Alexander E. White
Research Subjects
Research Interests
Alexander E. White has a total of 78 co-authors. Here are the top co-authors by number of shared publications:
- Rebecca B. Dikow
Data Science Lab, Office of the Chief Information Officer, Smithsonian Institution
7 shared publications - Paul B. Frandsen
Department of Plant and Wildlife Sciences and Bean Life Science Museum, Brigham Young University , Provo, UT 84602
6 shared publications - Michael Trizna
Smithsonian Institution Office of the Chief Information Officer
4 shared publications - Trevor D. Price
University of Chicago Biological Sciences Division 1
3 shared publications Alexander McKeeken
Entomology, Plant Pathology and Nematology University of Idaho Moscow ID USA
2 shared publications- Corey DiPietro2 shared publications
Latest Publications
- Open Access
- 24 April 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (4) , e0346139
Digitizing microscope slide-based natural history collections: A protocol using slide scanner technology
Natural history collections contain millions of microscope slides documenting global microscopic biodiversity, yet these materials remain largely undigitized and are vulnerable to deterioration and loss. Recent advances in slide scanner technology, originally developed for medical pathology, offer new opportunities for comprehensive digitization of slide-based collections. Here we present an optimized protocol for digitizing diverse microscope slide specimens, using the Hamamatsu NanoZoomer S20 slide scanner, developed while imaging slides at the Smithsonian National Museum of Natural History. We provide specimen-specific recommendations for scanning parameters, including scan area, focal points, Z-stack configuration, and file management workflows. Scanning times range from 41 seconds for small invertebrates to 18 minutes for palynological samples, with final compressed file sizes of 0.15-28 GB. High-resolution images (0.23 μm/pixel) captured diagnostic morphological features across all specimen types, including pollen, diatoms, radiolarians, plant and fungi tissues, and invertebrates. Using this method, we estimated that just the NMNH’s paleo-palynology slide collection contains approximately 4.3 billion individual specimens, 30 times more than the current estimated size of the entire NMNH collection. Slide scanning enables 3D data capture, facilitates remote collaboration, improves reproducibility of taxonomic identifications, and creates permanent digital records that mitigate risks of physical deterioration. This protocol provides practical guidance for institutions looking to digitize slide-based collections to preserve and unlock their full research potential.
- 24 April 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (4) , e0346139
Open AccessDigitizing microscope slide-based natural history collections: A protocol using slide scanner technology
Natural history collections contain millions of microscope slides documenting global microscopic biodiversity, yet these materials remain largely undigitized and are vulnerable to deterioration and loss. Recent advances in slide scanner technology, originally developed for medical pathology, offer new opportunities for comprehensive digitization of slide-based collections. Here we present an optimized protocol for digitizing diverse microscope slide specimens, using the Hamamatsu NanoZoomer S20 slide scanner, developed while imaging slides at the Smithsonian National Museum of Natural History. We provide specimen-specific recommendations for scanning parameters, including scan area, focal points, Z-stack configuration, and file management workflows. Scanning times range from 41 seconds for small invertebrates to 18 minutes for palynological samples, with final compressed file sizes of 0.15-28 GB. High-resolution images (0.23 μm/pixel) captured diagnostic morphological features across all specimen types, including pollen, diatoms, radiolarians, plant and fungi tissues, and invertebrates. Using this method, we estimated that just the NMNH’s paleo-palynology slide collection contains approximately 4.3 billion individual specimens, 30 times more than the current estimated size of the entire NMNH collection. Slide scanning enables 3D data capture, facilitates remote collaboration, improves reproducibility of taxonomic identifications, and creates permanent digital records that mitigate risks of physical deterioration. This protocol provides practical guidance for institutions looking to digitize slide-based collections to preserve and unlock their full research potential.
