ApertureLab Open the appsign-in required

Teach a model the seafloor before it ever goes to sea.

ApertureLab is a synthetic aperture sonar simulation and beamforming workbench. It turns a scene description into focused sonar imagery with pixel-accurate ground-truth labels on a single workstation, as training data for models of the ocean floor.

Run the beamformer in your browser

A back-projection beamformer runs in the page on 564 pings over a 9.4 m track; the aperture slider sets the length of the synthetic aperture, and the scene comes into focus as it grows.

Why synthetic aperture? · Everything else

↓
Beamformed sonar image showing a dense multi-object scene across mixed seafloor zones Hover for labels

Applications of labeled synthetic sonar imagery

Foundation models for synthetic aperture radar (SAR) exist because satellite imagery is inexpensive, abundant, and labeled. Sonar imagery requires ships, dive operations, and often restricted access, so labeled sonar imagery is scarce. ApertureLab generates it by physics-based simulation, which is the precondition for the three applications below.

Synthetic Datasets

Labeled imagery, by construction

Every simulated image carries ground-truth labels derived automatically from the scene file: object class, position, orientation, burial depth, seafloor type, shadow mask. The six annotation categories are written with the image in COCO format, together with the pixel size and scene extent needed for a georeference.

  • Oriented object footprints from real geometry and heading
  • Buffered polylines for cables and pipeline runs
  • Traced outlines for every seafloor type in the scene
  • Domain variation in depth, range, bottom type, aspect, and burial

Because a scene is a file, a dataset is a parameter sweep over scene files. The next thousand images cost a thousand times the first one, and none of them requires ship time.

Foundation Models

Pretraining data for SAS foundation models

Pretraining architectures such as ViT, MAE, and contrastive learning are mature; the input they lack for synthetic aperture sonar (SAS) is a training corpus of sufficient size. ApertureLab produces that corpus from parametric scenes that vary depth, range, bottom type, aspect, and burial.

  • Sim-to-real transfer for autonomous underwater vehicle perception
  • Environment-agnostic feature representations across seabed types
  • Edge-case and rare-target coverage unavailable in field collections
  • The sonar analog to ImageNet-scale pretraining

The pretraining methods established on camera and SAR imagery carry over to SAS once that corpus exists.

Vision-Language Models

Vision-language models on SAS imagery

The 2026 IGARSS paper reports that vision-language models (VLMs) classify SAS targets at 0.946 AUC with no domain-specific training, using only a text prompt that describes highlight-shadow geometry. Captions generated at render time supply the data for fine-tuning; every statement a caption makes about its scene is true by construction.

  • Auto-generated natural-language captions paired with every image
  • Query sonar archives by English phrase
  • Caption imagery for non-expert operators and analysts
  • A path to zero-shot object recognition across unseen target types

Domain fine-tuning on captioned simulated imagery is the next step from the 0.946 AUC zero-shot baseline.

Evidence

A labeled dataset from one English sentence

The entire specification, written through Claude Code “create a 60 by 100 m scene with a grid of different seafloor backgrounds and a target in the middle of each one, so I can crop 256 by 256 chips of the same object on different bottoms for ML training.”
ApertureLab grid scene: sixty seafloor targets laid out across ten varied bottom types in a 6-by-10 grid of distinct sonar seafloor textures. Hover for labels  ·  click to enlarge

The resulting scene, simulated and beamformed overnight on one workstation, holds sixty seafloor targets on ten bottom classes: clean sand, mud, rock, gravel, coarse gravel, silt, rippled sand, two rocky-sediment variants, and shelly sand hash. Roughness and reflectivity are jittered per cell, so no two cells match. Every green box, name, and seafloor zone outline shown on hover is generated from the scene file at simulation time rather than annotated by hand.

The image and its labels are computed from the same scene description, so each annotation matches the simulated geometry exactly.

Evidence

Simulated image and automatic labels

The figure shows one scene twice: on the left, the simulated sonar image; on the right, the object footprints and seafloor-type boundaries that accompany it. No label was annotated by hand. The labels are written from the scene file at simulation time, so they are exact by construction.

