Side-Scan Sonar Images of Marine Engineering Structures

Marine_PULSE contains 627 labeled side-scan sonar images across four categories: 323 pipeline or cable, 134 underwater residual mound, 88 seabed surface, and 82 engineering platform images. Xing Du from the First Institute of Oceanography compiled the dataset using instruments like EdgeTech4200FS and Klein-3000, with images processed via the PostSurvey software. The data is pre-split into training and test sets for machine learning applications.

Use Cases

• Train a convolutional neural network to classify sonar images into the four provided categories: pipeline or cable (POC), underwater residual mound (URM), seabed surface (SS), and engineering platform (EP).
• Benchmark pre-training strategies for object detection models using the 323 POC and 82 EP image labels as target objects.
• Analyze feature representations learned from data captured by multiple sonar instruments (EdgeTech4200FS, Benthos SIS-1624, etc.) to improve model robustness.
• Develop segmentation models to isolate pipeline or cable structures from the surrounding seabed surface background in the provided images.

Strengths

• Provides 627 pre-labeled images across four distinct marine structure categories.
• Data collection includes multiple sonar instruments, adding variety to the image characteristics.
• Dataset is already divided into training and test sets, facilitating immediate use in model development.

Limitations

• Modest total size of 627 images may be insufficient for training large, deep neural networks without augmentation.
• Class distribution is imbalanced, with pipeline or cable images (323) significantly outnumbering engineering platform images (82).
• Lacks metadata on geographic location, water depth, or specific collection dates for the images.

Provenance

Source

First Institute of Oceanography, author Xing Du.

Collection Method

Images obtained with various side-scan sonar instruments and processed using KNUDSEN's PostSurvey software.