- 11 June 2025
- Published by Western North American Naturalist in Western North American Naturalist
- Vol. 85 (2) , 205-215
Designing Fog Collectors to Understand Impacts of Changing Precipitation Patterns on the California Channel Islands
La niebla es una fuente crucial de humedad para las plantas en entornos costeros, especialmente en las Islas del Canal de California, donde es un recurso hídrico significativo. Los ecosistemas de los bosques de pino bishop en la Isla de Santa Cruz sirven como un ejemplo destacado de la importancia ecológica de la niebla, donde el goteo de la niebla de las agujas de pino proporciona agua esencial para las especies del sotobosque. Para comprender mejor el impacto de la precipitación de niebla, diseñamos e implementamos dispositivos recolectores de niebla conocidos como “arpas de niebla”. Nuestro diseño se centró en crear un arpa de niebla precisa, duradera y fácilmente transportable, optimizada para su despliegue en el campo. A través de pruebas paralelas de arpas de niebla en la Reserva Rancho Marino Kenneth S. Norris, un apropiado proxy continental para entornos inundados de niebla en la Isla de Santa Cruz, evaluamos la capacidad de las arpas de niebla para capturar patrones de precipitación de niebla con el tiempo. Nuestras arpas de niebla demostraron consistencia en la recolección diaria de precipitación de niebla, resaltando la viabilidad y el potencial de nuestras arpas de niebla para contribuir a nuestra comprensión de los cambios climáticos en entornos costeros, especialmente en las Islas del Canal de California.
- 11 June 2025
- Published by Western North American Naturalist in Western North American Naturalist
- Vol. 85 (2) , 205-215
Designing Fog Collectors to Understand Impacts of Changing Precipitation Patterns on the California Channel Islands
La niebla es una fuente crucial de humedad para las plantas en entornos costeros, especialmente en las Islas del Canal de California, donde es un recurso hídrico significativo. Los ecosistemas de los bosques de pino bishop en la Isla de Santa Cruz sirven como un ejemplo destacado de la importancia ecológica de la niebla, donde el goteo de la niebla de las agujas de pino proporciona agua esencial para las especies del sotobosque. Para comprender mejor el impacto de la precipitación de niebla, diseñamos e implementamos dispositivos recolectores de niebla conocidos como “arpas de niebla”. Nuestro diseño se centró en crear un arpa de niebla precisa, duradera y fácilmente transportable, optimizada para su despliegue en el campo. A través de pruebas paralelas de arpas de niebla en la Reserva Rancho Marino Kenneth S. Norris, un apropiado proxy continental para entornos inundados de niebla en la Isla de Santa Cruz, evaluamos la capacidad de las arpas de niebla para capturar patrones de precipitación de niebla con el tiempo. Nuestras arpas de niebla demostraron consistencia en la recolección diaria de precipitación de niebla, resaltando la viabilidad y el potencial de nuestras arpas de niebla para contribuir a nuestra comprensión de los cambios climáticos en entornos costeros, especialmente en las Islas del Canal de California.
- Open Access
- 25 October 2023
- Published by Pensoft Publishers in Research Ideas and Outcomes
- Vol. 9, e113334
Developing responsible AI practices at the Smithsonian Institution
Applications of artificial intelligence (AI) and machine learning (ML) have become pervasive in our everyday lives. These applications range from the mundane (asking ChatGPT to write a thank you note) to high-end science (predicting future weather patterns in the face of climate change), but, because they rely on human-generated or mediated data, they also have the potential to perpetuate systemic oppression and racism. For museums and other cultural heritage institutions, there is great interest in automating the kinds of applications at which AI and ML can excel, for example, tasks in computer vision including image segmentation, object recognition (labelling or identifying objects in an image) and natural language processing (e.g. named-entity recognition, topic modelling, generation of word and sentence embeddings) in order to make digital collections and archives discoverable, searchable and appropriately tagged. A coalition of staff, Fellows and interns working in digital spaces at the Smithsonian Institution, who are either engaged with research using AI or ML tools or working closely with digital data in other ways, came together to discuss the promise and potential perils of applying AI and ML at scale and this work results from those conversations. Here, we present the process that has led to the development of an AI Values Statement and an implementation plan, including the release of datasets with accompanying documentation to enable these data to be used with improved context and reproducibility (dataset cards). We plan to continue releasing dataset cards and for AI and ML applications, model cards, in order to enable informed usage of Smithsonian data and research products.
- 25 October 2023
- Published by Pensoft Publishers in Research Ideas and Outcomes
- Vol. 9, e113334
Open AccessDeveloping responsible AI practices at the Smithsonian Institution
Applications of artificial intelligence (AI) and machine learning (ML) have become pervasive in our everyday lives. These applications range from the mundane (asking ChatGPT to write a thank you note) to high-end science (predicting future weather patterns in the face of climate change), but, because they rely on human-generated or mediated data, they also have the potential to perpetuate systemic oppression and racism. For museums and other cultural heritage institutions, there is great interest in automating the kinds of applications at which AI and ML can excel, for example, tasks in computer vision including image segmentation, object recognition (labelling or identifying objects in an image) and natural language processing (e.g. named-entity recognition, topic modelling, generation of word and sentence embeddings) in order to make digital collections and archives discoverable, searchable and appropriately tagged. A coalition of staff, Fellows and interns working in digital spaces at the Smithsonian Institution, who are either engaged with research using AI or ML tools or working closely with digital data in other ways, came together to discuss the promise and potential perils of applying AI and ML at scale and this work results from those conversations. Here, we present the process that has led to the development of an AI Values Statement and an implementation plan, including the release of datasets with accompanying documentation to enable these data to be used with improved context and reproducibility (dataset cards). We plan to continue releasing dataset cards and for AI and ML applications, model cards, in order to enable informed usage of Smithsonian data and research products.
