Is the Random Trees Classifier in ArcMap 10.5 the same as what is named Random Forest in other applications/software? I see that the parameters are the same and the design seems the same, but I need a verification of which algorithm is behind the classifier before I can implement it in my research.

If this is the case, how do I acquire the metadata from each classification task run in ArcMap? I am specifically thinking of the out-of-bag (OOB) error and the variable importance (VI) estimates.

  • 1
    I would ask this question on GeoNet the ESRI forum, you are likely to get a response from ESRI staff, rather than here a generic forum for all GIS systems and related issues. – Hornbydd Aug 11 '18 at 16:41
  • I would also echo @Hornbydd on this. One of the major issues with closed source algorithms is that only the maintainers of the code or the official documentation can answer your question. Your best bet is to start within the ESRI forums. – Trevor J. Smith Aug 11 '18 at 18:40
  • Okay, thanks a lot for your advice! I have posted the question on GeoNet. – Martin M. Aug 12 '18 at 18:24

They look similar, but the developers can better explain it. However, you can get out-of-bag (OOB) error, and the variable importance (VI) estimates by using the "Forest Based Classification and Regression" Tool. Also, you might find the following link interesting:



Based on this quote: "Creates models and generates predictions using an adaptation of Leo Breiman's random forest algorithm" from the ESRI function documentation I would not assume that the algorythm is equivalent to the the original published method.

Unfortunately, the documentation does not state what type of "adaptation" is implemented but, in my experience ESRI does all kinds of hinky things. At first blush it seems like there are all sorts of things going on behind the scenes, as far as how the spatial data is digested into the model. In addition, it looks like they are using Strobl's Conditional variable importance parameter selection method which, I have found is not appropriate in ecological models or remote sensing classification.

What is not clear is if they have also implemented Strobl's proposed alternative node splitting approach. Where Strobl's work is completely valid, it was focused on genetic array data and their simulations were structured around nominal (binomial) responses and not continuous data. As such, their proposed node splitting algorythm and conditional variable importance measures are predicated on a combinatorics data structure of nominal indicator (x) variables. One must also keep in mind that, from any inferential or result(s) based model evaluation criteria for a dependent variable (y), their end result was the parameter selection itself, indicating gene expression, and not any type of prediction of a response variable. In relation to the application of Random Forests models in other disciplines (particularly ecology and remote sensing), this research is often taken way out of context and applied in very unintended ways.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.