A personal story of trees and machines
[Thanks to more help from a backer, the campaign has been extended until December 16th at midnight PST. Would you please continue to help spread the word and make contributions? Now, a backstory that also explains some aspects of the project. ]
Such time consuming steps force scientists to conduct their studies with fewer data points; thus the true potential of the dataset is never achieved. Automatic detection and measurement of stomata have the potential to solve this problem. (Jayakody et al. 2021 p. 2)
When I was a small child, I overheard a scientist in a documentary describe flows, hubs, and robustness in networks, the organizational pattern of all living systems. All of my hair stood on end. I felt chills and knew that science was what I wanted. (Maybe this was or wasn't prescient—I am also susceptible to autonomic sensory meridian response (ASMR).) Around the same time, a 4 cm tall drawing and a short paragraph about a coast redwood in an illustrated botany book enchanted me. I was growing up in a suburban neighborhood next to a capped landfill amid chaparral in Southern California; I tended a garden the size of a blanket at the foot of a palm tree and hardly travelled. Less than fifteen years later, in 2011, a fellow UC Berkeley undergraduate, Robert Stevenson, recruited me onto the redwood research team from within a class we both took: Mathematical Modeling for Biological Systems, the class where I learned my first programming language, MATLAB, and non-linear dynamics (aka chaos theory), the mathematical underpinning of systems and network theory.
When I told Robert about my previous undergraduate research with fig trees and my search for an honors thesis project, he replied that he knew exactly what I should do. He and I went to Cindy Looy's sunny, plant-filled lab; he and Cindy explained the redwood project; and I signed up with them nearly immediately.
I participated for the next two years in that Looy-Dawson collaboration. This included a Summer Undergraduate Research Fellowship, my honors coursework, a stint as a professional lab technician in the Looy Lab after I graduated in 2012, a trip to New Orleans to present results at the 2013 Botanical Society of America (BSA) conference, and a few weeks volunteering as a field assistant. In the field, I helped Looy Lab then-doctoral student Dori Contreras dig for Sequoioid fossils in New Mexico, and, since I met their prerequisite of experience with rope-based tree climbing techniques, climbed for one of the Dawson Lab’s studies about old-growth coastal redwoods. I got to collect branchlets from the base and top of the canopy pre-dawn and at noon to help assess how the trees in Big Basin were responding to that region’s acute drought. Robert and I both presented at the 2013 BSA conference. He focused on Paleozoic winged conifer seeds and I compiled our collaborative team’s work on coastal and giant redwood stomata. My original thesis only included S. sempervirens. All of this was a vast privilege and pleasure.
However, the techniques available for studying stomata required nearly notorious amounts of repetitive labor. Cindy designed the project to involve a community and regularly checked in with everyone to make sure they were coping OK with the monotony. I always told her I felt fine, but at the age of 21 when I considered a Ph.D. where it seemed I would endlessly peel the skins off of minuscule leaves and count how many stomata, pores, I found on them, I experienced a bodily reaction. As important as stomata really are, and no matter how much I loved the trees and our community, I just did not feel ready.
However, microscope image analysis requires scientists to count and measure a large number of stomata in order to find statistically significant pat terns, and this proves to be time consuming and cumbersome work if done manually. (Ibid p. 2)
Over the next decade, I got many of my wiggles out with projects like coordinating the U.S. Forest Service's first Urban Ranger Station and co-founding a gardening company inspired by systems theory. I eventually rediscovered my enjoyment of programming during the methods class of a master's program called Social Science: Environment and Community. Meanwhile, no one finished publishing the stomata paper(s) the Looy-Dawson team had started and machines took over most of what was repetitive. Cindy and I started talking in 2020 about finishing the project.
Three stages became mostly or fully automated. Cindy helped obtain a specialized microscope for the department to share: it has a robotic arm for generating microscopy panoramas, no easy job. Cuticles have textured depth to them and microscopes a limited range of focus, so photographers need to pull the microscope's focus and take multiple images at each frame to get everything crisply visible: no blurry stomata. Next, improved Adobe Photoshop software can now handle stitching these large, three-dimensional sets of pictures, eliminating yet another task we used to do by hand. Finally, computers can now count stomata. What might have taken teams years can get done in a fraction of the time. Stomata studies might begin to yield powerful information about plant life in the past and present.
