Is Eye Allometry the Silver Bullet for Measuring Marine Fishes with a Single Camera?
Reliable length frequency distributions are useful tools for the assessment and management of fish populations. However, in situ estimation of fish length with single-camera video systems may be inaccurate. The use of allometric changes of eye diameter (ED) and head height (HH) was proposed as a promising technique to estimate fish length from single-camera footage. However, the method, so far, has proven to be effective for a limited number of species and families, and the replicability among taxa remains uncertain. We described the allometric patterns of eye diameter and head height for a set of Osteichthyes and Chondrichthyes bearing heterogeneous morphological, taxonomical, and habitat-type features from northern Patagonia, Argentina. To assess the usefulness of the method, we regressed fish length on the HH:ED ratio for each of 12 common species by using artificial neural network models and compared the results with those obtained by standard regression models. We found relevant variability in allometry among the species analyzed and described four general cases. Furthermore, artificial neural network models outperformed conventional regression, although both models converged to similar results for several species. Overall, our findings highlighted the utility of the allometry method and suggested an association of both allometric patterns and model performance with fish morphology and habitat type. Whereas osteichthyan reef species were highly suitable for using this technique, models obtained for chondrichthyans and fusiform body-shaped species were less accurate. Our findings have practical implications for the estimation of fish length using single-camera systems and provide a framework to understand the differential suitability of the allometric approach.
Study site with details on the geographic location of the data collected.
Artificial neural network (ANN) modeling process. (A) General ANN architecture diagram; (B) example of the ANN backpropagation training and testing process with the loss function (error) optimization across 200 epochs for one species (Acanthistius patachonicus). Lr: learning rate; bs: batch size; n_h: number of neurons in the hidden layer; p: dropout.
Four examples (two Osteichthyes and two Chondrichthyes) representing different allometric cases in marine fishes, identified through the change in head height (green lines) and eye vertical diameter (red lines), as percentages of the total body length. The cases were defined on the basis of combining the type of eye and head height allometry, which were categorized separately as negative (–), isometric (=), or positive (+).
Artificial neural network (ANN) and standard regression (red and blue lines, respectively) models for the relationships between the ratio of head height to eye diameter (HH:ED ratio) and total length for 12 species of marine fishes from Patagonia, Argentina.
(A) Multiple correspondence analysis (MCA) showing the position of the species (numbers) in the first two dimensions. (B) Alluvial plots showing the relationship between types of allometry (i.e., Case), habitat type, and body shape in the studied species: block sizes represent the proportion of species for each level, and stream fields between the alluvial diagrams’ blocks represent the proportion of species corresponding to each level through all categories. 1—Genidens barbus; 2—Odontesthes platensis; 3—Callorhinchus callorynchus; 4—Percophis brasiliensis; 5—Pinguipes brasilianus; 6—Pseudopercis semisfasciata; 7—Scomber colias; 8—Sebastes oculatus; 9—Acanthistius patachonicus; 10—Diplodus argenteus; 11—Pagrus pagrus; 12—Mustelus schmitti.
Histograms and boxplots showing the length frequency distributions for nine Patagonian marine fishes sampled with Baited Remote Underwater Video Stations (BRUVS), estimated with three different methods: reference of known size in the visual field (Ref.), standard regression (Regr.), and artificial neural networks (ANN). The semitransparent pink shaded area indicates the length range used to calibrate the ANN and standard regression model (i.e., interpolation range) for each species. The number of individuals measured for each species (n) is shown.
Contributor Notes
Associate Editor: W. L. Smith