High quality specimen digitization is becoming standard across the sciences, is relevant for curation of natural history collections, and must become a priority for dendrochronology. This paper overviews the enduring role of imaging in dendrochronology, summarizes the potential relevance of gigapixel macro photography of polished specimens, offers a long-term review of a commercial imaging system, and reports our progress imaging entire collections of specimens at ultra-high resolution. Our gigapixel images of polished specimens have proven effective for digital analyses, archiving, and education, and we believe macro photography may prove a lower cost and more broadly accessible digitization alternative to microtomy and X-rays. We advocate for gigapixel macro photography as one accessible and adaptable paradigm to elevate reflected light imaging standards in dendrochronology.
High-resolution imaging for digital collections development and curation is becoming standard across the sciences (e.g.Baird 2010; Tegelberg et al. 2014; Short et al. 2018). This should be a front-line priority for dendrochronology, which hinges on histological scrutiny of xylem structures that encode individual- and population-scale records of cambium growth dynamics and environmental history (Schweingruber 1996). Assiduous recognition of individual cellular complexities is a fundamental prerequisite to precise crossdating of growth increments, particularly for collections from extreme environments. Measurement of cellular scale geometries can produce data signals of finer seasonal resolution and greater environmental system covariance than those from the traditional parameters of ring width, density, and reflected light intensity (e.g.Fonti et al. 2010; von Arx and Carrer 2014; Björklund et al. 2019; Edwards et al. 2020; Björklund et al. 2021). This emergent body of work underscores the scientific potential for high-resolution imaging to quantify shape characteristics for entire fields of cellular structures in woody plant growth rings.
Since the early years of North American dendrochronology, photography has been pivotal to credibility and communication of scientific results (e.g.Douglass 1919). At the first American dendrochronology conference, images were presented as primary evidence for accuracy of master chronology dating sequences (Glock 1937). A remarkable mosaic of photomicrograph prints was compiled to illustrate the key dating signatures of the continuous millennial-length chronology from the Flagstaff area (Douglass 1947). Photography of tree rings has been included for data visualization in myriad publications since.
In the latter half of the 20th Century, X-ray imaging transformed paleotemperature dendroclimatology (e.g.Polge 1970; Schweingruber and Briffa 1996). By the 1990s, video cameras improved increment width measurement systems with microscopes and stage micrometers (e.g.Munro et al. 1996; Sheppard et al. 1996). In our opinion, the most exciting specimen analysis progress of the last century includes quantitative wood anatomy (QWA) approaches that leverage meticulous microtomy specimen preparation (e.g.Spiecker et al. 2000; Fonti et al. 2010; von Arx and Carrer 2014; Edwards et al. 2020; Björklund et al. 2021), or alternatively, variant techniques with X-ray imaging (e.g.Evans 1994; Downes et al. 2002), including X-ray computed tomography (Van den Bulcke et al. 2009; De Mil et al. 2016). However, few institutions are equipped for success with these advanced and resource-intensive analytical approaches.
Our understanding is that most labs still rely primarily on reflected light with polished specimens, and that high-volume data development remains heavily dependent on traditional stage micrometry or digital measurement with flatbed scanner images using software such as WinDENDRO (https://regentinstruments.com) and CooRecorder (http://www.cybis.se/forfun/dendro/). Unfortunately, the Epson XL scanners so widely used for digitization are fundamentally inadequate for resolving the cellular structures that are critical to accurate crossdating (Figure 1) and are of increasing environmental interest (e.g.Edmondson 2010; De Micco et al. 2016; Edwards et al. 2020; Therrell et al. 2020). Considering the unique natural history collections curated by our community, and the fundamental importance of image resolution (e.g.Björklund et al. 2019), it is clear that the tree-ring community standards for reflected light imaging are due for an upgrade.
