to a changing climate varied widely among the Adiantum species. Under the more drastic climate scenario (SSP5-8.5), 12 of the 18 species were predicted to lose > 20% of their current range by 2070. At the same time, only three species would find a new suitable area amounting to > 20% of their current range. On balance, eight species would experience a net loss of > 50% of their range and one would experience a net gain of > 50%. Smaller net gains were predicted for three species and smaller net losses for six species (Fig. 6). Results for the less drastic climate scenario (SSP2-4.5) and for earlier years showed similar trends but more moderate numbers (Suppl. material 1: figs S8–S12). The two species with largest predicted gains but almost no losses of area were A. argutum and A. tetraphyllum, both of which have distribution areas that extend further south-southeast than those of other species do.

For many species, the southern and eastern parts of the distributional area are where the big losses of suitable area were predicted to happen. Otherwise, there does not seem to be any systematic relationship between current distributional area and the degree of expected change in suitable habitat area. For example, A. cajennense seems to be relatively resilient, unlike A. tomentosum, A. tuomistoanum, and A. terminatum, which have superficially similar distributional areas. Out of 18 species, 14 would face shrinkage of their habitat under extreme climatic conditions towards the end of the century (Fig. 7).

All SDM methods (GLM, GBM, and RF) reached good to excellent performance, and combining their results by optimal ensemble modelling (EM) increased performance further (ROC > 0.95 and TSS > 0.78). This agrees with results from previous studies from elsewhere in the world, which have found that ensembling reduces the uncertainty that is inherent in the individual models (Araújo and New 2007; Grenouillet et al. 2011; Montoya-Jiménez et al. 2022; Mahmoud et al. 2024). Using averaged climatic variables from five relevant global climatic models (GCMs), is also expected to reduce uncertainty in the predictions (Forester et al. 2013). To more rigorously assess the accuracy of forecasting from one time period to another, it has been recommended that temporally independent test data be used (Crimmins et al. 2013; Piirainen et al. 2023). This was unfortunately not possible in our case, because splitting the occurrence data into separate time slices would have exacerbated the spatial biases inherent in all Amazonian species occurrence data. The preliminary steps, such as variable importance, correlation, and category-wise selection of the top important predictors, can be performed to select the set of species-specific predictors. For example, in our study, we found that the microclimate (understory temperature) category has low importance for each species; therefore, it could be merged with the category of climatic predictors. Also, a sufficient set of best predictors should be used in the models, as the smaller number of predictors possibly reflects predictors artefacts and their distribution in the predicted habitat maps.

The species differed in which of the environmental variables were included in their SDMs. The variable that most consistently had a high predictive power, and was selected in all models, was soil base cation concentration (SBCC). This agrees with earlier studies that have emphasized the importance of SBCC for species distributions (Figueiredo et al. 2018; Flores et al. 2020; Zuquim et al. 2020) and community composition (Tuomisto et al. 2003a, 2003b, 2016, 2019). The SBCC response curves we obtained from the SDMs indicate similar SBCC niche positions as the published Huisman-Olff-Fresco (HOF) models for the same species (Tuomisto et al. 2024). The main difference is that niche widths in the HOF models were generally narrower, so a larger proportion of the species with intermediate niche position obtained unimodal response curves. This may reflect methodological differences: our models used raster-based rather than field-measured SBCC values, factored in several climatic and other variables in addition to SBCC, and used additional occurrence data from GBIF. Overall, the results support the notion that nutrient availability is such an important niche dimension for plants in Amazonia that its inclusion substantially improves species distribution models (Roe et al. 2022).

Other soil-related variables were selected in many of the SDMs as well, but none in all of them, and their general predictive power was lower than that of SBCC. It is noteworthy that cation exchange capacity (CEC) provided much poorer predictive power than SBCC did, and CEC was included in the SDMs of only half of the species. Unlike SBCC, CEC is available as a raster layer in SoilGrids (Hengl et al. 2017), so it is often used as a soil variable in SDMs and other studies. However, CEC does not quantify the availability of nutrient cations to plants, so its relevance to plants is limited and its use as a main soil predictor can lead to underestimating the importance of soil properties (Moulatlet et al. 2017). The same is true of the soil type probabilities of SoilGrids; only Ferralsols and Podzols entered the SDMs of at least half of the species, and a few other soil types (e.g., Acrisols and Luvisols) had relatively high predictive power for one or a few species. Clay content, silt content, and soil organic carbon content were also important for a few species only.

