Physics-informed simulation framework for realistic sonar image generation and statistical validation

Technical Summary: ACOUSIM – A Physics-Informed Simulation Framework for Realistic Sonar Image Generation and Statistical Validation

Problem Statement
The development of robust perception algorithms for underwater applications (e.g., seabed mapping, mine countermeasures, and object detection) is constrained by the scarcity of diverse, high-fidelity real-world sonar datasets. Acquiring real sonar imagery is operationally expensive, logistically complex, and often restricted by security protocols. While controlled tank experiments offer a cost-effective alternative, they fail to capture realistic environmental variability, seabed textures, and complex object-shadow interactions.

Consequently, researchers increasingly rely on synthetic data generated via learning-based models (GANs, diffusion models) or simulation-driven frameworks. However, the validation of these synthetic datasets remains a critical gap. Current evaluation methods predominantly rely on qualitative inspection or indirect downstream performance metrics (e.g., classification accuracy). Standard image quality metrics like PSNR, SSIM, FID, and KID are calibrated for natural optical images and fail to capture sonar-specific intensity and texture statistics. Furthermore, mixing real and synthetic data during training often obscures the intrinsic domain gap. There is a lack of systematic, quantitative methods to assess how closely synthetic sonar data aligns with real data distributions without coupling the analysis to learning-based performance.

Methodology: The ACOUSIM Framework
To address these limitations, the authors present ACOUSIM (ACOustic SIMulation and Validation Platform), a physics-informed framework designed for sonar image generation and standalone statistical validation. The framework operates in three distinct stages:

1. Physics-Inspired Scene Modelling:
   Instead of modeling full acoustic wave propagation, ACOUSIM utilizes a Gazebo-based environment to emulate first-order visual characteristics of sonar imagery. The simulation explicitly controls acquisition-relevant parameters:

   • Platform Altitude: Varied between 10–20 meters.
   • Illumination and Shadowing: Directional light sources are adjusted via roll, pitch, and yaw to simulate object shadowing and highlight regions.
   • Seabed and Objects: Planar seabed surfaces with varied textures simulate heterogeneous backscatter, while 3D object models introduce shadow and highlight regions.
   • Image Formation: Approximated as $I(x, y) = L(x, y) \cdot R(x, y)$, where $L$ encodes illumination/shadowing and $R$ denotes surface reflectance.

2. Noise Modelling and Image Processing:
   To approximate coherent acoustic scattering and background disturbances, the clean synthetic images are degraded using:

   • Additive Gaussian Noise: $N(0, \sigma^2)$.
   • Multiplicative Speckle Noise: $S(x, y)$.
     The final noisy image is formulated as $I_S(x, y) = (I(x, y) + N(0, \sigma^2)) \cdot S(x, y)$.
     Crucially, the framework includes a physics-consistent subsetting step. Since raw synthetic outputs often contain excessive seabed background compared to real datasets (where targets are cropped), synthetic samples are spatially cropped and zoom-adjusted to match the object scale, framing, and viewing characteristics of real sonar imagery.

3. Standalone Statistical Validation:
   The framework evaluates the alignment between real ($P$) and synthetic ($Q$) distributions using global intensity and local texture (Local Binary Patterns - LBP) histograms. Three complementary metrics are employed:

   • Kullback–Leibler (KL) Divergence: Measures directional information loss.
   • Jensen–Shannon (JS) Divergence: Provides a symmetric, bounded measure.
   • Earth Mover's Distance (EMD): Quantifies the minimum transport cost, particularly sensitive to distribution shifts caused by shadow displacement.
     Robustness is assessed via a stratified five-fold analysis.

Key Contributions

• ACOUSIM Platform: A physics-informed, standalone simulation and validation platform that is independent of generative models.
• Statistical Validation Scheme: A novel evaluation methodology based on global intensity and LBP texture distributions, quantified via KL, JS, and EMD metrics, specifically tailored for sonar data.
• Rigorous Evaluation: Systematic evaluation on two public real-world datasets (SeabedObjects-KLSG-II and Sonar Common Target Detection - SCTD) using dataset-level analysis and stratified robustness assessment.

Experimental Results
The framework was tested against the KLSG-II and SCTD datasets, focusing on two object classes: "Plane" and "Ship."

• Texture Alignment: The simulation demonstrated strong local texture alignment across all classes and datasets. KL divergence values for LBP distributions remained consistently below 0.07, and EMD values were below 0.016. This indicates that the framework effectively preserves micro-texture patterns and spatial gradient distributions.
• Intensity Alignment:
  • Plane Class: Showed moderate-to-strong global intensity alignment (KL values of 0.1804 for KLSG-II and 0.2500 for SCTD). Distribution-level visualization revealed substantial overlap in the mid-intensity range.
  • Ship Class: Exhibited substantially higher divergence, particularly in the SCTD dataset (KL = 0.9173), corresponding to low alignment. This is attributed to the elongated hull geometry and complex shadow regions of ship targets, which introduce intensity variations not fully reproduced by the current simulation.
• Robustness: Stratified five-fold analysis confirmed stable global intensity alignment for the plane class but highlighted higher variance for the ship class, reflecting sensitivity to viewpoint and shadow geometry variations.

Significance and Claims
The authors claim that ACOUSIM establishes a reproducible, distribution-level baseline for sim-to-real sonar evaluation. By decoupling synthetic data evaluation from machine learning models, the framework provides a transparent and interpretable benchmark for comparing real versus synthetic sonar images.

The paper positions this work as one of the first systematic efforts to evaluate synthetic sonar realism through direct statistical distribution alignment without relying on learning-based performance metrics. The framework supports the development of reliable datasets for underwater image analysis and lays the groundwork for future physics-grounded digital twins for underwater sensing systems. The authors maintain a modest scope, noting that while the framework handles plane-class targets well, ship-class complexity remains a challenge due to shadow geometry, and future work will focus on extending object and seabed diversity.