Our group has explored one approach by deploying off-the-shelf macro photography equipment and computational techniques to compile microscope quality images with comprehensive subject coverage, sharp focus, and pixel counts in the range of one billion or more (i.e. gigapixel macro photography; “GMP”). Here, we summarize the mechanics of GMP, review a commercially-available imaging system, and evaluate the relevance and prospects of this approach in relation to other operational paradigms for developing secondary data from tree rings.
GIGAPIXEL MACRO PHOTOGRAPHY
Macro photography is the practice of imaging relatively large subjects (e.g. tree-ring specimens) at reproduction ratios that are life size or larger, accomplished using interchangeable lens camera systems, dedicated macro lenses, and bright light sources. Advances with computational techniques have changed what is possible with digital macro photography. A single ultra-high resolution image can now be distilled from tens to thousands of overlapping exposures captured with fine, precise, and orthogonal movement of the camera system in three dimensions relative to the subject of interest (Figure 2). Robotic automation of camera system movement is ideal for ensuring consistency in high-quality results with computational post processing. Digital post processing that preserves geometric proportions without distortion and measurement scale calibration ensures feasibility of precise and accurate measurements with individual and mosaic images.
Multi-focus fusion techniques (Kaur and Kaur 2016), colloquially described as “focus stacking”, are used to create composite images with depth of field greater than possible with an individual exposure. The composite is derived from a stack of exposures obtained as the camera moves sequentially in the z-dimension, altering the position of the focal plane relative to the subject surface (Figure 2). This process is well suited to subjects with subtle or dramatic variations in surface height, and ultimately circumvents the miniscule depth of field limitation inherent to the physics of optical light transmission at high magnification. Focus stacking algorithms compile a composite image from those pixels in the stack with the sharpest focus, as determined through complex analyses of pixel matrices. We have had good success with the commercial software Zerene Stacker (https://zerenesystems.com).
Mosaic stitching concatenates sequential exposures that overlap in the x-y plane (i.e. across the subject surface (Figure 2)). Resulting images have proportions that are undistorted, rectilinear, and suitable for calibrated measurement across the x-y plane. Because most known panoramic landscape photography stitching softwares apply geometric projections that create distortion with significant scaling implications, we developed an automation routine to leverage open-source computational techniques for distortion-free mosaics from sequences of focus-stacked images. The scale-invariant feature transform algorithm (Lowe 1999) finds key points in the overlap zone of adjacent images. The Fast Library for Approximate Nearest Neighbor algorithm set (Muja and Lowe 2014) is used to compute the alignment matrix between image pairs, and mosaic compilation of the final image follows. The general mechanics of GMP described here transcend any particular hardware and software solution, and apply to both reflected light and transmitted light imaging. Analogous image capture and post processing workflows are used for focus stacking and stitching images with commercial microscopy systems, although usually for much smaller subjects than full-length increment cores and cross sections.
THE GIGAMACRO SYSTEM: A LONG-TERM REVIEW
Following brief experimentation with ad hoc, custom-fabricated systems for manual and automated camera movement, we were fortunate to purchase a GIGAmacro Magnify2 Robotic Imaging System™ (Four Chambers Studio LLC, Napa, CA, USA; Figure 2). GigaMacro (GM) operation begins with physical configuration of the lens, lighting, and subject placement, followed by exposure and stitching of a few test images to verify satisfactory outcomes of resolution and subject illumination. The user then builds an imaging job queue, specifying the three-dimensional camera movement extents for individual subjects. The GM handles automation of camera system movement and exposure, and data file transfer to local disk storage. Our post-processing routine automates multi-focus fusion and mosaic stitching as described above, plus data backup to online data storage. With our current setup, an efficient technician can complete imaging and post-processing for a batch of twenty cores with two or fewer hours of hands-on user time, four to eight hours for image acquisition, and roughly twelve hours for automated post processing. These rough estimates are hardware dependent and sensitive to physical dimensions of the subjects being imaged. The person-time requirements for this approach are less than for digitization with a flatbed scanner.
