Physics-Informed Simulation Framework for Realistic Sonar Image Generation and Statistical Validation

Kamal Basha S, Athira Nambiar · 2026

Read Physics-Informed Simulation Framework for Realistic Sonar Image Generation and Statistical Validation by Kamal Basha S, Athira Nambiar (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-19712.txt.

Introduction

Kamal Basha S and Nambiar present ACOUSIM, a physics-based framework for scoring synthetic sonar images. Synthetic sonar data offer a cheap alternative to real capture, yet lack strict validation. A Gazebo environment builds sonar-like images by setting seabed texture, shadowing, altitude and noise. Realism is scored against two public sonar sets using global intensity and local texture patterns gauged by three divergence gaps. Texture match proves strong, with KL below 0 point 07 across all classes, while plane brightness beats ship brightness, which suffers from shadow complexity.

Catalog record

CREATORKamal Basha S, Athira Nambiar
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20618
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.19712

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-Informed Simulation Framework for Realistic Sonar Image Generation and Statistical Validation Kamal Basha S, Athira Nambiar Abstract Synthetic sonar datasets offer a scalable alternative to costly real-world acquisition, yet their utility remains limited by the absence of rigorous quantitative validation. We present ACOUSIM (ACOustic SIMulation and Validation Platform), a physics-informed framework that evaluates the statistical alignment between synthetic and real sonar imagery without relying on generative models. A Gazebo-based environment generates sonar-like images by explicitly controlling seabed texture, illumination-driven shadowing, platform altitude, and noise. Realism is quantified against two public sonar datasets such as SeabedObjects-KLSG-II and Sonar Common Target Detection (SCTD) using global intensity and local texture (LBP) distributions, assessed via Kullback–Leibler divergence, Jensen–Shannon divergence, and Earth Mover’s Distance. Results show strong texture alignment (KL < < 0.07) across all classes, with plane-class intensity alignment outperforming ship-class due to shadow geometry complexity. ACOUSIM establishes a reproducible, distribution-level baseline for sim-to-real sonar evaluation, directly supporting reliable dataset validation for underwater image analysis. † † address: Department of Computational Intelligence, SRM Institute of Science and Technology, Chennai, India Figure 1 : High-level overview of ACOUSIM . 3D object models are placed in a Gazebo-based underwater environment to generate synthetic sonar-like images, which are statistically validated against real sonar datasets. 1 Introduction Sonar image analysis plays a critical role in ocean and offshore applications such as seabed mapping, underwater inspection, mine countermeasures, and object detection for autonomous and remotely operated underwater systems [ 17 ] . Both civilian and defense sectors rely on high-quality sonar imagery to support safe and reliable underwater operations. However, the development and validation of robust perception algorithms remain constrained by the limited availability of diverse and high-fidelity real-world sonar datasets. The acquisition of real sonar imagery in marine environments is associated with high operational cost, logistical complexity, and strict security restrictions [ 16 ] . Field experiments require specialized vessels, trained personnel, and calibrated equipment, making large-scale and repeatable data collection impractical. Controlled tank experiments, while comparatively cost-effective, provide limited environmental variability and do not adequately capture realistic seabed textures, object–shadow interactions, and range-dependent intensity variations. As a result, publicly available sonar datasets remain limited in both quantity and diversity, restricting systematic evaluation of underwater perception systems To address these limitations, recent studies have explored learning-based sonar image generation techniques, including generative adversarial networks (GANs), diffusion models, and style-transfer approaches [ 11 , 10 ] . In parallel, simulation-driven frameworks such as S3Simulator [ 3 ] , and other virtual underwater sensing platforms have been developed to generate synthetic sonar imagery under controlled environmental and sensor configurations. While these approaches improve visual realism and dataset scalability, their validation is often limited to qualitative inspection or indirect performance gains in downstream machine learning tasks. Quantitative assessment of how closely synthetic images align with real sonar data distributions remains comparatively underexplored. However, evaluation almost universally relies on qualitative inspection or indirect downstream metrics, leaving a fundamental question unanswered: how statistically close is synthetic sonar data to real? Standard image quality metrics PSNR, SSIM, FID, KID [ 15 , 1 ] are calibrated for