None of the climatic variables emerged as a universally superior predictor. Three of them were selected in more than half of the species-specific SDMs, and every model included at least one of these: mean diurnal temperature range (bio2), temperature seasonality (bio4) and precipitation seasonality (bio15). Total annual precipitation (bio12) was the most important climatic factor for only one species. These results suggest that it is the degree of seasonality, rather than average conditions, that are most critical. This is consistent with the general notion that ferns are sensitive to desiccation and reach highest abundance and species richness in conditions where scarcity of water is not a limiting factor (Kessler et al. 2011; Tuomisto et al. 2014; Qian et al. 2022). Even though trees are generally considered less drought-sensitive, their species richness in Amazonia has been found to be constrained by seasonal drought as well (Esquivel-Muelbert et al. 2017).

Taken together, our results support earlier notions that soils are at least as important as climate for the broad-scale distributional patterns of plants in Amazonia (ter Steege et al. 2006; Figueiredo et al. 2018; Tuomisto et al. 2019; Zuquim et al. 2020; Luize et al. 2024). Therefore, taking edaphic factors into account in SDMs is important to improve the accuracy of species distribution maps over this major biome.

The point maps based on species occurrences suggested that some Adiantum species are more widely distributed than others, and the raster maps produced by the species distribution models (SDMs) clarified differences in distribution patterns between the species. Overall, our SDM results agree with earlier studies that have found most Adiantum species to be adapted to relatively nutrient-rich soils (Lehtonen et al. 2015; Tuomisto et al. 2024). In particular, our results suggest that biogeographical distribution patterns within Amazonia reflect, to a large degree, the environmental niches of plant species and the spatial distribution of the habitats that are suitable for them. For example, the Andean element in the western Amazonian flora is not just defined by spatial proximity to the Andes, but also by the special environmental (especially edaphic) conditions prevalent in western Amazonia. Similarly, whether a species is widespread across Amazonia or absent from central Amazonia is linked to its ability to grow on nutrient-poor soils.

The six species with the smallest distributional areas fell into three categories: Guianan (A. adiantoides), Andean (A. macrophyllum, A. platyphyllum, A. scalare, A. poeppigianum), and southern (A. tetraphyllum). For all the Andean species, soil base cation concentration (SBCC) was the most important variable in the SDMs, and the probability of species occurrence was basically zero at SBCC < 1 cmol(+)/kg. This agrees both with the results of an earlier study that reported these species to have high optima for SBCC (Tuomisto et al. 2024) and with the overall geological setting of Amazonia: western soils are formed in relatively young sediments of Andean origin, which are less weathered and more nutrient-rich than soil parent materials in other parts of Amazonia (Hoorn et al. 2010; Quesada et al. 2011; Zuquim et al. 2023).

The Andean species had some GBIF occurrence points in lower montane forests, so their modelled distributions were truncated at our 600-m elevation limit. However, these species would still have small ranges in an Amazonian context even if the higher-elevation areas with occurrence points had been included in the modelling, because the surface area of the potentially suitable elevational belt is relatively small. The potential range of the single southern species might have been underestimated to a larger degree, because modelling was only done over closed-canopy forest. Extensive areas of southern Amazonia were excluded from analysis either because they predominantly represent drier biomes, or because they have already been deforested. For the single Guianan species, our analyses probably covered its entire distributional range.

A distinctive distributional pattern can be characterized as circum-Amazonian: the distributional areas of the species in this group (A. dolosum, A. glaucensens, and A. argutum) surround Amazonia south, east, and north, but are missing from the central-western parts. These species had intermediate optima and relatively narrow niches for SBCC (Tuomisto et al. 2024), but some of the climatic variables emerged as the most important predictors in the SDMs. These species, together with the single species with a southern distribution, seem to be better adapted to the relatively seasonal climates along the perimeter of the Amazonian rain forest biome than the other species in our dataset.

