Tools for tree ring identification

[sandyie]

1. WinDENDRO

• https://regentinstruments.com/assets/images_windendro/WinDENDRO_2023.pdf
• System produced by Regent Instrument （canadian company)
• An optical scanner system
• Can be offered as a complete system (hardward scanner + software) or software along (3 different versions)
• Since we already have our own device, we only need to purchase the software in basic or regular tier
• Main difference between basic and regular tier are:
• Photo editing (in ring validation)
• Ring-width graphic, cross-dating, skeleton plotting, detrending (all the graphic analysis function are available for regular or higher ver)
• analyse tree-rings from wood disks, cores, X-ray films and images produced by filmless digital X-ray systems (support jpeg format)
• 3 steps: take image -> ring detection -> ring validation

PRO:

• specifically designed for dendrometrists
• Measure parameters:
• Annual tree-ring widths
• Minimum, maximum, average density
• Earlywood density
• System is updated regularly (mostly annually) and it has been updated since 1991 (mature system)
• Ready-to-used system
• Cover special cases such as narrow ring, ring contrast, skip cracks, photo editing … etc
• Contain cross-dating and density analysis functions
• Image can be edited manually (add parameters, select early vs. late wood, adjust boundary, skip gaps… etc) and the process is relatively easy according to a tutorial video of the software

CON:

• ring validation is still semi-automatic, user need to browse the image and adjust misidentified rings. When modification are done by user, the ring numbers and years are automatically updated
• No free-trial version available online, some reviews states that the system is unintuitive. Some training may be required to use the system
• No pricing available online
• A review from Carlson School of Management states that the system “costs upwards of $8,000”
• Can contact Regent for pricing
• The system introduction contain details about pipeline of the software, we can try to recreate it with lower cost (https://regentinstruments.com/assets/images_windendro/WinDENDRO_2023.pdf)

2. CDendro and CooRecorder

• https://www.cybis.se/forfun/dendro/
• https://www.cybis.se/forfun/dendro/helpcoorecorder7/index.htm
• programs used for e.g. dendrochronology, climate and forestry research all over the world
• Price : CDendro + CooRecorder TOGETHER: 68 US$ per installation (excl any VAT if applicable) ; school licence: 1360 US$ for 20 individual licenses
• CDendro: crossdate wood samples, Tree-Ring Measurement, Data Visualization, Statistical Analysis
• CooRecorder: measure ring widths with standard office scanner

PRO:

• Pricing is relatively fair
• Contain tutorials, forum, and wiki about how to use
• More resources online and screenshot of the software look more intuitive
• CooRecorder allows the user to ensure that rings are measured orthogonally, a feature that WinDendro lacks
• do not register tree ring widths, only save point coordinates which are then process; allow for the return to an old registration and further adjusting it

CON:

• Less information about how each step are implemented
• CooRecorder also seem to be semi-automatic ( require user to set certain parameters and make manual adjustment)
• No free trial version
• Anyone using our system is then required to purchased their own CDendro and CooRecorder since charges is per installation (computer)

3. Image Processing Software + Deep learning

• https://besjournals.onlinelibrary.wiley.com/doi/epdf/10.1111/2041-210X.14183
• A paper about using CNN for tree-ring detection and measurements
• CNNs have been widely used across many recognition tasks such facial recognition

PRO:

• The application for tree-ring measurements is freely accessible asa Docker container at quay.io/repository/treeringgenomics/image-processing and the image preprocessing code are also available in github
• All the steps are documented in the original paper, so we can make our own adjustments if needed
• Can customize to automate more steps AND include special cases for certains tree species if needed
• The “software” is free and other users using our system are not required to purchased any additional program, but we need to create our own tutorials for them

CON / ISSUE:

• The model proposed by the paper is trained on a datasets of 1705 images with 4092 annotated rings, where 179 were unseen validation images with 419 rings
• The entire datasets are not linked in the paper, we may not have access to the datasets and need our owns set of data
• CNN models can be large, and require a server to run the model (maybe UBC Sockeye or Digital Research Alliance of Canada; having a large model on AWS may cause high charges since it require lots of computation)
• Need to make adjustments of the model and the image processing pipeline since the application did not work perfectly on the our sample image
• Also need to do testings by ourselves

Sandy’s note:

• The paper for the deep learning method is relatively straightforward to read and understand
• Since the code for the deep learning method is available in a container, we can try exploring it more; initial exploration yield relatively satisfying result
• Since we implemented the photo stitching codes already, I can further expand the pipeline to include photo stitching + photo pre-processing
• If the approaches does not work or require too much time and effort, then we can explore the commercial products (since there are no free trial)?

[lizzieinvancouver]

@sandyie OMG -- you read my mind on JUST one of these (see meeting notes here) ... I think you are way ahead on this. Nice work!

I suggest you purchase ONE license for Coorecorder for now (on the corporate card) and try it out versus the deep learning approach and report back. One query on the third option ... would the space on midge possibly be enough (64 threads)? Or not?

[lizzieinvancouver]

@sandyie also reported to me by email:

I also explored some mobile apps that claim to be able to count tree rings but work poorly so I didn't include them as an option.

And asked:

Can I have more information about the certain species of trees that our software must support? so that I can try to customize the model and test with those species. Please let me know if you have any questions or other ideas about which tools to use for the identification.

For tree species, we need Abies amabalis, Doug fir, western red cedar and one other at Mount Rainier (please get @wangxm-forest to confirm). We also have Alnus rubra, Betula and some Quercus from Boston, check with @DeirdreLoughnan

[DeirdreLoughnan]

@sandyie thanks for your work on this, it is interesting to see the tradeoffs between existing methods.

Lizzie's list of the Boston species is correct, but more specifically we have: Alnus rubra, Betula papyrifera, Betula populifolia, Quercus alba and Quercus rubra.

[wangxm-forest]

@sandyie Thank you for working on this. The tree species at Mt. Rainier are: Abies amabilis (Pacific silver fir), Abies lasiocarpa (Subalpine fir), Pseudotsuga menziesii (Douglas fir), Tsuga heterophylla (Western hemlock), Tsuga mertensiana (Mountain hemlock), Thuja plicata (Western redcedar), Callitropsis nootkatensis (Yellow cedar)

We probably won't be able to core all these species though, but it we could include all these species, that would be great.