Utilizing Geometric Morphometrics & Artificial Intelligence to Identify Early Life Stages of Klamath Suckers to Support Hatchery Operations

Authors: Josh Rasmussen, Jane Spangler, Michelle Jackson

Year: 2022

Abstract

The artificial propagation program for the endangered shortnose sucker, Chasmistes brevirostris, and Lost River sucker, Deltistes luxatus, relies on larvae collected from the natural spawn to generate numbers each year. The resulting cohorts include individuals from the two listed species as well as Klamath largescale sucker, Catostomus snyderi. These species are closely related but readily distinguished with external morphology as adults. However, classifying Klamath sucker larvae and juveniles by morphology is extremely difficult and likely unreliable. To date, attempts to identify early life stage Klamath suckers has been largely qualitative. However, even quantitative studies have struggled to separate the three species due to hybridization (Markle et al. 2005, West. N.A. Nat. 65:473-489). To promote quantitative, reliable identification of juvenile Klamath suckers in the Klamath Falls National Fish Hatchery, we collected high quality ventral head (i.e., lips and mouth) digital images, digitized six landmarks to capture lip and mouth morphology, and standardized the landmarks using General Procrustes Analysis. The standardized data were used to generate relative warps to identify how many groups were present and inputted into a neural network to determine how best to describe those groups. Discovered groups in our data corresponded predominantly to the three species, indicating that morphological differences existed even at relatively smaller sizes, even though such differences are challenging to see with the naked eye. The neural network model adequately identified individuals in the validation set to the correct species. This approach can provide real-time species identification at a fraction of the cost of genetic alternatives, while potentially leading to a greater understanding of the dynamic of hybridization in conjunction with genetic analyses. Automation of image and data collection with a handheld digital device (e.g. a tablet) is necessary to make the methodology useful in real-time. Refinement of the neural network model (e.g., optimizing the number of hidden layers and nodes) is needed to maximize model utility.