BIO-AI: STAR: IUSC: Morphological Barcoding, a novel, machine learning-based approach to solving the taxonomic impediment

Taxonomists working since the time of Linnaeus have described between 1.2 million and 1.5 million species. A 2011 study predicted that the number of eukaryotes alone on Planet Earth is 8.7 million species. The number of described fish species is between 32,000 and 36,000 species. The 2011 study predicted the actual number of ray-finned fishes alone to be 100,000 species. At the rate taxonomists are presently describing new fish species, it will take 217 years to describe all the fishes estimated to live in the world’s rivers, lakes and oceans. The problem is that traditional methods of taxonomy, whether involving fishes or other kinds of organisms, are laborious, time consuming, and frequently don’t yield characters useful for diagnosing populations under study as quantifiably different from populations of described species they are being compared to. The premise of the proposed project is that data extracted from images of fish specimens with AI methods will generate larger amounts of data for newer, more useful kinds of data characters than those traditionally gathered by ichthyologists and will aid fish taxonomists in their task of diagnosing undescribed fish species much more rapidly. We call this new method of taxonomic diagnosis morphological barcoding. This interdisciplinary project will use a resource of 236,435 bounding-boxed fish specimen images representing 15,262 fish species from 4,006 genera and 604 families — associated with an existing platform called FishAIR — in experiments designed to demonstrate the effectiveness of the morphological barcoding approach. The feature data we will be capturing include segmentation data, histograms, color palette data, pixel strips, outlines and automated landmarks placed at homologous points on the body of a fish specimen to determine the position and area of features like the eyes, head, nares (nostrils), mouth, trunk and fins of specimen images, all recorded as x, y coordinates. Machine-learning tools developed by collaborators at Virginia Tech will be used to segment the external body features named above and assign unique colors to each feature. The morphological barcoding workflow will 1) collect image data and metadata from FishAIR, 2) extract data on as many morphological features as possible, transforming raw data into meaningful features that can be used by the machine learning algorithms; 3) Perform several ML feature selection tasks and statistics to identify the most distinctive morphological characters(species profiling); 4) generate barcodes using the results of feature selection and species profiling processes to create a unique sequence of characters for each species, starting with species that are best represented in our dataset; and 5) test the barcodes and modify the processes to improve accuracy and efficiency. A postdoc will be trained in machine learning methods and will gain hands on experience with all aspects of the morphological barcoding workflow, including new species diagnosis and publishing descriptive taxonomy papers. We will host a workshop to train ichthyologists, other biologists, ML scientists, and biodiversity data managers and engineers how to use all our morphological barcoding resources. AI methods, a NSF priority, are integral to the work of this project. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.