Automated aquatic biodiversity monitoring using deep learning on the Tigris River: Species identification and ecosystem assessment

Abstract

Aquatic ecosystems play a crucial role in biodiversity and ecological stability but are increasingly threatened by climate change, pollution, habitat degradation, and invasive species. Traditional monitoring methods are labor-intensive, costly, and limited in spatial and temporal coverage. This study integrates deep learning techniques with biodiversity monitoring to enhance species identification, abundance estimation, and ecosystem assessment in freshwater environments. Focusing on the Tigris River, Iraq, we developed convolutional neural network (CNN) - based models to automate species detection and classification from underwater imagery. Our multi-tiered data collection approach, which includes direct field sampling, remote sensing, and citizen science, yielded a dataset of over 8,000 images across six camera locations. The Faster R-CNN model achieved a mean average precision (mAP) of 88% for fish identification, while U-Net segmentation models demonstrated 99% accuracy in organism detection, significantly outperforming traditional methods. The application of optimized deep learning models significantly enhanced the accuracy and efficiency of aquatic biodiversity monitoring. The Faster R-CNN model, after hyperparameter optimization and transfer learning, achieved an accuracy of 88% in species identification, outperforming baseline models that averaged around 75%. The optimization techniques, particularly data augmentation and early stopping, improved the model's robustness to environmental variations, such as high turbidity and poor lighting conditions. Unlike traditional methods that rely on expert identification, the deep learning model provided automated, scalable, and real-time monitoring capabilities, reducing the need for labor-intensive field surveys. Additionally, the model demonstrated higher precision in detecting species that are typically misclassified in traditional statistical models, thereby offering a more reliable approach to biodiversity conservation and ecological assessments. These results underscore the potential of deep learning to provide scalable, automated, and highly accurate biodiversity assessments. Our findings demonstrate how artificial intelligence can revolutionize ecological conservation, offering a cost-effective and reliable solution for biodiversity monitoring. The study also emphasizes the importance of interdisciplinary approaches in addressing global biodiversity loss and advancing conservation strategies.