Building on a modified version of rpart which can be found here, this package implements advanced functionalities for random forests (Breiman, L. Random Forests. Machine Learning 45, 5–32, 2001) which make this technique suitable for statistical downscaling of precipitation, as analyzed in Legasa et al. 2022: A Posteriori Random Forests for Stochastic Downscaling of Precipitation by Predicting Probability Distributions, published in Water Resources Research. The key elements of RandomForestDist are:
-
The inclussion of several split functions intended for predictand variables that are non-normally distributed. In Legasa et al. 2022 , we focus on the two-parameter gamma distribution (Deviation and Log Likelihood). However, other distributions can be easily added through the modified rpart package.
-
A new approach called "a posteriori" which has proven to accurately capture the whole probability distribution of the predictand Y given the predictors X, allowing thus for the generation of reliable stochastic predictions.
Please refer to the notebook included in the package for examples of use.
For RandomForestDist to work, the modified version of rpart needs to be installed first:
devtools::install_github("MNLR/rpart")
Afterwards, install the additional dependencies from CRAN and the package itself:
install.packages(c("progressr", "qmap", "fitdistrplus"))
devtools::install_github("MNLR/RandomForestDist")
Note that, in case devtools is not already available, it can be installed from CRAN using the command install.packages("devtools").