How do random forests work?

First, several small samples of the entire data set are taken. Any observations that are not taken are called “out-of-bag” samples.

Bootstrapped samples from random forests.

Random forests methods work well for response variables and models of various kinds. For example, the problem could involve classification (e.g., whether a tree is alive or dead) or regression (e.g., the height of a tree). After the bootstrapped samples are drawn, a set of regression trees are built.

Then, the mean square error (MSE) is calculated for each regression tree using the out-of-bag samples for that tree. After all of those calculations, the MSE is averaged across all regression trees.

What is the key output from random forests?

Many studies in forestry contain dozens, if not hundreds, of potential predictor variables. Each variable in an RF model will contain an importance score. The importance score is calculated by evaluating the regression tree with and without that variable. When evaluating the regression tree, the MSE will go up, down, or stay the same.

If the percent increase in MSE after removing the variable is large, it indicates an important variable. If the percent increase in MSE after removing the variable is small, it’s less important.

In a forestry example, Deo et al. 2017 used random forests to evaluate hundreds of remote sensing variables from a Landsat time-series and lidar. The authors identified several remote sensing variables that led to estimates of change in forest aboveground biomass:

Variables are ranked from most to least important in random forests output.

What are the strengths and limitations of random forests?

The strengths of random forests include:

• They can handle all types of data (e.g., categorical, continuous).
• They often perform similarly or better than regression methods.
• They work well with data sets containing a large number of predictor variables.

The limitations of random forests include:

• They are often considered a “black box” method because models are not easily interpreted.
• They require extensive model tuning for the best results.

Conclusion

Random forests are powerful and versatile with forestry data because they only use a random subset from a large data set. These methods have increased in popularity in recent years and can be applied to a variety of classification and regression problems.