Optical imaging and machine learning to identify and enumerate early developmental stages of fish species

Larval fish are a key component for biomonitoring aquatic system health but are notoriously difficult to identify. While recent advances in optical imaging and artificial intelligence (AI) have assisted in identifying organisms, most applications for fish are limited to adults, not juveniles. Here, we evaluate the capabilities of AI to automatically identify and enumerate larval fish for use in taxonomic identification and aquaculture. To accomplish this, we used an optical imaging system and trained machine learning (ML) models, support vector machines (SVM) and convolutional neural networks (CNN), on eggs and at least two early larval stages of three fish species, channel catfish (CCF), Florida largemouth bass (FLLMB), and koi. When considering all species and life stages, accuracies ranged from 0.34 to 0.65 for SVM and CNNs, respectively; however, when non-target items (i.e., bubbles and detritus) were included, accuracies dramatically increased to 0.92 to 0.94 (SVM and CNN). Experiments to differentiate life stages for each species individually showed accuracies ranging from 0.53 to 0.74, increasing to 0.88 to 0.97 with non-targets. Generally, CNNs outperformed SVMs, whereas the number of predicted classes reduced accuracy. Our results suggest that while current optical imaging technologies can successfully differentiate larval fish from non-target items, their ability to differentiate amongst larval fish species and life stages would benefit from improvement to optics, ML algorithms, and a more extensive training library. This work also suggests tradeoffs exist between taxonomic accuracy and sample volume processing between small-volume, desktop systems and high-throughput, field-scale systems. • Optical imaging and machine learning to identify larval fish and eggs. • Convolutional neural networks generally outperformed support vector machine. • Accuracy could be improved with better object detection software. • Larger image training libraries could improve accuracy. • High accuracy (>97%) in separation of targets from non-targets.