How to Separate Evergreens from Deciduous Trees in Deep Learning Output

Question

I ran the Tree Detection deep learning package against our 2022 leaf-off aerial photography to produce an output of ~116,000 tree boxes within my study area. I calculated area, mean greenness, mean infrared, and mean LiDAR intensity for each tree crown box using zonal statistics as table. After making some selections of evergreen tree boxes and inspecting histograms, scatterplots and box plots, I was able to come up with a query to select evergreens that works fairly well:

Area <=1000 AND Intensity >=40 AND MeanIR >=84

Note: green from the 3-band imagery was not useful due to a large number of nearly white deciduous trees having high green values, so IR from 2021 imagery was used.

This was an iterative trial and error process and I was wondering what else I could do to help me arrive at the magic numbers. I can probably train a new model using a set of carefully selected evergreen tree boxes, but for now I'm looking for suggestions on this process.

The green boxes in the attached video show all trees, while the yellow boxes shown later are the evergreens resulting from the query.

ITPACTreeDetect.mp4

AllTreeBoxes.png

GBacon · Answer

Thanks. The leaf-off imagery showed better separation between the evergreens and deciduous trees, especially with IR. I still use NAIP for general land cover though. Though interesting, I was looking for something higher res than NASS.

GIS-Cambria · Answer

What if you also used NAIP imagery? I think that is typically leaf on. There may also be a companion classified dataset for that.

USGS NASS is 30 meter pixels, probably too large to help but....The NASS is classified for evergreen, deciduous, and mixed forest.

GBacon · Answer

Thanks! I find it very interesting too. Love to hear what you find.