CleanCam: a labelled image dataset for camera-cleaning decisions in aquaculture monitoring

CleanCam is a benchmark dataset for underwater camera viewport fouling severity assessment in aquaculture. The dataset is designed to distinguish viewport-attached contamination from general water-column degradation, such as turbidity, haze, suspended particles, and low contrast, since these conditions can look similar in individual frames but imply different operational responses.

The frozen release contains 22,572 RGB JPEG images in total, including 18,972 real images sampled from fixed underwater monitoring videos collected over 20 effective collection days at Research Institute for Aquaculture No1 in Hai Phong, Vietnam, and 3,600 split-consistent synthetic images used to supplement underrepresented moderate-to-heavy fouling cases. Images are annotated using a five-level ordinal severity protocol focused on viewport fouling, where clean-but-turbid frames remain in Level 1 and higher levels require stable viewport-attached evidence across time.

This release includes:

• real and synthetic image folders organized by severity label
• metadata tables for all images
• official deterministic capture-disjoint train/validation/test splits
• synthetic provenance information
• release manifests and documentation

The metadata provides image-level information such as image ID, file path, label, camera, session, date, elapsed seconds, and capture identifier. For synthetic images, metadata additionally records parent image linkage and generator parameters.

CleanCam is intended as a reproducible resource for maintenance-aware underwater vision, camera-health monitoring, and reliable long-term aquaculture imaging systems.