Automatic, operational, high-resolution monitoring of fish length and catch numbers from landings using deep learning

Abstract

Informed fishery management decisions require primary input data such as the fluctuations in the number of fish landed and fish length. Obtaining these data can be costly if conducted by hand, which is the case for length data in most fisheries. This cost often implies reduced sample sizes, which may introduce biases and lead to information loss at, for example, the boat level. The recent boost in artificial intelligence applied to fisheries provides a promising way to improve the assessment and management of stocks. We present an operational system using a deep convolutional network (Mask R-CNN) coupled with a statistical model that automatically estimates the number and the mean fork length of dolphinfish (Coryphaena hippurus) caught in a Mediterranean fishery with a resolution of each landed fish box from each boat. The system operates on images of fish boxes collected automatically at the centralized fish auction. The statistical model corrects for biases due to undetected fish using the convolutional network and estimates the mean fork length of the fish in a box from the number of fish and the box weight, allowing for high-resolution monitoring of fishery dynamics during the entire fishing season. The system predictions were empirically validated and showed good accuracy and precision. Our system could be readily incorporated into assessment schemes. We discuss how this type of monitoring system opens new opportunities for improving fishery management.