Brazil nut journey under future climate change in Amazon

seven different distribution modeling algorithms, which are: Artificial Neural Networks (ANN) [34]; Classification Tree Analysis (CTA) [35]; Function Discriminant Analysis (FDA) [34]; Generalized Additive Models (GAM) [36]; Generalized Boosted Regression Models (GBM) [37]; Generalized Linear Models (GLM) [38]; Random Forest (RF) [39]. For each modeling method, we ran 10 replicate models with 75–25% training-test propor- tion samples, respectively, re-sampled randomly by bootstrap. The 500 pseudo-absence sam- ples were selected using the ’disk’ function, considering the limit of a radius of 1˚ (111.11 km) in the surroundings of each geographical occurrence for selection. To evaluate the quality of the generated models, we used two metrics: True Skill Statistics (TSS) [40] and Receiving Operating Characteristic (ROC) [41] (S1 Table). For each model gen- erated, a measure of the importance of climatic variables for the construction of the models was also computed (S2 Table). Finally, for ensemble composition, we selected only models with TSS values equal to or greater than 0.6. Our consensual distribution model, represented by the average model among the best rotated models, was then designed under present and future climatic conditions for the different scenarios and periods. The result comprises a mea- sure of climate adequacy, ranging from inadequate (0) to the most appropriate (1000) climati- cally between the different periods and climate scenarios evaluated. Calculation of trend of gain or loss of climatic suitability

To assess the trend of changes in climate suitability for Bertholletia excelsa by the end of the century, we used the VoCC package [42] in R [33]. We use the ’tempTrend’ function that esti- mates coefficients of a linear regression and projects them over the geographic space, showing loss (negative values) or gain (positive values) of suitability for the species over the periods and climatic scenarios evaluated. Definition of the critical threshold of suitability for extinction The distribution models have in their representation in a continuous form of suitability for the species. Here, ranging between zero (0) and 1000. One of the key points of the post-modeling steps is to define from which value along the gradient the species will be present (persistence or potential colonization) or absent (local extinction). As a critical threshold parameter for the extinction of the species in a pixel, we used the minimum training presence (MTP) observed among all records of occurrence of the species, which is represented by the value equal to 550. In other words, if a pixel, in any period or future scenario assessed, has a value of less than 550 in terms of climatic suitability, it is declared that the species would be locally extinct. This strat- egy is widely adopted to define critical thresholds because it has a strong biological support [43]. Simulation of future spatial dynamics with MigClim Unlike static models, which are used to evaluate the effects of step-by-step environmental changes, here we use the MigClim package [44] in R [33]. The main function of this package is

to incorporate spatially explicitly the dispersion (movement) of the species over the study area as a function of events of changes in suitability for the time interval evaluated. In this sense, it was possible to quantify three fundamental results about the geographical distribution of a spe- cies under pressure of environmental changes (here climatic): [1] Persistence—when a particu- lar pixel maintains, throughout the evaluated period, minimally satisfactory values of suitability that support the maintenance of the species; [2] Extinction—occurs when an evalu- ated pixel loses suitability for a critical climate threshold; & [3] Colonization—occurs when there is an increase in suitability, because of ongoing changes, in unoccupied pixels next to PLOS ONE Effects of climate change on Brazil nut PLOS ONE | https://doi.org/10.1371/journal.pone.0312308 November 13, 2024 4 / 13