Importance of the predictors of a random forest model

Question

I want to investigate how a vegetation type (4 categories) is explained by several environmental factors (height above sea level, disturbance magnitude, disturbance frequency, etc) - maybe some predictors could be excluded, maybe one predictor is the most relevant? Therefore I identified the vegetation type in the field in 160 locations (points, random sampling) and gathered the info of the environmental factors.

My idea is to build a random forest model (in R with the randomForest package) based on 70% of the points and then test the prediction accuracy against the other 30%. The importance of the predictors I want to evaluate by building different models (excluding some predictors) and then compare the accuracy of the model.

Would that be a appropriate way to do such an analyses?

Is it necessary to do the 70/30 split (i guess my model would get better if i use all points for training)?

In the next step I want to use the model to make a prediction for an area (cell wise) where I have raster maps of all environmental variables.

Answer 1

R is already doing something similar, and you can extract what it deems the most important using the varimp function. However, in many practical situations I've found it is of somewhat limited value. For instance, I can often eliminate the most important variable with little loss in overall performance, probably because other variables can step in and fill the gap. But you're overall premise is correct - if your performance suffers whenever variable X is removed, it was probably important.