currently I'm doing predictive modeling of soft-bottom-communities (more than 2.500 point features) with at least 15 explanatory rasters (for example bathymetry and it's derivates) within the German EEZ of the North Sea. I ensured the equal configurations (amount of trees, no. of variables per tree, split ratio of train and test data) for the forest-model in R trained with the package randomForest and the tool forest-based classification and regression in ArcGIS Pro.
As the nature of the algorithm, the results always vary a little bit, but they are quite reproducible. As described in the tool of ArcGIS Pro, 2/3 of the test data is used per tree to train, in R the default value for bootstrapping is quite the same (sampsize = 63.25%, source = https://uc-r.github.io/random_forests).
But by comparing the results of ArcGIS Pro and R there is a quite high discrepancy between each other considering the predicted rasters and the importance of the explanatory raster. Therefore I conclude the following question: Is it possible, that the algorithm as respectively implemented in R and ArcGIS Pro vary's and produce different results? Or did I miss to set up any further parameter to equal the results more?
For demonstration I show you the respectively predicted rasters in ArcGIS Pro and R.
I looking forward to your help.