Another group of species (A. humile, A. pulverulentum, A. obliquum, and A. petiolatum) had a bimodal distribution as they were present in large areas both parallel to the Andes and in the Guianas but were missing from a broad swath of central and southern Amazonia. These had similar or slightly higher SBCC optima than the circum-Amazonian species (Tuomisto et al. 2024) but differed especially in that SBCC was more important and climate variables less important in their SDMs. Their distributions seem to avoid both the poor soils of central Amazonia and the seasonal climates of southern/southeastern Amazonia.

The final group of species had the largest distributional areas across Amazonia (A. paraense, A. tomentosum, A. tuomistoanum, A. cajennense, and A. terminatum). These have the lowest SBCC optima of all the studied Adiantum species (Tuomisto et al. 2024), and the commonness of infertile soils in Amazonia can be assumed to explain their broad distributions. SBCC was included in all their SDMs, but climatic variables obtained higher importance, and these species were modelled to be absent from the most seasonal southern parts of Amazonia. Given that each one of these distributional patterns can be linked to specific responses along environmental gradients, we propose that they are not specific to Adiantum ferns, but similar patterns will be found in Amazonian plants in general.

According to our models, climate change can be expected to affect the distributions of Amazonian Adiantum species quite dramatically already within the next 50 years. Under the more extreme business-as-usual climate scenario (SSP5-8.5), 14 out of the 18 modelled species would experience net loss of suitable habitat. For eight of them, habitat area would be reduced by more than 50%; five of these especially hard-hit species are currently among the most widespread and common in the entire genus (Tuomisto et al. 2024). This is contrary to the results of Miles et al. (2004), who found that widespread species were more resilient to climate change than narrowly distributed species.

Losing the common fern species from more than half of their current range would make a big impact on the community composition and species richness of the forest understory. Similarly worrisome results have been obtained for other plants both in Amazonia (Miles et al. 2004; Gomes et al. 2019; Zuquim et al. 2020; de Morais et al. 2024; Marques et al. 2024) and in other forested areas in the tropical Americas (Lima et al. 2022; Silva et al. 2024). For example, (Gomes et al. 2019) estimated that by 2050, more than 4000 tree species could become threatened, and the average tree species richness in Amazonia could decrease by 37% under the more pessimistic of the two climate scenarios they considered.

Ferns are generally considered efficient dispersers, but we do not know if they are able to disperse far enough fast enough to take advantage of a new suitable habitat at the same pace as current habitat becomes unsuitable. If not, the actual net loss of suitable habitat area will be even larger than appears from Fig. 7. Successful dispersal may be especially slow or unlikely for species whose suitable edaphic habitat occurs as isolated patches (Corlett and Tomlinson 2020). For more dispersal-limited plant groups than ferns, the problem will be even worse.

However, not all species react to a changing climate in the same way. The suitable habitat for one of our species (A. tetraphyllum) was forecast to increase by > 50%. This is the only species in our data whose main distributional area is at the southern border of Amazonia and extends well into the neighboring drier biomes. This means that the species can tolerate drier and more seasonal conditions, which was also evident in its positive response to increasing understory temperature in our SDM. Its predicted future expansion in southern Amazonia is no doubt due to the climate models predicting change towards drier and warmer conditions in this area. The other species with a sizable (~20%) net increase in suitable area (A. argutum) is one of the circum-Amazonian species, which is also adapted to seasonal climates and currently extends further south and southeast than the other Adiantum species.

Over the past decades, changes in land use and land cover (deforestation, logging, and fires) have had a prominent influence on Amazonian forests (dos Reis et al. 2021; Brandão et al. 2022; Lapola et al. 2023). The combined impacts of climate change and anthropogenic activities on habitat suitability and species diversity are far more severe than those of climate change alone. Especially the relatively drier southern and eastern parts of Amazonia suffer from deforestation and local climate change caused by it (Gatti et al. 2021; Flores et al. 2024). Our analyses did not explicitly consider the impact of future land-use change, fragmentation, degradation, fire, and dispersal barriers when modelling species distributions, but it is obvious that forest understory plants will be essentially wiped out from any deforested areas. When SDMs do not account for these factors, they produce incomplete, overestimated, and biased representations of how species are distributed. Therefore, including predictions of deforestation and fires in the SDM models would produce more realistic, but at the same time more worrying, future predictions (e.g., Gomes et al. 2019; de Morais et al. 2024).