Model-R / Back-end

Funções de modelagem de distribuição de espécies
1 stars 1 forks source link

R-model backend

This repository holds all the data and functions necessary to perform environmental niche models (ENM) using dismo and associated packages both in console and in the graphic interface Model-R.

The folder ENM has the following structure:

  1. At the root of ENM are the scripts where all the modeling procedures are executed and documented
  1. /data holds the species occurrences file, FLORA_occs_final.csv. The species occurrences com from the NeoTropTree project for plants. This is a final file that was already cleaned to adequate the original species names to the current correct names of the Flora of Brazil (16/02/2016). The file columns are the scientific name, latitude and longitude.
  1. /env has the predictor variables. These come from the 19 bioclimatic variables in Worldclim and the six topography variables in CGIAR. They have been transformed into a set of six "eigenvariables", a PCA was performed for the entire area and the first six components were spatialized, retaining 95% of the variation. The resolution of these predictors is 1km x 1km and the final extent is the BAF.
  1. /fct has all the functions needed to perform the ENMs.

    • 'modelos.R' defines functiondismo.mod, that allows to choose which predictors (as a raster::stack object) and which algoritms are used. The algorithm options are maxent, bioclim, domain, mahalanobis distance, glm, random forests and two svm implementations. It also allows to choose the number of partitions (default=3), and a seed for reproducibility. The output folder is where models will be stored. The projection to other datasets is partially implemented. The number of pseudoabsences can also be modified (default=500). dismo.mod produces models for every partition but does not join them.

    • Joining partitions is performed by the function final_model in final_model.R. Two ways of joining partitions are used:

      1. Models are selected according to the value of their maximum TSS value, then they are cut but the threshold that maximizes TSS. The final model is a mean of these binary models, cut again to match a majority consensus (at least 50% of the selected models predict each final area). This option is called Final.mean.bin7 due to historical reasons,
      2. Models are selected according to their maximum TSS value, then a mean is performed and this mean is cut by the mean of the thresholds that maximize each selected model. This option is called Final.bin.mean3.

Any post-processing operation may be performed manually, such as generating potential richness or checking the variation between models (by calculating the standard deviation between algorithms, for example).

More information about the file structure and operations performed can be found in the repository wiki.