I am using Random Forest to predict different forest attributes (e.g. biomass, density, age class, etc.) using remote sensing and climate data. We want to predict species composition in each site, expressed as % for 4 groups of species (4 continuous responses). These percentages must sum to 100% in each site. I initially trained 4 different regression random forests independently for each species group. Some models have higher prediction accuracy (lower RMSE) than others, which is to be expected as some species are easier to identify with the sensors. The problem is that combining predictions from the 4 RF models yields a total predicted percentage higher than 100%. I just realized it's possible to use a multivariate random forest to predict all 4 responses simultaneously. My questions are:
- Will multivariate RF constrain the 4 predicted responses in a site to sum to 100%? If not, is it possible to impose such a constraint?
- If not, would it make sense to simply scale the predictions back to 100% myself (prediction x 100/ total %)?
- For the species groups that have a good predictive performance in the univariate models, how will the multivariate predictions compare? Will they be as good or can they be impaired by estimating them alongside harder to predict species groups?