natural optical images and cannot capture sonar-specific intensity and texture statistics. Mixing real and synthetic data during training further obscures the intrinsic domain gap [ 14 , 8 ] . This paper addresses this gap by presenting ACOUSIM (ACOustic SIMulation and Validation Platform), a physics consistent simulation and validation framework for sonar image generation with a specific focus on statistical distribution alignment rather than learning-based performance metrics. In this work, the term physics-informed refers to explicit control over acquisition-relevant parameters that influence sonar image appearance, including platform (camera) height in the range of 10– 20 m, roll, pitch, and yaw variations of the illumination source to simulate object shadowing, and the inclusion of Gaussian and speckle noise to approximate characteristics of coherent acoustic imaging. Rather than modeling full acoustic wave propagation, these parameters are used to reproduce first-order visual effects observed in real sonar imagery in a controlled and interpretable manner. To the best of our knowledge, this work represents one of the first systematic efforts to evaluate synthetic sonar image realism through direct statistical distribution alignment with real sonar datasets, without coupling the analysis to learning-based performance metrics. Statistical validation is performed using Kullback–Leibler (KL) divergence [ 6 ] , Jensen–Shannon (JS) divergence [ 9 ] , and Earth Mover’s Distance (EMD) [ 12 ] . By decoupling synthetic data evaluation from machine learning models, this work establishes a reproducible and interpretable benchmark for real versus synthetic sonar image comparison. The proposed framework, referred to as ACOUSIM, provides a foundational step toward physics-aware dataset validation and supports future development of digital twins for underwater sensing systems. The main contributions of this work are summarized as follows: • ACOUSIM : a physics-informed sonar simulation and standalone statistical validation platform, independent of generative models. • Statistical validation scheme based on global intensity and LBP texture distributions, quantified with KL, JS, and EMD metrics. • Evaluation on two real sonar datasets (KLSG-II, SCTD) using dataset-level analysis and stratified five-fold robustness assessment. 2 Related Work The development of sonar image datasets for underwater perception and analysis has been actively explored through three primary directions: (i) Acquisition of real sonar data, (ii) Learning-based sonar image synthesis, and (iii) Simulation-driven synthetic data generation. Real sonar datasets are commonly acquired using side-scan sonar (SSS) systems deployed in controlled underwater surveys and seabed inspection missions. Existing datasets include shipwreck segmentation benchmarks [ 13 ] , semisynthetic training datasets for underwater object classification [ 2 ] , and Sonar Common Target Detection Dataset [ 18 ] . These datasets provide annotated sonar imagery for tasks such as target detection, classification, and segmentation, with validation typically performed using expert-labelled ground truth and downstream perception benchmarks. While real-data acquisition provides physically authentic sonar measurements, large-scale collection and annotation remain expensive, time-consuming, and operationally constrained. Learning-based sonar synthesis approaches attempt to generate realistic sonar imagery from limited real data using generative models. Existing methods include CycleGAN-based sonar translation frameworks [ 5 ] , hybrid ray-tracing with GAN refinement approaches [ 7 ] , multi-view GAN-based sonar generation [ 11 ] , and diffusion-model-based sonar enhancement frameworks [ 10 ] . These approaches have demonstrated improved visual realism and promising downstream recognition performance for tasks such as classification, detection, and segmentation. Evaluation is typically performed using perceptual similarity metrics such as FID and KID together with downstream task-level accuracy measures. Simulation-driven approaches focus on reproducing sonar image characteristics through analytical, geometric, or physics-inspired acoustic modelling. Unreal Engine-based underwater sonar simulators [ 14 ] , ray-acoustic modelling frameworks [ 8 ] , and S3Simulator [ 3 ] generate synthetic sonar imagery under controlled environmental and sensor configurations. Validation is commonly performed using target recognition benchmarks, shadow consistency analysis, or limited statistical comparisons against real sonar imagery. Despite these advances, validation of synthetic sonar datasets across existing literature remains predominantly indirect, relying on perceptual quality metrics or downstream task performance. Systematic validation based on direct statistical distribution alignment between real and synthetic sonar image domains remains comparatively underexplored. 3 Methodology ACOUSIM consists of three stages (Fig. 1 ): (i) physics-inspired scene modelling, (ii) synthetic image generation with noise, and (iii) standalone

Read the complete text