The scenProc feature to detect holes from imagery often returns polygons with holes, especially in area with dense forests and some clearings in the forest. When the ProcessHoles option is set to true the holes will be processed for the output of the detect feature step as well.

But I have to give a warning, for big and complex polygons the output is sometimes corrupt. I will try to fix that issue, but for now you can improve the output by first applying a split grid step to slice your imagery into chunks. When the detect features step is run on those chunks the output is better in general. See an example configuration file below:

IMPORTGDAL|..\geo\imagery\gtif_wgs84_4band\321778.tif
 #
 SPLITGRID|0.001
 #
 DETECTFEATURES|FTYPE=RASTER|filter_ndvi_test.tfc|OBJ|FOREST
 #
 MERGEGRID
 #
 EXPORTSHP|OBJ=FOREST|trees_test

As you can see I used quite a small grid (even smaller than the autogen grid). In the end I use a merge grid step to combine all the cells again and export them as one shapefile. That shapefile I can then use in my further processing to autogen.

An added benefit of the SPLITGRID step is that the performance of the feature detection gets better. When I ran it on the image as a whole it took around 4000 seconds, but now with the small chunks it takes only 300 seconds. So for better performance using split grid is also a good advice.