Identifying escaped farmed salmon from fish scales using deep learning

TL;DR

This paper shows how deep learning can accurately identify escaped farmed salmon from fish scales using a large and diverse dataset.

Contribution

The study introduces a deep learning model trained on a large, diverse dataset of fish scales to identify escaped farmed salmon with high accuracy.

Findings

The model achieved an F1 score of 0.95 on a large independent test set.

The model's predictions matched genetic reference samples and known farmed-origin scales closely.

Deep learning generalizes robustly across ecological and methodological contexts.

Abstract

Escaped farmed salmon are a major concern for wild Atlantic salmon (Salmo salar) stocks in Norway. Fish scale analysis is a well-established method for distinguishing farmed from wild fish, but the process is labor and time intensive. Deep learning has recently been shown to automate this task with high accuracy, though typically on relatively small and geographically limited datasets. Here we train and validate a new convolutional neural network on nearly 90 000 scale images from two national archives, encompassing heterogeneous imaging protocols, hundreds of rivers, and time series extending back to the 1930s. The model achieved an F1 score of 0.95 on a large, independent test set, with predictions closely matching both genetic reference samples and known farmed-origin scales. By developing and testing this new model on a large and diverse dataset, we demonstrate that deep learning generalizes robustly across ecological and methodological contexts, supporting its use as a validated, large-scale tool for monitoring escaped farmed salmon.

Figures