Denise Klinkenbuß, Olivia Metz, Jessica Reichert, Torsten Hauffe, Thomas A. Neubauer, Frank P. Wesselingh, Thomas Wilke
Malacologia 63 (1), 95-105, (10 September 2020) https://doi.org/10.4002/040.063.0109
KEYWORDS: Bivalvia, morphometry, 3D scanning, species identification, Caspian Sea
In recent years, 3D analyses, new indices to describe the complexity of morphological structures and sophisticated machine learning approaches have advanced morphometrical analyses to assist species determination. However, the applicability of these modern approaches to the determination of cryptic species or fossil taxa has rarely been investigated.
In this study, fossil and subfossil specimens of the four modern Dreissena species in the Caspian Sea are used to test the performance of 3D-based morphological approaches for machine learning assisted species identification. Specifically, 3D scans of the shells were used to construct 3D models for calculating “traditional” shell dimensions and “modern” shell complexity parameters. Finally, two machine learning approaches were applied to test the determination performance of shell measurements vs. shell complexity and individual vs. combined shell parameters.
The results show that (i) there is no superior machine learning approach to species determination based on shell characters, (ii) shell complexity parameters are not per se more suitable for species identification than shell dimensions, (iii) a combination of shell parameters increases determination performance and reduces their species dependence and (iv) shell characters alone do not allow precise determination of all Dreissena species studied.
These findings suggest that the most appropriate machine learning approach, the most informative shell characters and the best combination of characters need to be tested individually for different data sets. However, considering that it is difficult even for experts to distinguish Dreissena species based on shell characters, the machine learning assisted classification in the current study has performed comparatively well. Future analyses based on machine learning may therefore help experts to process large sample sizes efficiently and non-specialists to assess species level information with reasonable certainty.