Improving the directed segmentation would be the most straightforward and beneficial way to improve performance.
The indications are clear that the random search used for directed segmentation is suboptimal. The most obvious of these is the huge difference in accuracy between TClass when producing individual classifiers and when producing a voting ensemble. The fact that repeated executions of random segmentation improve accuracy dramatically indicate that the random segmentation algorithm could choose synthetic events better.
Here we make several suggestions for improving directed segmentation.