akridge/MOUSS_fish_imagery_dataset_grayscale_small

Dataset Card for Modular Optical Underwater Survey System (MOUSS) Imagery - Small Set

This dataset contains grayscale underwater imagery collected by NOAA's Modular Optical Underwater Survey System (MOUSS), specifically for object detection of fish. The dataset is intended for training and evaluating models like the YOLOv8n-based Fish Detector on grayscale underwater footage.

Dataset Details

Dataset Description

This dataset is composed of black-and-white underwater footage captured by the MOUSS system. The dataset provides labeled grayscale images for developing and testing object detection models focused on detecting fish.

• Curated by: Pacific Islands Fisheries Science Center (PIFSC)
• Shared by: NOAA Open Data
• License: Public Domain (NOAA Open Data)

Dataset Sources

• Data Repository: NOAA Open Data Dissemination (NODD) PIFSC Cloud

Uses

Direct Use

This dataset is designed for testing and training object detection models, particularly for detecting fish in grayscale underwater imagery. It can be used in marine biology research, real-time monitoring, and post-processing of underwater footage.

Out-of-Scope Use

The dataset is not suitable for color imagery or tasks unrelated to fish detection in marine ecosystems. It should not be used for tasks requiring full-spectrum image analysis or where higher accuracy for non-grayscale footage is needed.

Dataset Structure

The dataset consists of grayscale images with corresponding annotations (bounding boxes) around fish. Images are split into 80% for training and 20% for validation, enabling effective training of detection models such as the YOLOv8n architecture used in the Fish Detector.

Dataset Creation

Curation Rationale

The MOUSS imagery dataset was created to provide a focused resource for researchers and data scientists working on underwater object detection, specifically for fish detection in grayscale environments. This small subset offers a balance between sufficient sample size and practical use for model development.

Source Data

Data Collection and Processing

Images were captured using NOAA’s MOUSS system during surveys. The dataset consists of grayscale (black-and-white) underwater images to support the training of models on a specific environment with limited color information.

Personal and Sensitive Information

This dataset contains no personal or sensitive information, as it consists solely of underwater imagery.

Bias, Risks, and Limitations

Bias

The dataset is focused on grayscale underwater footage, which may limit the model’s generalization to other environments, such as color underwater footage or different aquatic ecosystems. Furthermore, the limited number of species in the dataset could introduce biases in identifying different types of fish or marine life in other contexts.

Recommendations

While this dataset can effectively train fish detection models for grayscale footage, users should be cautious when applying the model to other environments or color imagery. Additional training or domain adaptation may be necessary to avoid misclassifications or errors.

Model Details for Associated YOLOv8n Fish Detector

• Model Architecture: YOLOv8n
• Task: Object Detection (Fish Detection)
• Footage Type: Grayscale (Black-and-White) Underwater Footage
• Classes: Fish
• Training Method: Unsupervised learning with limited labeled data
• Key Metrics:
  • Precision: 0.863
  • Recall: 0.869
  • mAP50: 0.936
  • mAP50-95: 0.856
• Training Data: 80% training and 20% validation split, using grayscale underwater images from this dataset.
• Training Parameters: 50 epochs, 0.001 learning rate, 416x416 image size.

Citation

Pacific Islands Fisheries Science Center, 2024: Bottomfish Fishery-Independent Survey in Hawaii (BFISH) - Experimental Camera Surveys (2011-2015), https://www.fisheries.noaa.gov/inport/item/55928.

Dataset Card Contact

For questions or more information, please contact: Michael Akridge ([email protected])