A simulated synthetic aperture sonar scene beside the same scene overlaid with object bounding boxes and traced seafloor-zone outlines
Object class, position, orientation, burial depth, and the traced outline of each seafloor type, written out with the image in COCO format.

Walkthrough

Scene construction from an empty window to a labeled image

The recording, sped up, follows one build from an empty window to a focused SAS image covering 100 m in range by 60 m along track, with a pixel-accurate label on every object. The pages listed below describe each stage of the process.

Explore

Where to go deeper.

Each page below walks through one capability and ends in the imagery it produces.

The physics

Why synthetic apertureFrom one transducer element to a synthetic aperture, in seven steps; beamform a scene in your browser at the end. Why baseband?How a sonar echo on a 300 kHz carrier becomes complex baseband samples without losing anything, and why back-projection must restore the carrier phase; nine interactive steps ending with images from ApertureLab's beamformer. The surface of focusWhere a synthetic aperture sonar image is in focus, and why objects standing above the seabed are displaced and blurred; an interactive scene with images from ApertureLab's simulator and beamformer. The plane of focus in circular SASOn a circular track the depth plane is a true plane of focus, and anything off it is drawn as a ring; an interactive scene with images from ApertureLab's simulator and beamformer. Hardware faultsWhat a tone, a weak receiver, band-limited noise and ping-to-ping phase errors do to the raw data and the image; an interactive page with images from ApertureLab's simulator and beamformer. From bathymetry to SASFour planned survey lines flown over real deep-water bathymetry, with the simulated sonar images placed on the map and measured readouts of what each line sees and misses. Kirchhoff facetsObjects simulated as point scatterers and as Kirchhoff triangular facets on one scene, why faceted objects image as outlines and glints, and the roughness models that extend the facet approach. From a Blender scene to SASAn authored Blender scene and the physics-based synthetic aperture sonar image ApertureLab forms from it on the same 100 meter patch of seabed; click the figure to zoom into the sonar. SimulationTwo independent physics engines, one interchangeable output, each checking the other against a published paper. Image formationRaw returns are not a picture. Back-projection, micronavigation, and what turns echoes into a focused scene.

The data layer

Scenes and seafloorsA seafloor authored as a node graph; deterministic, reseedable, and free of any rectangular footprint. LabelsEvery image ships with pixel-accurate ground truth, emitted from the scene file in COCO format. DatasetsTwo releases of 1000 labeled HISAS-class sonar swaths on Hugging Face; v1.1 adds terrain height maps, boulder tables and georeferenced images.

The tools

The workbenchOne codebase, desktop and browser. Author a scene, run the physics, inspect the image. StudioDescribe a scene in English; an agent builds it, runs the simulator, and shows you the sonar image. MCP ServerConnect Claude, GPT or a local model to ApertureLab on your own workstation through the Model Context Protocol; the model authors scenes and the simulator runs on your GPU. CloudThe same scene file, one GPU instance per scene, so a batch finishes in about the time of one.

The record

FidelityResolution, sidelobes, spectra, speckle statistics and phase-center delays of the simulator and beamformer, measured against prediction. PerformanceHow long one image takes, end to end, against scene density, and against the time the sonar spent collecting it. Research29 papers across 14 years, mapped onto the stage of the processing chain each one advances. GalleryWhat the seafloor looks like through ApertureLab. Object galleryEvery object and seafloor type the simulator can place, each imaged in synthetic aperture sonar, with its geometry and true physical size.

Isaac D. Gerg, Ph.D.

Rothrock Research LLC · gergltd.com

Underwater machine learning depends on two disciplines that are rarely held by one team: the acoustics of the sonar and the deep learning applied to its imagery. My research over twenty years has covered the complete signal chain. It runs from acoustic wave scattering and IQ time-series recording, through beamforming and image formation, to deep embeddings for detection, segmentation, and compression.

ApertureLab applies that experience in software, from the scattering model to the labeled image. I welcome conversations with groups building AI systems that perceive the physical world.

29 peer-reviewed papers across 14 years, spanning the same chain · isaac.gerg@gergltd.com