- Open Access
- 27 September 2023
- Published by Pensoft Publishers in ARPHA Preprints
- Vol. 4, e113335
Developing responsible AI practices at the Smithsonian Institution
Applications of artificial intelligence (AI) and machine learning (ML) have become pervasive in our everyday lives. These applications range from the mundane (asking ChatGPT to write a thank you note) to high-end science (predicting future weather patterns in the face of climate change), but because they rely on human-generated or mediated data, they also have the potential to perpetuate systemic oppression and racism. For museums and other cultural heritage institutions, there is great interest in automating the kinds of applications at which AI and ML can excel, e.g., tasks in computer vision including image segmentation, object recognition (labeling or identifying objects in an image), and natural language processing (e.g. named-entity recognition, topic modeling, generation of word and sentence embeddings) in order to make digital collections and archives discoverable, searchable, and appropriately tagged. A coalition of staff, fellows, and interns working in digital spaces at the Smithsonian Institution, who are either engaged with research using AI or ML tools, or working closely with digital data in other ways, came together to discuss the promise and potential peril of applying AI and ML at scale and this work results from those conversations. Here we present the process that has led to the development of an AI Values Statement and an implementation plan, including the release of datasets with accompanying documentation to enable these data to be used with improved context and reproducibility (dataset cards). We plan to continue releasing dataset cards, and for AI and ML applications, model cards, in order to enable informed usage of Smithsonian data and research products.
- 27 September 2023
- Published by Pensoft Publishers in ARPHA Preprints
- Vol. 4, e113335
Open AccessDeveloping responsible AI practices at the Smithsonian Institution
Applications of artificial intelligence (AI) and machine learning (ML) have become pervasive in our everyday lives. These applications range from the mundane (asking ChatGPT to write a thank you note) to high-end science (predicting future weather patterns in the face of climate change), but because they rely on human-generated or mediated data, they also have the potential to perpetuate systemic oppression and racism. For museums and other cultural heritage institutions, there is great interest in automating the kinds of applications at which AI and ML can excel, e.g., tasks in computer vision including image segmentation, object recognition (labeling or identifying objects in an image), and natural language processing (e.g. named-entity recognition, topic modeling, generation of word and sentence embeddings) in order to make digital collections and archives discoverable, searchable, and appropriately tagged. A coalition of staff, fellows, and interns working in digital spaces at the Smithsonian Institution, who are either engaged with research using AI or ML tools, or working closely with digital data in other ways, came together to discuss the promise and potential peril of applying AI and ML at scale and this work results from those conversations. Here we present the process that has led to the development of an AI Values Statement and an implementation plan, including the release of datasets with accompanying documentation to enable these data to be used with improved context and reproducibility (dataset cards). We plan to continue releasing dataset cards, and for AI and ML applications, model cards, in order to enable informed usage of Smithsonian data and research products.
- Top Cited PapersOpen Access
- 13 June 2022
- Published by Wiley in Methods in Ecology and Evolution
- Vol. 13 (8) , 1640-1660
Deep learning as a tool for ecology and evolution
Deep learning is driving recent advances behind many everyday technologies, including speech and image recognition, natural language processing and autonomous driving. It is also gaining popularity in biology, where it has been used for automated species identification, environmental monitoring, ecological modelling, behavioural studies, DNA sequencing and population genetics and phylogenetics, among other applications. Deep learning relies on artificial neural networks for predictive modelling and excels at recognizing complex patterns. In this review we synthesize 818 studies using deep learning in the context of ecology and evolution to give a discipline‐wide perspective necessary to promote a rethinking of inference approaches in the field. We provide an introduction to machine learning and contrast it with mechanistic inference, followed by a gentle primer on deep learning. We review the applications of deep learning in ecology and evolution and discuss its limitations and efforts to overcome them. We also provide a practical primer for biologists interested in including deep learning in their toolkit and identify its possible future applications. We find that deep learning is being rapidly adopted in ecology and evolution, with 589 studies (64%) published since the beginning of 2019. Most use convolutional neural networks (496 studies) and supervised learning for image identification but also for tasks using molecular data, sounds, environmental data or video as input. More sophisticated uses of deep learning in biology are also beginning to appear. Operating within the machine learning paradigm, deep learning can be viewed as an alternative to mechanistic modelling. It has desirable properties of good performance and scaling with increasing complexity, while posing unique challenges such as sensitivity to bias in input data. We expect that rapid adoption of deep learning in ecology and evolution will continue, especially in automation of biodiversity monitoring and discovery and inference from genetic data. Increased use of unsupervised learning for discovery and visualization of clusters and gaps, simplification of multi‐step analysis pipelines, and integration of machine learning into graduate and postgraduate training are all likely in the near future.
