Applications of deep convolutional neural networks to digitized natural history collections

Abstract

Natural history collections contain data that are critical for many scientific endeavors. Recent efforts in mass digitization are generating large datasets from these collections that can provide unprecedented insight. Here, we present examples of how deep convolutional neural networks can be applied in analyses of imaged herbarium specimens. We first demonstrate that a convolutional neural network can detect mercury-stained specimens across a collection with 90% accuracy. We then show that such a network can correctly distinguish two morphologically similar plant families 96% of the time. Discarding the most challenging specimen images increases accuracy to 94% and 99%, respectively. These results highlight the importance of mass digitization and deep learning approaches and reveal how they can together deliver powerful new investigative tools.

The history and impact of digitization and digital data mobilization on biodiversity research

“…Others have addressed emerging research angles, including the supplementation of existing datasets with related digital layers to enhance niche and species distribution modelling [54]; the use of 3D data for generating and testing new hypotheses; the implementation of convolutional neural networks (CNN) and deep learning in the analysis of image data for taxonomic determination [55][56][57] and specimen curation [58], the delineation of traits in specimen images and the determination and identification to genus or species of sediment-deposited pollen grains [59].…”

Using a Convolutional Siamese Network for Image-Based Plant Species Identification with Small Datasets

“…Because the scope of this experiment is limited to leaf image datasets, further research is needed to address the same issues when other types of images are used. For example, herbarium sheet datasets have been used to identify plants with CNNs [14,29]; however, for certain regions, datasets are rather small. Additionally, we are currently working on the problem of identifying tree species in a context where very few images are available worldwide, namely, tree species identifications based on wood cut images.…”

Taxonomic Classification of Ants (Formicidae) from Images using Deep Learning

“…Related deep learning studies CNNs have been used in plant identification (Lee et al 2015; Lee et al 2016; Dyrmann et al 2016; Barré et al 2017; Sun et al 2017), plant disease detection (Mohanty et al 2016) and identification of underwater fish images (Qin et al 2016), all with high accuracy (71% – 99%). Applied examples with high accuracy include classification of different qualities of wood for industrial purposes (79%) (Affonso et al 2017), identifying mercury stained plant specimens from non-stained (90%) (Schuettpelz et al 2017), and identification of specimens using multiple herbariums (70% – 80%) (Carranza-Rojas et al 2017). Especially studies like the last two are important for natural history collections, because such applications can benefit research, speed up identification and lower costs.…”