Estimation of the fish number in farming cage from the fish finder echo images via machine learning

Abstract

Determining the precise number of farmed fish in a preserve is crucial, as it informs the quantity of feed required. However, especially for ocean-based preserves, this measurement poses a significant challenge. While a fish finder is useful for investigating the state of fish schools in the water, the clarity of the echo image is insufficient for individuals to count each fish. This study aims to address this issue by developing a novel method for fish counting, leveraging combined simulation and machine-learning techniques applied to the echo images from the fish finder. Two key innovations underpin our approach. First, we employ a convolutional neural network (CNN) to extract features from echo images with varying fish numbers to estimate the fish count from a portion of fish school images. Second, we generate virtual echo images via simulation to create a training dataset for the CNN. This step is necessary because CNN image classification requires a large number of images, and preparing such a vast collection of echo images with accurate labels through real-world fish finder observations in a fish preserve is time-consuming. We tested our method in six farming cages housing yellowtail (Seriola quinqueradiata). Our method could estimate the fish count with errors ranging from −0.86 % to 6.89 %. While our method requires further refinement, the results suggest the potential for a new fish-counting system.