- 13 June 2022
- Published by Wiley in Methods in Ecology and Evolution
- Vol. 13 (8) , 1640-1660
Top Cited PapersOpen AccessDeep learning as a tool for ecology and evolution
Deep learning is driving recent advances behind many everyday technologies, including speech and image recognition, natural language processing and autonomous driving. It is also gaining popularity in biology, where it has been used for automated species identification, environmental monitoring, ecological modelling, behavioural studies, DNA sequencing and population genetics and phylogenetics, among other applications. Deep learning relies on artificial neural networks for predictive modelling and excels at recognizing complex patterns. In this review we synthesize 818 studies using deep learning in the context of ecology and evolution to give a discipline‐wide perspective necessary to promote a rethinking of inference approaches in the field. We provide an introduction to machine learning and contrast it with mechanistic inference, followed by a gentle primer on deep learning. We review the applications of deep learning in ecology and evolution and discuss its limitations and efforts to overcome them. We also provide a practical primer for biologists interested in including deep learning in their toolkit and identify its possible future applications. We find that deep learning is being rapidly adopted in ecology and evolution, with 589 studies (64%) published since the beginning of 2019. Most use convolutional neural networks (496 studies) and supervised learning for image identification but also for tasks using molecular data, sounds, environmental data or video as input. More sophisticated uses of deep learning in biology are also beginning to appear. Operating within the machine learning paradigm, deep learning can be viewed as an alternative to mechanistic modelling. It has desirable properties of good performance and scaling with increasing complexity, while posing unique challenges such as sensitivity to bias in input data. We expect that rapid adoption of deep learning in ecology and evolution will continue, especially in automation of biodiversity monitoring and discovery and inference from genetic data. Increased use of unsupervised learning for discovery and visualization of clusters and gaps, simplification of multi‐step analysis pipelines, and integration of machine learning into graduate and postgraduate training are all likely in the near future.
- 20 November 2021
- Published by Wiley in Global Ecology and Biogeography
- Vol. 31 (2) , 280-293
Key roles for the freezing line and disturbance in driving the low plant species richness of temperate regions
Aim: At the macroscale, climate strongly correlates with species richness gradients, resulting from differences in in‐situ diversification and dispersal. One historical explanation for the pattern is that regions spanning temperate climates contain few species because past disturbances have generated high extinction rates, and species from tropical regions are unable to easily colonize temperate regions. We test these postulates for Himalayan plants, which span subtropical to temperate climates over steep elevational gradients.Location: Himalaya.Time period: Present day.Major taxa studied: Angiosperms.Methods: We use a comprehensive survey of 31 floras to document the elevational and geographical distributions of native Himalayan plants, augmented by field studies of trees in both the east and west Himalaya. We use grade of membership models to cluster species according to locations shared and phylogenetic analysis to evaluate diversification rates.Results: Species fall into four cohesive biotas, organized by climate. Points of turnover between biotas occur where the mean minimum temperature of the coldest month is approximately 0 °C (2,000–2,500 m), and at the point of occasional annual freezing (1,000–1,500 m); these boundaries run the length of the Himalaya. The patterns are retained when we consider whole clades rather than species. All plants (and the subsets trees, herbs and shrubs) belonging to the biota above the 2,000–2,500 m line have higher recent speciation rates than those lower down.Main conclusions: We attribute the high rate of recent speciation in temperate climates to high rates of turnover, creating ecological and geographical opportunity. The high elevation biota has few species, but spans the largest area, implying species numbers are far from any carrying capacity, at least with respect to accumulation of allopatric forms. This study thus links climatic restrictions of clades to differences in diversification rates, and by inference species numbers.
