A COMPARISON OF DEEP LEARNING MODELS FOR EELGRASS MEADOWS MONITORING.

Summary

This report intends to comprehensively compare how YOLO instance segmetation models performe in complex environments with sophisticated components such as eelgrass detection and density estimations. It studies their use in ecological monitoring and management, while also examing the efficiency of using them in contrast with traditional studies. 

Biological target:

The objective species was Zostera marina, common name eelgrass. However, the models were also trained for other species and taxonomic groups. 

Methodology:

The images used in this project were obtained with Remotely Operated Vehicles deployed in Sweden's Skagerrak sea during October 2024 and April 2025. From the videos recorded, up to 260 random images were extracted and then annotated, reaching a total of 1014 annotations distributed between 18 classes. The platforms used were Biigle for annotating, SUBSIM (Swedish platform for subsea image analysis) for model training and Weights and Biases for performance analysis. Integration scripts can be found in GitHub (https://github.com/PauSowa/Model-Implementation-for-Eelgrass-Meadows-Mapping). 

Best Performing Model based on Metrics - K_IS_R

Best Performing Model based on Ground truth (eelgrass detection) - KB_IS_R

Best Performing Model based on Ground truth (density estimation) - K_IS

Contents: 

Report: A Comparison Of Deep Learning Models For Eelgrass Meadows Monitoring. 

Model: KB-IS-R 

Other models and dataset are available upon request with correct permissions.