# How to associate the probability output of a Random Forest classifier with reality

I am currently working with the national forest inventory of Mexico. A sample sceheme that collected data concerning the state of forests on points spread out all over the country. One of these variables is a cualitative one concerning ground damage/degradation due to many causes. E.g. pasture activities. This variable also has associated an ordinal degree of "damage" if there is any, 1: 1%-20%, 2: 20%-40% ... (I dont recall the size of the intervals exactly). As a quick excercise I took this pasture variable and ignored the degree of damage. I just made a binary variable that indicates damage is present / not present. Then I trained a Random Forest classifier using many remote sensing, topographic and climatological variables. These variables are available wall to wall in the country so I can later use the model to generate a map. If I train the classifier using a balanced sample between 1's and 0's and do a hard-classification I get a precision with CV of around 75%. I'm more interested in generating a probability of degradation map so I didnt balance the sample and generated a soft-classification. This would be the probability of class 1 (degradation is present):

Does this excercise make sense to anyone? My true concern is the interpretation. How would one associate these probabilities with what is really going on in the country? Or does anyone have an example of when one would want to associate a soft-classification with a real life phenomenon? To me it makes much more sense to think fuzzily. But I am lost when it comes to describing what is going on.

• Did you try to validate your classification? Do you have samples of data which allow you to compare results with your map? For example, databases providing a threshold value involving number of cattle produced per hectare (or acre). If cattle/ha > 1 = 0 (not degraded), if not = 1. Your way of thinking makes sense to me, but you need to validate your model somehow. Apr 10, 2013 at 1:17
• here Spatio-temporal avalanche forecasting with support vector machines they try to predict the probability of an avalanche, and they plot it on a map Apr 10, 2013 at 4:57
• These are nice but brief comments. I really appreciate them but I am looking for a little more detail. Any ideas? Apr 11, 2013 at 15:33