Improve tree extraction method

[المشرف العام]

I want to extract trees from a given normalized Digital Surface Model (nDSM) in SAGA GIS. My workflow is the following:

1. Applying a Gaussian Filter to the nDSM with search radius with standard deviation=1, search mode=circle and search radius=5m
2. Reclassifying grid values to exclude vegetation below 5m
3. Applying an inverse watershed segementation algorithm (searching maxima; not joining the elements)

Somehow, my result yields too many trees a some site, too few at others. Some "real" trees even show up two seed points.

Here's a comparison my found trees to the orthophoto (dark green=trees or pastures. light green=tree shadows/bare):



Besides a little shift due to georeference inaccuracies: Any ideas how to improve the workflow? I already tried to vary the parameters (resampling the 1m resoluton to 10cm, gaussian radius 2-5m, reclassifying threshold=15m (as to extract tree crowns only), joining the segments based on seed to seed difference...)



أكثر...