Segment anything on crowns and then black out bad pixels in the bounding box.

[bw4sz]

At test time we can use segment anything but on deepforest box.

[bw4sz]

Can we also look at an 1) a three band representation, 2) HSI embedding pixel classier to assist the flow algorithm.

[bw4sz]

Since we are already creating bounding boxes, we could also try to create a mask RCNN type network that uses species aware masks. We could generate thousands of weak annotations. Both from the point data we have, as well as from predictions. I don't know if the LiDAR could be added as well.