COMPUTER VISION FOR AUTOMATED PHENOTYPING: WEIGHT PREDICTION FOR GENOMIC EVALUATION OF GROWTH AND RESILIENCE INDICATORS (LnVar) IN YELLOWTAIL KINGFISH Seriola lalandi
Introduction
Only a very small portion of global aquaculture production is derived from structured breeding programs. A major bottleneck in implementing aquaculture breeding programs is the difficulty of removing aquatic animals from their water environment to measure phenotypes. Over the last decade, the development of machine learning (ML) and computer vision (CV) techniques has provided non-invasive alternatives using image and video analysis to be applied in animal production systems, opening a new possibility for high-throughput phenotyping with minimal disturbance to the animals. This study uses computer vision to predict animal weight directly from images for estimating phenotypic data in Yellowtail Kingfish (Seriola lalandi) under commercial conditions, simplifying the repeated measurements that are required for phenotypes such as growth rate the LnVar resilience indicator. Repeated measurements are combined with genotypes to apply genomic analyses of these image-derived measurements.
Materials and Methods
A total of 1,600 PIT-tagged kingfish juveniles were randomly distributed into two treatment tanks at The Kingfish Company (Kats, the Netherlands) and monitored via monthly biometrics over five months. During each event, automated digital imaging (Dorset ID, Aalten, the Netherlands) was performed, and a subset was physically weighed for model calibration. Subsequently, a deep-learning segmentation model processed the images to extract pixel-based area and volume (Figure 1). Quadratic regression models predicted weights for the unweighed population using extracted pixel traits, subsequently allowing the calculation of the daily growth coefficient (DGC) and the LnVar resilience proxy (Mengistu et al., 2022). Genomically, 1,140 individuals were genotyped (Benchmark Genetics, Bergen, Norway). Post-quality control, 25,160 retained SNPs were used to construct a genomic relationship matrix (GRM) via Calc_grm (Calus & Vandenplas, 2016), which was utilized in ASREML to estimate variance components and heritabilities through a univariate model accounting for the treatment effect
Results
By physically weighing a calibration subset of 100 fish per standard timepoint, the quadratic regression models applied to these calibration subsets yielded coefficients of determination () ranging from 0.914 to 0.977 and low Mean Absolute Percentage Error (MAPE) between 2.25% and 5.01%, successfully predicting the weights of the remaining unweighed population (up to 770 individuals per treatment per timepoint). Over the entire experiment, a total of 1,494 manual scale measurements estimated 6,299 individual weights. Heritabilities () for predicted body weight progressively increased from 0.39 ± 0.05 at early stages (WT 1) to 0.48 ± 0.05 at harvest (WT 5). Growth traits demonstrated similarly robust heritabilities, with 0.40 ± 0.05 for DGC, whereas resilience (LnVar) exhibited a low heritability of 0.04 ± 0.02 (Table 1).
Discussion
The computer vision approach utilized in this study successfully predicted the entire population-level weights by weighting only 17%±7.36% of the sampled animals per timepoint, providing a significantly faster and reliable procedure. Regarding the genomic results, while overall growth measurements exhibited moderate heritabilities of approximately 0.4, the LnVar resilience index demonstrated near-zero heritability, rendering its evaluation suggestive at best. Study of growth variation and resilience remains of significant interest, which strictly requires frequent longitudinal phenotyping. The integration of computer vision combined with artificial intelligence in aquaculture provides a framework for rapid and precise phenotyping as demonstrated in this study. Nevertheless, to meaningfully improve animal welfare and eliminate handling-induced stress, future technologies must evolve to capture images without sedatives and, preferably, without any physical handling of the fish. As these computer vision systems advance toward fully non-invasive, in-water monitoring, they will become critical for evaluating complex resilience traits under commercial conditions.
Table/Figure 1
Figure 1
Image processing pipeline for automated phenotyping
Table 1. Variance estimates and heritabilities of weight, growth, and resilience
Trait
N
VA se
VE se h2 ± se
Weight 1
1100
302
50.8
473.5
26.7
0.39 (0.05)
Weight 2
1100
2514.2
403.6
3992.7
217.2
0.39 (0.05)
Weight 3
1132
9114.4
1383.1
12218.5
678.8
0.43 (0.04)
Weight 4
1097
29884.2
4480.4
35238.7
2051.2
0.46 (0.05)
Weight 5
1103
69974.1
10428.3
75019.3
4478.8
0.48 (0.05)
DGC 1-5
1063
0.248
0.04299
0.367
0.0217
0.40 (0.05)
LnVar
1140
0.0576
0.0287
1.296
0.0575
0.04 (0.02)
DGC: Daily Growth Coefficient; LnVar: Logtransformed variance if weight deviations;
Acknowledgment
This study received funding as public-private partnership from the Dutch Ministry of Agriculture, Nature and Food Quality and The Kingfish company (grant number LWV20.37)
References
Vandenplas, J., & Calus, M. P. L. (2025). Calc_grm – a program to compute pedigree, genomic, and combined relationship matrices. Animal Breeding and Genomics Centre, Wageningen, The Netherlands.
Mengistu, S. B., et al. (2022). Fluctuations in growth are heritable and a potential indicator of resilience in Nile tilapia (Oreochromis niloticus). Aquaculture, 560, 738481.