A research team from the Aerospace Information Research Institute of the Chinese Academy of Sciences (AIRCAS), together with its collaborators, has developed a novel synthetic aperture radar (SAR) interpretation method that identifies the specific scattering sources on three-dimensional (3D) targets responsible for prominent radar reflections visible in SAR images.

By integrating a customized differentiable SAR simulator with inverse optimization, the method correlates bright scattering features in two-dimensional (2D) SAR imagery with precise 3D structural characteristics. This advancement strengthens target analysis, enhances image interpretation capabilities, and supports physically grounded SAR simulation for remote sensing applications.

The findings were recently published in Journal of Remote Sensing.

SAR technology is widely recognized for its ability to capture high-resolution imagery under all weather conditions and at any time of day. Despite these advantages, SAR images remain challenging to interpret from a physical perspective: intricate 3D structures are compressed into 2D signatures, which frequently appear as intense scattering signals such as bright spots or linear features. Conventional interpretation methods are often labor-intensive, computationally expensive, or constrained in terms of physical interpretability and general applicability.

Differentiable rendering has opened new avenues in this field. Even so, SAR research still requires forward models capable of accurately simulating shadowing effects and complex scattering behavior to achieve high-fidelity results.

To address this technical bottleneck, the research team investigated approaches to establish direct correlations between bright scattering signatures in 2D radar images and their corresponding real-world 3D structures. The team subsequently created the novel method, which maps strong scattering information directly to 3D target geometry via a tailored differentiable SAR simulator.

The core innovation of the approach lies in constructing what the researchers define as prominent scattering regions to explicitly visualize this mapping process. Rather than relying on manual annotation or pure black-box fitting, the simulator adopts physically meaningful operators. By adjusting the scattering intensity attributes of target vertices, it quantifies the contribution of different surface positions to scattering signals observed from a specific viewing angle. The framework has been validated on both simple radar reflector scenes and complex T72 tank models.

The simulator operates through two interconnected stages. In the forward simulation stage, it transforms the target into the coordinate system of the SAR sensor, calculates facet-level scattering geometry, applies shadow constraints to determine surface visibility, and aggregates scattering data to generate a simulated SAR image. In the inverse mapping stage, the simulated image is compared with ground-truth SAR datasets. Backpropagation and gradient descent algorithms are then applied to update the scattering intensity attributes of target vertices, while maintaining the original geometric structure of the target.

To preserve full differentiability throughout the simulation process, the team adopted the Blinn–Phong model for scattering approximation, replacing more complex non-differentiable electromagnetic formulas. The ground-truth SAR image data were generated using RaySAR across 24 viewing angles, with a resolution of 124 × 124 pixels and a 0.2-meter range and azimuth sampling interval.

This research demonstrates that SAR interpretation can evolve beyond basic 2D target recognition to pinpoint the 3D structures generating dominant scattering signals. Accordingly, the proposed method not only improves simulation fidelity but also delivers a more physically interpretable analytical framework.