PFMGAN: A Generative Adversarial Network with Physics Fusion for DTM Generation from Martian Monocular Images

High-resolution digital terrain models (DTMs) are essential for Martian scientific exploration and engineering missions. However, traditional methods for Martian DTM generation are constrained by limitations in coverage and production efficiency, while the implicit treatment of illumination in existing deep learning methods leads to reconstructions lacking physical interpretability. To address this, we propose a Physics Fusion generative adversarial network based on the Mix Transformer architecture (PFMGAN) that incorporates physical prior knowledge to reconstruct high-precision DTMs from monocular images. This design significantly improves the accuracy and robustness of DTMs by enabling the network to effectively interpret grayscale variations caused by surface albedo and terrain undulations. Experimental results across various Martian landforms demonstrate that, by explicitly inputting solar angles and leveraging an Albedo-Aware Attention (AAA) module, PFMGAN achieves superior accuracy and robustness compared to other baseline models, with up to a 50% improvement in reconstruction accuracy in complex terrains. Furthermore, multi-scale (0.25–6 m/pixel) experimental results indicate that the proposed model is highly adaptable to reconstruction tasks across different spatial scales, consistently delivering high-quality topographic products. The results demonstrate the immense potential of PFMGAN for large-scale, high-precision Martian terrain reconstruction by leveraging the vast archive of monocular imagery.