**4. Discussion**

#### *4.1. Variable Selection and Model Precision*

The results of this study show that it is possible to establish a clear relationship between the probability of occurrence of both species and a reduced number of variables related to soil and climate parameters.

For *E. grandis*, the values of occurrence increase as a function of the percentage of clay content decrease and increase as the depth of the surface horizon increases. These soil texture and depth variables are closely related to eucalypts growth given their major influence on water availability [46,47], nutrients availability [48], and the volume of soil explored by the roots. The distribution of the eucalypt subgenus has been attributed to different responses to water availability. This, among other reasons, is due to the development capacity of a broad root system by modifying the ratio of root tissue to growth tissues and the depth of root penetration based on water availability [49]. For Uruguay, similar results have been reported [50], leading to the conclusion that the available water in the soil is one of the variables that better explain *E. grandis* growth. Likewise, [51] found a negative effect of the clay and silt presence on *E. globulus* growth. Isothermality, which indicates the temperature stability through the year [52], has also been reported as one of the variables explaining the growth of *Eucalyptus cloeziana* [53] and *E. grandis* [54]. The strong dependence of *E. grandis* on soil parameters was also verified by [55] evaluating plantations from 7 to 17 in the state of Paraná, southern Brazil.

The positive effect of north-west to north-east aspects on the probability of occurrence could be related to the greater exposure to solar radiation and the higher temperatures to which the planted trees in these orientations are exposed [56,57], although the positive effect of these variables arises from the interaction with variables such as soil moisture, temperature, or slope [58]. The relationship between orientation and radiation and the effects on the growth of eucalypts has been studied although according to Paton (1980) cited by [49] eucalypt in general are relatively less sensitive to the number of hours of light than to room temperature. However, [59] they argue that for the conditions of latitude such as those of Uruguay (30◦ to 33◦ S) in the north-northeast aspect there are greater hours of light which determines a greater period of photosynthesis as well as an occurrence of higher temperatures during the months of winter.

For *E. dunnii* the effects of thickness of the A horizon about the probability of presence can be explained by the same causes as those mentioned above. The positive effect of the temperature during April may be explained by the combined effect of a high relative water content in the soil (with greater precipitation than evapotranspiration) and intermediate temperatures that would promote tree growth. The increase in temperature during summer contributes to greater evapotranspiration, which may cause a deficit in the potentially available water in the soil during this period, although the precipitation has a relatively uniform distribution during the year [60]. This is consistent with the results obtained by [61] in the sense that the inclusion of climate variables with monthly records allows a greater predictive power of the species compared to the use of annual average values. According to these authors, this is due to the increasingly frequent occurrence of extreme weather events and therefore the importance of knowing the behavior of the species in such events.

In both species, higher values and less variation of the AUC were obtained with ensemble models compared to the use of individual models, which confirms the advantage of using the former to predict the occurrence of species [20,62,63]. Values of TSS above 0.85 (both with the assembled model and with the RF model) represent excellent predictive power [64]. These two models have shown very high capacities for predicting habitat in several species [37,65]. The work [66] obtained slightly lower values for this index (0.66 and 0.78) when analyzing the habitat prediction of *E. sideroxylon* and *E. albens*. According to the scale used by [67], several models in the group analyzed showed good to very good precision in habitat prediction, with the highest values corresponding to the RF model, for both species. This evaluation scale considers the following categories: <0.20 = poor, 0.21–0.40 = limited, 0.41–0.60 = moderate, 0.61–0.80 = good, and 0.81–1.00 = very good concordance. In general, greater values of

this index were obtained for *E. grandis*, with all the evaluated models. ANN also returned acceptable accurate predictions with good–very good values for K, TSS and AUC, though it also presents a large variability on the predictions which might be linked to the hidden relationships built within the model [28] which suggest that the results must be interpreted with caution. Surprisingly, MaxEnt gave a moderate accurate model prediction when is one of the most use and accurate method to predict species distribution [68].
