Hi Lisa,

Cool project! "I've noticed nearly no difference between a scale of 125, versus a scall of 5,000,000" --> Please contact support, I think if you provide them with the data and the Rule Set they will quickly be able to help you out. In rough theory, yes a larger Scale Parameter should result in larger objects, but the MRS results depends on different parameters like bit depth of input data, # of input layers, complexity etc..

Seeing your project, the first idea that came to my mind is that you try to classify the larger trees (intense green) and if you have those you could compute a relational feature which for example counts the number of these trees in a buffered area of x meters. This would result in a value distribution that looks like a heat map. This feature than could be used to create your desired object (your last screenshot). This is at least an approach that I have used in the past to create such outlines. You also could compute the distance from those trees and use this to create such an outline. There are as always numerous ways to get to your goal.

Hope this helps?

Cheers,

Matthias

Matthias! Wow! Let me first say thank you for all your quality training content. I have watched many, many of your videos, many, many times! I truly appreciate you sharing your knowledge. I have applied much of what I've learned from you and it has enabled me to be successful in my geobia endeavors. 

I will give your idea a go and contact support. I have several potential layers to work with (solar radiation, topographic wetness index, DHM). So far I've been using multi resolution segmentation with the 4 bands of NAIP and DHM equally weighted. I'll think about expending that. Or maybe experimenting with weighting. 

I found this definition in the docs "Relational features are used to compare a particular feature of one object to those of related objects of a specific class within a specified distance" Can this concept be used on objects before they are classified? Maybe it could work if I just select "unclassified". I spend some time experimenting.

Thanks very much!
Lisa

Thank you very much. 

Please also try a template matching approach, I assume the low standard deviation of the trees and their sparse distribution will also lead to success.

If you have elevation data, you can also first segment/classify all elevated objects and then use the spectral characteristics to identify the trees.

Cheets,

Christian

Thank you Christian. I watched the webinar on template matching where beach umbrellas are identified. Would you think for something like I'm doing, I would make it a much bigger pixel width and height, encompassing an area in each sample that represents a section of savanna with multiple trees in it?

(ツ) No ... Please create a template for the single trees, once you identify them using template matching, you can then aggregate the trees into larger image objects (there are several ways to do this, e.g. using a Grow&Shink approach). I believe once you have the single trees, you have solved the main challenge. 

I follow, thank you! This is similar to what I was trying with Matthias' approach, there I am identifying single trees with the surface model values. 

*Edit - which I realize now is your tree crown delineation tutorial. =)