- 20 November 2021
- Published by Wiley in Global Ecology and Biogeography
- Vol. 31 (2) , 280-293
Key roles for the freezing line and disturbance in driving the low plant species richness of temperate regions
Aim: At the macroscale, climate strongly correlates with species richness gradients, resulting from differences in in‐situ diversification and dispersal. One historical explanation for the pattern is that regions spanning temperate climates contain few species because past disturbances have generated high extinction rates, and species from tropical regions are unable to easily colonize temperate regions. We test these postulates for Himalayan plants, which span subtropical to temperate climates over steep elevational gradients.Location: Himalaya.Time period: Present day.Major taxa studied: Angiosperms.Methods: We use a comprehensive survey of 31 floras to document the elevational and geographical distributions of native Himalayan plants, augmented by field studies of trees in both the east and west Himalaya. We use grade of membership models to cluster species according to locations shared and phylogenetic analysis to evaluate diversification rates.Results: Species fall into four cohesive biotas, organized by climate. Points of turnover between biotas occur where the mean minimum temperature of the coldest month is approximately 0 °C (2,000–2,500 m), and at the point of occasional annual freezing (1,000–1,500 m); these boundaries run the length of the Himalaya. The patterns are retained when we consider whole clades rather than species. All plants (and the subsets trees, herbs and shrubs) belonging to the biota above the 2,000–2,500 m line have higher recent speciation rates than those lower down.Main conclusions: We attribute the high rate of recent speciation in temperate climates to high rates of turnover, creating ecological and geographical opportunity. The high elevation biota has few species, but spans the largest area, implying species numbers are far from any carrying capacity, at least with respect to accumulation of allopatric forms. This study thus links climatic restrictions of clades to differences in diversification rates, and by inference species numbers.
- Open Access
- 25 June 2021
- Published by California Digital Library (CDL) in EcoEvoRxiv
Deep learning as a tool for ecology and evolution
- 25 June 2021
- Published by California Digital Library (CDL) in EcoEvoRxiv
Open AccessDeep learning as a tool for ecology and evolution
- Open Access
- 6 January 2021
- Published by Wiley in Global Ecology and Biogeography
- Vol. 30 (3) , 685-696
Dispersal syndromes drive the formation of biogeographical regions, illustrated by the case of Wallace’s Line
Aim: Biogeographical regions (realms) reflect patterns of co‐distributed species (biotas) across space. Their boundaries are set by dispersal barriers and difficulties of establishment in new locations. We extend new methods to assess these two contributions by quantifying the degree to which realms intergrade across geographical space and the contributions of individual species to the delineation of those realms. As our example, we focus on Wallace’s Line, the most enigmatic partitioning of the world’s faunas, where climate is thought to have little effect and the majority of dispersal barriers are short water gaps.Location: Indo‐Pacific.Time period: Present day.Major taxa studied: Birds and mammals.Methods: Terrestrial bird and mammal assemblages were established in 1‐degree map cells using range maps. Assemblage structure was modelled using latent Dirichlet allocation, a continuous clustering method that simultaneously establishes the likely partitioning of species into biotas and the contribution of biotas to each map cell. Phylogenetic trees were used to assess the contribution of deep historical processes. Spatial segregation between biotas was evaluated across time and space in comparison with numerous hard realm boundaries drawn by various workers.Results: We demonstrate that the strong turnover between biotas coincides with the north‐western extent of the region not connected to the mainland during the Pleistocene, although the Philippines contains mixed contributions. At deeper taxonomic levels, Sulawesi and the Philippines shift to primarily Asian affinities, resulting from transgressions of a few Asian‐derived lineages across the line. The partitioning of biotas sometimes produces fragmented regions that reflect habitat. Differences in partitions between birds and mammals reflect differences in dispersal ability.Main conclusions: Permanent water barriers have selected for a dispersive archipelago fauna, excluded by an incumbent continental fauna on the Sunda shelf. Deep history, such as plate movements, is relatively unimportant in setting boundaries. The analysis implies a temporally dynamic interaction between a species’ intrinsic dispersal ability, physiographic barriers, and recent climate change in the genesis of Earth’s biotas.