Close adherence to an operations protocol has facilitated replicable and high-quality results between users. For 5.15-mm increment cores, we often use the 3:1 macro lens reproduction ratio with a 36×24mm “full-frame” imaging sensor oriented in the landscape position, resulting in images that cover a 12×8mm area of the subject. This allows us to image the entire surface of a full-length 5.15mm increment core with a single row of images along the x-axis (Figure 2), accounting for subtle bends common to many core mounts. For the 5:1 reproduction ratio, full-frame images cover a 7.2×4.8mm area of the subject, so we orient the 5.15-mm core along the y-axis to capture the entire specimen surface in a single column of exposures. Final image resolution and effective resolving power will depend on camera sensors and optical lens specifications. To provide an example, our 22 megapixel full-frame camera paired with the 3:1 (5:1) reproduction ratio produces a final resolution of 468p/mm (780p/mm), or 11,887 dots per inch (19,812 dpi). These correspond to pixel resolutions of 2.14 and 1.28 microns, respectively. For a 500mm long 5.15mm diameter core specimen, first planed with the core microtome (Gärtner and Nievergelt 2010) and then polished with microfinishing abrasives, a typical 3:1 (5:1) imaging job would require around 12 (25) images in the z-stack, one row (column) of 60 (159) exposures. Considering the 30% overlap required between frames, that results in a total of 720 (3975) exposures to produce a final mosaic with 0.89 (2.18) gigapixels. Automation of camera movement and post processing is essential.
The stacking and stitching workflow described above is straightforward for an individual row or an individual column of exposures, which works well for imaging core specimens and radii of cross-sections typically analyzed for time-series measurement. For larger subjects, creating distortion-free, rectilinear mosaics from multiple columns and multiple rows (i.e. for a complete tree-ring cross-section), requires 50–66% exposure overlap in the x-y dimensions, and photogrammetry software such as Agisoft Metashape (https://www.agisoft.com).
System Strengths and Limitations
The GM is capable of producing incredible mosaic photographs of entire tree-ring specimens (Figure 1). Image quality, flexibility with subject types and sizes, automation, and the modular nature of the GM system are its greatest strengths. The system works with a wide range of off-the-shelf camera bodies and lenses that produce image files in public domain formats free of proprietary restriction (e.g. RAW, TIFF, JPEG). Users can select and swap camera bodies and lenses to suit budget and resolution needs, and to track significant technology advances while avoiding system obsolescence.
Durability of the GM system components has been impressive, requiring only basic maintenance and end-of-life replacement. Our camera shutter failed after ten times its expected maintenance lifespan of 150,000 actuations. One lens aperture eventually required service, as it actuates with each exposure. We degraded one set of macro flashes and two ring flashes by not allowing a sufficient rest time between individual exposures. Three years after purchase, the GM has proven an unequivocally reliable workhorse for automated GMP.
We expect some operational limitations to be improved in future platform iterations. Currently, the greatest inefficiency of the GM software is that individual jobs require a consistent number of z-dimension images across all x-y positions, so large subjects with uneven surface heights require 10–30+ images for each z-stack instead of the key three to seven exposures typically useful for accurate focus stacking at an individual x-y position. Resolving this issue would dramatically reduce exposure counts, imaging and post processing time, and the data file storage volumes that ensue. Neither of our variable magnification macro lenses transmit electronic metadata on exposure reproduction ratio, nor do they include mechanisms to lock into specific magnification levels between the minimum and maximum. We work around these issues by imaging a calibration scale with each batch, and using gaffers tape to inhibit lens creep.
