We are in the process of building and evaluating OLS and spatial error models, and we look forward to comparing their results and performance to random forests given that each model has its own properties and weaknesses.
Performance Results
Due to the independent nature of the majority of our parallelism, little communication was required (outside of the reading and writing of large files to disk) and, consequently, our scaling efficiency was relatively high. Figure \ref{258787} below shows the scaling of the raster cropping and resampling, the DEM transformations, and the random forest training steps. All reported scaling is strong in the sense that we did not vary the overall problem size but instead were only varying the number of processors assigned to the problem.