- 6 January 2021
- Published by Wiley in Global Ecology and Biogeography
- Vol. 30 (3) , 685-696
Open AccessDispersal syndromes drive the formation of biogeographical regions, illustrated by the case of Wallace’s Line
Aim: Biogeographical regions (realms) reflect patterns of co‐distributed species (biotas) across space. Their boundaries are set by dispersal barriers and difficulties of establishment in new locations. We extend new methods to assess these two contributions by quantifying the degree to which realms intergrade across geographical space and the contributions of individual species to the delineation of those realms. As our example, we focus on Wallace’s Line, the most enigmatic partitioning of the world’s faunas, where climate is thought to have little effect and the majority of dispersal barriers are short water gaps.Location: Indo‐Pacific.Time period: Present day.Major taxa studied: Birds and mammals.Methods: Terrestrial bird and mammal assemblages were established in 1‐degree map cells using range maps. Assemblage structure was modelled using latent Dirichlet allocation, a continuous clustering method that simultaneously establishes the likely partitioning of species into biotas and the contribution of biotas to each map cell. Phylogenetic trees were used to assess the contribution of deep historical processes. Spatial segregation between biotas was evaluated across time and space in comparison with numerous hard realm boundaries drawn by various workers.Results: We demonstrate that the strong turnover between biotas coincides with the north‐western extent of the region not connected to the mainland during the Pleistocene, although the Philippines contains mixed contributions. At deeper taxonomic levels, Sulawesi and the Philippines shift to primarily Asian affinities, resulting from transgressions of a few Asian‐derived lineages across the line. The partitioning of biotas sometimes produces fragmented regions that reflect habitat. Differences in partitions between birds and mammals reflect differences in dispersal ability.Main conclusions: Permanent water barriers have selected for a dispersive archipelago fauna, excluded by an incumbent continental fauna on the Sunda shelf. Deep history, such as plate movements, is relatively unimportant in setting boundaries. The analysis implies a temporally dynamic interaction between a species’ intrinsic dispersal ability, physiographic barriers, and recent climate change in the genesis of Earth’s biotas.
- Open Access
- 9 November 2020
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 117 (47) , 29268-29270
Deep learning in deep time
- 9 November 2020
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 117 (47) , 29268-29270
Open AccessDeep learning in deep time
- Open Access
- 1 June 2020
- Published by Wiley in Applications in Plant Sciences
- Vol. 8 (6) , e11352
Generating segmentation masks of herbarium specimens and a data set for training segmentation models using deep learning
Premise: Digitized images of herbarium specimens are highly diverse with many potential sources of visual noise and bias. The systematic removal of noise and minimization of bias must be achieved in order to generate biological insights based on the plants rather than the digitization and mounting practices involved. Here, we develop a workflow and data set of high‐resolution image masks to segment plant tissues in herbarium specimen images and remove background pixels using deep learning.Methods and Results: We generated 400 curated, high‐resolution masks of ferns using a combination of automatic and manual tools for image manipulation. We used those images to train a U‐Net‐style deep learning model for image segmentation, achieving a final Sørensen–Dice coefficient of 0.96. The resulting model can automatically, efficiently, and accurately segment massive data sets of digitized herbarium specimens, particularly for ferns.Conclusions: The application of deep learning in herbarium sciences requires transparent and systematic protocols for generating training data so that these labor‐intensive resources can be generalized to other deep learning applications. Segmentation ground‐truth masks are hard‐won data, and we share these data and the model openly in the hopes of furthering model training and transfer learning opportunities for broader herbarium applications.
- 1 June 2020
- Published by Wiley in Applications in Plant Sciences
- Vol. 8 (6) , e11352
Open AccessGenerating segmentation masks of herbarium specimens and a data set for training segmentation models using deep learning
Premise: Digitized images of herbarium specimens are highly diverse with many potential sources of visual noise and bias. The systematic removal of noise and minimization of bias must be achieved in order to generate biological insights based on the plants rather than the digitization and mounting practices involved. Here, we develop a workflow and data set of high‐resolution image masks to segment plant tissues in herbarium specimen images and remove background pixels using deep learning.Methods and Results: We generated 400 curated, high‐resolution masks of ferns using a combination of automatic and manual tools for image manipulation. We used those images to train a U‐Net‐style deep learning model for image segmentation, achieving a final Sørensen–Dice coefficient of 0.96. The resulting model can automatically, efficiently, and accurately segment massive data sets of digitized herbarium specimens, particularly for ferns.Conclusions: The application of deep learning in herbarium sciences requires transparent and systematic protocols for generating training data so that these labor‐intensive resources can be generalized to other deep learning applications. Segmentation ground‐truth masks are hard‐won data, and we share these data and the model openly in the hopes of furthering model training and transfer learning opportunities for broader herbarium applications.