SHR_DSCAM

YOLOV11 Object Detection Model for Southern Hydrate Ridge Imagery

This YOLOV11 model was localized and fine-tuned on images from the National Science Foundation's Ocean Observatories Initiative (NSF OOI) Regional Cabled Array (RCA) Digital Still Camera at Southern Hydrate Ridge. It focuses on identifying higher-level taxonomic groups within benthic ecosystems, making it particularly suitable for ecological studies and marine habitat monitoring in methane seep environments.

Dataset

The dataset used for training includes images collected at Southern Hydrate Ridge, part of the NSF OOI RCA. This location provides unique and ecologically valuable imagery of methane seep sites, including biogenic structures such as bacterial mats and associated fauna.

The model was trained on a set of 12 different classes, focusing on higher-level taxonomy compared to the original FathomNet model, which contained more fine-grained labels. The classes used in training are:

Anoplopoma (Sablefish)

Asteroidea (Sea Stars)

Chionoecetes (Snow Crabs)

Eptatretus (Hagfish)

Euphausia (Krill)

Liponema (Sea Anemones)

Microstomus (Dover Sole)

Poralia (Jellyfish)

Rajiformes (Skates)

Sebastes (Rockfish)

Vesicomyidae (Clams)

Zoarcidae (Eelpouts)

It is important to note that the classes Sebastes and Vesicomyidae are significantly overrepresented in the dataset. This may influence model predictions, leading to better detection performance for these taxa compared to others. Users should keep this in mind when evaluating model outputs.

Training Details

This model was trained using the FathomNet/MBARI-315k-yolov8 as a base, with adjustments made to adapt the model to higher-level taxonomic detection and specific environmental characteristics at Southern Hydrate Ridge. The training focused on balancing the various classes present in the dataset to improve accuracy, despite some taxa being overrepresented.

Training was performed on images at the original resolution, allowing for high fidelity of features required for accurate detection of benthic fauna in deep-sea imagery.

Metrics

The model has been evaluated using accuracy as a primary metric for performance. Further evaluation metrics such as precision, recall, and F1 score can be found in future updates of the modelcard, pending further benchmarking efforts.

Applications

This YOLOV11 model is well suited for:

Ecological monitoring of methane seep environments.

Detection and classification of benthic megafauna at the taxonomic levels relevant to the Southern Hydrate Ridge.

Assisting researchers in automated annotation of underwater imagery, reducing manual labor and improving data collection efficiency.

License

This model is released under the MIT License, allowing for open use and modification, with the intention of fostering collaboration and innovation in marine ecology and deep-sea research.