Lighting remains a pernicious challenge. Macro photography requires very bright light and it can be tricky to orient the two macro flashes for consistent and even illumination across an image frame. Ring flashes should alleviate this issue, though the ones we have tried rapidly overheated with high-speed use. Careful macro flash positioning and testing prior to each imaging batch can mitigate the issue, but a post-processing routine that automates brightness and color calibration and exposure adjustment gradients should be developed. The lighting issue injects uncertainty about optimal adaptation of GMP mechanics for developing continuous time series of reflected light intensity (Sheppard et al. 1996; McCarroll et al. 2002; Sheppard and Singavarapu 2006), though there are lingering questions regarding the interpretation of blue light as a robust and precise proxy for dendroclimatic inference (e.g.Wilson et al. 2014; Björklund et al. 2021).
Gigapixel macro photography of polished surface specimens delivers resolution, focal depth of field, dynamic range, and subject interpretability that are vastly superior to what is possible with flatbed scanners (Figure 1; full resolution images online: https://doi.org/10.13020/rep3-nf86). We have entirely migrated our data development workflow to GMP for imaging, digital measurement of ring width, and some QWA. With millions of individual exposures, we have imaged thousands of tree-ring specimens from dozens of site-level collections. Compared to secondary data such as ring width, density, or cell-wall thickness, the digital facsimile data we are creating are more akin to the primary data of our science, i.e. the physical wood specimens. A collections-scale archive of high-quality GMP images should be well suited to evolving analytical paradigms and future research applications. We believe macro photography may prove a lower cost and more broadly accessible digitization alternative to microtomy and X-rays.
Our image archive has proven useful for remote research collaboration, education, and outreach, both before and during the COVID-19 pandemic. For example, we used GMP images to instruct the introductory group at the World-Dendro 2018 fieldweek in Bhutan, and we have supported small and large institutions, including the University of Arizona Laboratory of Tree-Ring Research, to facilitate authentic online learning experiences with dendrochronology.
Beyond the clear benefits of a high-quality image archive, automated GMP delivers greater time efficiency to quality assured measurement data, particularly for lab groups that depend on relatively inexperienced students for research assistance. Unlike traditional stage micrometry, points placed digitally in an image software can be rapidly revisited for modification during the crossdating and measurement quality control processes. Compared to lower resolution flatbed scans, the GMP images routinely resolve microscopic details of cellular-scale anatomy that are crucial to rigorous crossdating and robust scientific interpretation (Figure 1). This alleviates the need to return to the microscope and physical specimens for re-evaluation of anatomy not sufficiently resolved in flatbed scanning images (Maxwell et al. 2011 ).
Near-flawless preparation of specimen surfaces is paramount for two dimensional imaging with micron scale pixel resolution. Various methods have been championed for high-quality outcomes, including the “v-cut” (Douglass 1941), an ultra-precise fly cutter (Spiecker et al. 2000), and thin section microtomy (e.g.Gärtner et al. 2015). Contemporary QWA analyses with two-dimensional images require binary image segmentation, whereby pixels with cellular structures of interest are differentiated from all other pixels. Effective segmentation has historically required high contrast images of intact and undistorted cellular structures with relatively homogeneous pixel brightness values. These can be quite challenging to produce for many wood types, even for expert microtomy technicians, and especially for those less experienced. Our understanding is that the vast majority of small and large research laboratories have high-volume workflows that rely primarily on sanding and finish polishing of wood specimen surfaces. We use aluminum oxide abrasive papers and films, with final grain sizes in the 1–10 micron range. This approach facilitates impressive surfaces and images that allow subjective evaluation of microscopic anatomical details needed for crossdating and measurement. However, these carefully polished wood surfaces do contain slight degradation or distortion of cell walls, irregular filling of void spaces with wood dust, and heterogenous pixel values for the cellular structures of interest. This is illustrated clearly for the tracheid lumen areas and earlywood vessel areas in Figure 1.
