Linking fish activity and turbidity through visual and sensor data fusion and deep learning

Monitoring underwater environments is crucial for industrial applications, providing data that can be used for reporting against corporate sustainability and environmental goals. This study presents a novel approach to integrating high-resolution underwater imaging and high-tech water quality sensing with deep learning models to detect fish, estimate turbidity in Nephelometric Turbidity Units (NTU), and analyze their interactions. An IP-based underwater camera and two advanced water quality sensors were deployed at the Port of Mackay (northern Queensland, Australia) to collect synchronized visual and water quality data. A significant portion of collected images lacked valid turbidity values due to camera and sensor synchronization issues. To address this, we developed a custom Convolutional Neural Network (CNN) model for image-based turbidity estimation. Additionally, YOLOWorld-based prompt-able object detectors were used and evaluated for fish detection, with YOLOWorld-v1 Large emerging as the best choice, achieving 89.7 % accuracy without any training. Our proposed CNN water turbidity estimation model gained root mean square error of 1.6 NTU. Using these deep learning models, we found a non-linear correlation between fish count and water turbidity with an R² of 0.93. This finding is aligned with previous research and highlights the complex interplay of environmental factors in marine ecosystems, while showcasing how technological advances can streamline ecological studies. Downstream applications of this technology could include permanently installed underwater cameras in port waters that record real-time data. Management responses could then be automatically triggered when water quality parameters exceed threshold levels, providing early warnings and enabling timely actions to protect marine ecosystems.