The opening and closing of stomatal pores directly impact both CO2 intake and water transpiration rate of a plant [4–6]. Hence, plant scientists study stomata behaviour to learn more about plant water stress as well as surrounding environmental changes [7–9]. In addition to studying living plants, scientists also use plant fossil cuticles to uncover climate change patterns by analysing stomatal density, size and behaviour [10–12]. (Ibid p. 1)
Our team has had back and forths about if we publish a first paper with of the manual technique data followed by a machine-assisted one, or merge the project stages. How much longer will small data sets for stomatal studies remain acceptable or relevant in scientific journals? The automation field is still developing and occasionally not how we would wish. Stomata counting software that was released in 2020 now runs on already legacy—outdated—languages and libraries (Tensorflow 1.5.1 and Keras 2.1.5). Little programming "bugs" can crop up. The existing software doesn’t yet do everything we would want for this project, such as measure the pore’s angle relative to the leaf axis. Distributed teams around the world are collaborating. It's all open source. I'm a novice in the field.
I chose to start with Python, a multipurpose language, when I wanted to expand my programming skills earlier this year. I let a preference for design draw me into full-stack software engineering from there but still pined for data analysis. When the scope of this redwood project became more clear I added some machine learning to my coursework. I have strong math skills so I felt prepared: machine learning utilizes linear algebra and calculus. Some skills transfer from software engineering—things like Git, Command Line, and patience for debugging. For this redwoods project, I plan to at least be able to speak intelligibly to humans who talk to stomata-counting computers if not to the computers themselves. Again, distributed teams around the world are collaborating on open-source software and I'm learning.
When I reviewed the redwood literature to catch up and help make sure that this study still contributes something new and useful, I read through a decade of work that I might have helped write if I had pursued that Ph.D., including the finding that coastal and giant redwoods sequester more carbon than any other forest types. That dive felt surreal, but I don’t regret my decisions. I have loved my jobs—a life full of trees in many contexts—and feel delighted to pursue this research today, including working with machines.
I hope that you support the study and keep learning along with us. The funding campaign will continue until midnight PST on December 16th. I'm grateful to the backer and Experiment teams for the time extension.
Reprise:
I admit to my struggles so far with crowd-fund campaigning, a different kind of work with trees and machines. Right now, this project has about five hundred page views. By Experiment's calculations in the researcher guide, projects of this scale need nearly three thousand to be successfully funded. Up until this year, I focused my digital networks to close friends and local communities. I sometimes avoided social media entirely. For example, my gardening company successfully ran on word-of-mouth recommendations alone and I commuted to some sites with a wheelbarrow. I’m more accustomed to using computers to network with plants and their ecosystems. While I've appreciated connecting with a beautiful, larger online human world for this campaign—including a cool dendrochronologist community on Twitter—I've been shy and slow. I've also been busy. In addition to this research, I study software engineering, facilitate a systems theory learning group with Capra Course, and tutor k-12 students in language arts and STEM. (Everyone on this team has been working a lot! We might have all tried coasting on charismatic megaflora during this campaign...)
Even with the struggles, this blend of work and communities feels meaningful to me, including in the context of climate collapse. Thank you for being here, together. Now, according to Experiment's researcher guide, 80% of projects that succeed in raising 20% of their funding target ultimately get funded. With the time extension, this campaign is in OK shape at 25%. I'm hoping for the best and really do hope you help us out.
With deep respect,
Meriel
Reference:
Jayakody, Hiranya, et al. “A Generalised Approach for High-Throughput Instance Segmentation of Stomata in Microscope Images.” Plant Methods, vol. 17, no. 1, Dec. 2021, p. 27. DOI.org (Crossref), https://doi.org/10.1186/s13007-021-00727-4.
0 comments