So are reflected light GMP images of polished wood surfaces suitable for QWA? Probably not with the current generation software. Preliminary testing with ROXAS (von Arx and Carrer 2014) and ImageJ (Jevšenak and Levanič 2014) indicates that segmentation with these images was not comprehensively accurate and precise at mapping the cellular structures of interest. Alternate surface preparation techniques with additives can preserve integrity of cellular structures and improve segmentation outcomes (e.g.Spiecker et al. 2000; Fonti et al. 2009; Gaertner et al. 2015). Segmentation errors can also be corrected through manual correction, though this would be much more feasible for ring-porous angiosperms with relatively few earlywood vessels than for gymnosperms with tens to hundreds of thousands of tracheids. On the other hand, emerging technologies with advanced computational image algorithms are encouraging (e.g.O'Meara 2021). We are cautiously optimistic that in the next few years, efforts with computer vision techniques, artificial intelligence, robust training datasets, and massive levels of specimen replication may facilitate effective segmentation with reflected light images of polished surface specimens. Pending that outcome, our reflected light image archives would be ready for the next generation of QWA tools.
Our gigapixel mosaic images are near or beyond the ceiling of fluid handling capability for most current generation desktop computers and the tree-ring image softwares publicly available as of April 2021, including WinDENDRO, CooRecorder, and ROXAS. These high quality softwares have proven incredible resources for the research community. Pending interest in the GMP approach, incompatibility could be resolved as computing technologies mature and as developers resolve the pixel count limitations common to most softwares. Meanwhile, these issues can be resolved through operational cropping of image subsets into more easily digestible units. Though expanded discussion is beyond the scope of this paper, our group is developing open-source software tools to facilitate rapid online access to ultra-high resolution images, with browser-based tools for micrometry, and potentially, next generation QWA.
We view GMP as a low-cost and high-volume workflow complement to the transformative techniques in QWA requiring meticulous microtomy (Gärtner et al. 2015) or X-ray computed tomography (Van den Bulcke et al. 2019), which are time-intensive and have thus far been relegated to higher-cost and lower volume workflows at well-resourced institutions. Although the U.S. $70,000 purchase cost of the GM system may be impractical for small institutions and lab groups in the developing world, less expensive robotics alternatives are feasible and camera equipment technology and availability will continue to improve. We aspire to increase community access to GMP and are continuing to explore lower-cost options for automated image acquisition systems. The modular nature of the GMP mechanics, equipment, and data stream should facilitate continued experimentation with emerging imaging technologies. This should be critical to avoiding dead ends with obsolescence inherent to many proprietary imaging alternatives. As cloud computing becomes more widely embraced, well-resourced facilities should serve as centralized hubs for data development that can be shared at very little cost to internet connected devices.
To conclude, ultra-high resolution macro photography has great potential to advance digital collections development, curation, and analysis in dendrochronology. Tree-ring metadata standards have improved (Jansma et al. 2010; Brewer et al. 2011), but the specimen digitization standards across much of the community have lagged far behind. Considering the enduring role of dendrochronology in the Earth and environmental sciences (e.g.Babst et al. 2017), its potential relevance for science-informed societal decision-making (e.g.Rice et al. 2009), and its accessibility as an educational gateway for environmental systems thinking (e.g.Davi et al. 2019), it is critical that we move towards standards of open science. High-quality imaging is a logical step, and GMP is one flexible approach worthy of consideration.
This work was conducted through collaboration with the UMN Advanced Imaging Service for Objects and Spaces, and was made possible by support from the UMN Office of the Vice Provost for Research, the UMN Office of Undergraduate Research, the UMN College of Liberal Arts, the UMN Liberal Arts Technologies & Innovation Services, NSF DEB Award #1655144, and NSF AGS Awards #1602633 and #1903504. We are grateful to Bethany Coulthard, Jesper Björklund, Bryan Black, Kristen Griffin, Julie Edwards, Kurt Kipfmueller, Paul Krusic, David Stahle, and Georg von Arx for conversations that informed this work. The manuscript benefited substantively from insightful feedback from two anonymous reviewers and guidance from Associate Editor Gretel Boswijk.