Multi Tree Analysis

[EdwinP503]

Good evening guys!

There is a workflow I am currently working on that requires me to analyze a collection of images with multiple trees in a fixed background. I created a binary mask where I isolate the objects using the CMYK color space ('K' channel works the best to threshold the image and keep the targeted objects), but now I need to identify each individually and run shape analysis on them. A rectangular ROI along the fixed background would be ideal and roi_objects to select the trees with the 'partial' option. Looking at the seed-analysis and multi-plant workflows I'd say I would create a list of id for each object and loop through them to do morphology analysis for each of them.

My questions are:

1- Do I need to create a ROI grid? if so, how can I automate the number of columns based on the number of trees? or if I set the value to be 6 for example, the grid will be created as long as there are <6 objects?

2- I'm thinking of creating a conditional loop that continues to run as long as there are not more than 6 objects detected; which variable saves the number of objects in the ROI? There is a chance some branches are cut off and could be considered a separate object, is it possible to join those pieces to the main stem or can I apply skeletonize and prune to all the objects in the image at once?

3- There are cases where some branches from different trees overlap, is there a function from PlantCV or OpenCV used to tell what branches belong to the main stem?

4- The CMYK color space I believe was added recently, and it gives me the following warning when I run it for the first time in the kernel.
/Users/Edwin_Polanco/opt/anaconda3/lib/python3.8/site-packages/plantcv/plantcv/rgb2gray_cmyk.py:44: RuntimeWarning: invalid value encountered in true_divide

Could this affect the overall purpose of the workflow? The warning was constantly showing when running the workflow but I was able to avoid this to happen by ignoring the warning and reset it after the function, but if there is a different way I can improve it I welcome any suggestions.
ignore_divide = np.seterr(divide='ignore', invalid='ignore')
k_channel = pcv.rgb2gray_cmyk(rgb_img=balanced_img, channel='K')
reset_divide = np.seterr(divide='warn', invalid='warn')

The project is a work-in-progress, and any information is greatly appreciated!

Cheers,
Edwin Polanco

[nfahlgren]

Hi @EdwinP503!

1- Do I need to create a ROI grid? if so, how can I automate the number of columns based on the number of trees? or if I set the value to be 6 for example, the grid will be created as long as there are <6 objects?

We have a few approaches for this kind of problem (summarized in this tutorial: https://plantcv.readthedocs.io/en/stable/tutorials/multi-plant_tutorial/). pcv.roi.multi is probably the best option for this setup as long as the overall positions of the trees is relatively similar between images. You can definitely create six ROIs with the multi ROI tool and it is okay if there are < 6 trees. @jdavidpeery is working on a related method for automatically defining ROI grids in #911.

2- I'm thinking of creating a conditional loop that continues to run as long as there are not more than 6 objects detected; which variable saves the number of objects in the ROI? There is a chance some branches are cut off and could be considered a separate object, is it possible to join those pieces to the main stem or can I apply skeletonize and prune to all the objects in the image at once?

This is usually what I do also. I loop over the ROIs and within that loop run pcv.roi_objects per ROI. Then check if the output area is greater than 0, then run pcv.object_composition and other analysis steps.

We don't currently have a way to join disconnected stem segments unfortunately and those will cause an issue. pcv.roi_objects has the option to output only the largest object within the ROI. That would keep the most connected stems but would drop any disconnected pieces.

3- There are cases where some branches from different trees overlap, is there a function from PlantCV or OpenCV used to tell what branches belong to the main stem?

We also do not have a method for this yet. The most common case we have is when two secondary segments overlap and create a cycle with the primary stem. We have discussed some potential options for disambiguating the cross-over junction but have not implemented anything yet.

4- The CMYK color space I believe was added recently, and it gives me the following warning when I run it for the first time in the kernel.
/Users/Edwin_Polanco/opt/anaconda3/lib/python3.8/site-packages/plantcv/plantcv/rgb2gray_cmyk.py:44: RuntimeWarning: invalid value encountered in true_divide

I will have to check into this. It is likely caused by either a division by zero or a calculation with an infinity value. This should be mostly harmless (it is only a warning) and would only affect the outputs of isolated pixels (wherever a valid value of K cannot be calculated).

[EdwinP503]

Thank you @nfahlgren!

The position and spacing of the objects will vary throughout the collection of image and the number of objects as well. I was able to get the loop started to iterate through the length of the objects variable in pcv.find_objects without creating the ROI following the seed-analysis-workflow. Besides using pcv.analyze_object for each tree, I want to do a morphology analysis to get branch angles as well. But pcv.morphology.segment_insertion_angle takes quite some time to run, would it be a good idea to have both analysis in the same loop or would the system' requirements to run the workflow of 100+ images be more demanding?