Delta Smelt, an endangered fish species, plays a critical role in maintaining biodiversity, genetic diversity, and ecological balance in California. Accurate sex identification of Delta Smelt is crucial for optimizing breeding and operational strategies and enhancing the conservation of this species. Currently, sex identification of Delta Smelt primarily relies on egg stripping during the sexually mature stage. Determining their sex is almost impossible without the presence of eggs or milt in the field.
This study presents a novel approach to recognize sex of Delta Smelt by utilizing deep neural networks to recognize differences on phenotypic appearance. We collected data from Delta Smelt across three life stages—before the spawning season (Section 1), during the spawning season (Section 2), and at the end of the spawning season (Section 3). The framework consists of two key steps: first, segmenting the fish from the background using a fundamental visual model, followed by applying a classification model to distinguish the sex of the Delta Smelt. Segment Anything Model, as a visual fundamental model, is employed to segment fish from the background with the visual prompt, reducing the cost of label annotation and model training. Transfer learning-based fine-tuning techniques are used to train the parameters of final convolutional and fully connected layers for the classification task.
The model achieves an accuracy of 92.75% and an F1 score of 92.44% on the test dataset for Delta Smelt in Phase 3. We further examine the interpretability of the model in recognizing the sex of Delta Smelt. This method provides a promising and potential approach to sex recognition based on phenotypic visual features, offering a valuable tool for the conservation and management of Delta Smelt populations, simultaneously providing inspiration for sex identification for other fish species.