--- title: "Variability in results" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Variability in results} %\VignetteEngine{knitr::knitr} %\VignetteEncoding{UTF-8} --- <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.0.0/css/all.min.css" /> ### <i class="fa-solid fa-address-card"></i> Definition When modelling species distribution, there is most of the time an **uncertainty** of what is actually modelled : - the *fundamental niche* is most of the time unknown - and there might be an uncertainty in the *realized niche*, except in the case of an exhaustive sampling. What can be evaluated is how much the modeling results are **consistent** with what is known and observed, and how much **variability** is present in the results in function of modelling choices. </br> ### <i class="fa-solid fa-chart-column"></i> Variability - within the evaluation / importance of variables Single and ensemble models can be evaluated by several available evaluation metrics, and the importance of variables can be calculated through several repetitions (see `metric.eval` and `var.import` parameters in [`BIOMOD_Modeling`](../reference/BIOMOD_Modeling.html) and [`BIOMOD_EnsembleModeling`](../reference/BIOMOD_EnsembleModeling.html)). Variability in evaluation and importance values can come from the parametrization of different elements of the modelling : - **observed dataset**, + through the number of repetitions of calibration / validation splitting, + *illustrating the robustness of the computed model based on the data* - **pseudo-absence dataset**, + through the number of repetitions of PA sampling, + *to check the choice of the pseudo-absence strategy* - **modelling technique**, + through the models selected, among 10 models available, + *to spot the most adapted modelling methods* </br> ### <i class="fa-solid fa-map"></i> Variability - within the predictions Making projections, either for single or ensemble models, can produce two additional sources of variability in results that can be explored through two parameters : - **For single models**, `build.clamping.mask` in [`BIOMOD_Projection` function](../reference/BIOMOD_Projection.html) : + it produces a map of the studied area with, in each pixel, the number of variables whose value is outside the range of values used to calibrate the models. + *It identifies potential extrapolation areas.* - **For ensemble models**, `prob.cv` in [`BIOMOD_EnsembleForecasting` function](../reference/BIOMOD_EnsembleForecasting.html) : + it calculates the coefficient of variation between models' projection, + *identifying areas of disagreement in predictions*.