A Monocular Depth Estimation Framework for Improved Underwater Fish Biomass Assessment

Abstract

Accurate and automated fish weight estimation is a critical component of modern aquaculture, enabling optimized feeding strategies, improved fish welfare, and effective production planning. A common approach is the use of computer vision techniques to estimate fish mass while the fish are swimming, thereby eliminating the need for manual handling and reducing stress on the fish. This paper presents a computer vision pipeline that integrates monocular depth estimation using MiDaS with YOLOv11 object detection, trained on a real-world underwater dataset of Nile tilapia covering multiple age groups. The new method requires only a monocular camera, eliminates the need for manual fish orientation, and enables fully automatic weight prediction through a CatBoost regressor. Experiments conducted across different fish age groups and tank conditions demonstrate consistent and robust performance. In this context, YOLOv11 achieves an F1-score of 0.93 at an IoU threshold of 0.5. Additionally, the weight estimation model attains a mean absolute error (MAE) below 4 g and an $R^{2}$ value exceeding 0.95. These results suggest that the proposed pipeline approach delivers accurate and reliable weight estimation without the need for using stereo-cameras, making it very well-suited for practical deployment in real-world aquaculture environments.