idg101/Aperture_Lab_Synthetic_Aperture_Sonar_v1

ApertureLab Synthetic SAS Dataset

Version 1.0 (September 2026). Author: Isaac Gerg. Made with ApertureLab; samples, statistics and the generation pipeline are described on the dataset page.

1000 simulated synthetic aperture sonar (SAS) images, each an 80 m along-track by 200 m range swath from a HISAS 1030-class 100 kHz sonar on a straight track, beamformed by time-domain back-projection at 2.5 cm pixels and delivered as dynamic-range-compressed (DRC) TIFF LZW images with COCO polygon annotations.

The seafloor is procedurally varied per scene (sand, silt, mud, gravel and rock patchworks, ripple fields with per-scene wavelength and direction, rocky outcrops, pockmarks, trawl scars, fractal relief). Each scene holds zero to six labelled targets, exactly 1000 truncated cones and 1000 wedges in total, plus unlabeled-class clutter: twelve parametric shapes (pyramid, cone, frustum, dome, drum, slab, hexagonal prism, capsule, cross, random polyhedron, sediment mound, torus) and ten mesh objects (tire, oil barrel, keg, anchor, fish trap, ladder, debris, propeller, pallet, rock pile). Every object is seated on the local seabed with substrate-coupled burial and scour.

Files

Path	Contents
images/si_XXXX_*_tdbp_drc.tif	DRC image, 8000 x 3200 px, TIFF LZW, along-track up, range right
images/si_XXXX_*_tdbp_coco.json	COCO polygons for that image, with world extents and pixel size
coco_all.json	merged COCO over the whole set
manifest.csv	one row per scene: split, seafloor family, clutter tier, counts
objects.csv	one row per object: role, class or shape, position, heading, size, substrate, burial, grazing angle
scenes/si_XXXX.yaml	the full simulator scene description, editable and re-runnable with ApertureLab

Splits are a column in manifest.csv: indices 0-799 train, 800-899 validation, 900-999 test. The pockmarked seafloor family appears only in the test split as a small out-of-distribution check.

Dataset viewer

The viewer shows the two tabular summaries, named explicitly in the configs: block of this card so that they are the only files it reads. The scenes subset is manifest.csv, one row per scene, and it is the default. The objects subset is objects.csv, one row per placed object. Each is presented as a single split called full, because the train, validation and test assignment is a column of manifest.csv rather than a separate file.

The COCO annotations are not a tabular split. They are per-image JSON files stored beside the images in images/, one file per image, together with the merged coco_all.json at the repository root. Read them with a COCO loader or with json.load, and read the images directly from images/.

Counts

Images	1000 (800 train, 100 validation, 100 test), 8000 x 3200 px, 2.5 cm pixels
Seafloor families	uniform 331, patchwork 288, ripple 219, rocky 153, pockmarked 9 (test only)
Targets	2000: 1000 truncated cones, 1000 wedges; 150 scenes hold none
Clutter objects	2479 across 22 classes (10 meshes, 12 parametric shapes), in tiers L0 (0-1), L1 (1-4), L2 (4-8) per scene
COCO annotations	5503 over 27 categories (targets, clutter, trawl scars, pockmarks)
Target substrates	sand 815, gravel 445, mud 300, silt 283, rock 157
Target burial	median 0.11 of object height; 249 past 0.33; max 0.70 (mud)
Grazing angle	6 to 53 degrees

Known issues in v1.0

• Burned-in information box. The scene runner rendered an ApertureLab information box into the lower-right corner of every image (about 1280 x 440 px, the far range at the start of the track). Labels under it are still correct. It is not present in the scene files, so a re-render with the box disabled produces clean images.
• Scatterer density. The images were simulated at 10,000 scatterers per square metre (the _simtype=ps_pm2=10000 tag in each file name). The scene files carry the batch default and were corrected to 10,000 after the first upload; a re-render from a scene file reproduces the image geometry and labels exactly and the speckle statistically.
• Mixed-type heading_deg in the COCO attributes. In every COCO file, an annotation's attributes.heading_deg is a number for a wedge and an empty string for a truncated cone, which is axisymmetric and carries no heading. A reader that infers one type per column, such as Apache Arrow, rejects the file on that change. The workaround is to treat an empty string as null when parsing. The same value is an empty cell in objects.csv, where it is harmless because CSV is untyped. Version 1.1 writes JSON null instead.

License

Distributed under the ApertureLab Synthetic SAS Dataset License 1.0 (see LICENSE.md): free for any use, commercial or not, by anyone except the parties named in its Exclusion clause, with attribution and the requirement that redistributions carry the same exclusion. Please cite:

Gerg, I. (2026). ApertureLab Synthetic SAS Dataset, version 1.0: 1000 HISAS-class swaths with cone and wedge targets. https://huggingface.co/datasets/idg101/Aperture_Lab_Synthetic_Aperture_Sonar_v1