To meet the demands of exploration and scientific research on extraterrestrial bodies, acquiring crater detection information and cataloged data is essential as the foundation for measurement and analysis. Although remote sensing payloads capture data containing craters, these images feature a large number of targets, high target density, and diverse morphologies. Detecting craters in collected image data is essential for further scientific research, measurement, and localization on celestial surfaces. The discovery of new craters and the establishment of a complete crater cataloging database are prerequisites for the study of craters and related downstream tasks. Existing detected and cataloged celestial crater datasets cover only parts of the Moon and Martian surfaces. With the advancement of future deep-space exploration, artificial intelligence methodologies are anticipated to supersede conventional manual identification approaches and emerge as the predominant method for celestial crater detection. However, AI-driven approaches depend on well-annotated training datasets, and manual annotation poses inherent challenges such as high labor intensity and technical complexity. This imperative has propelled simulation-based methodologies for generating synthetic remote sensing imagery of craters into a critical research focus for constructing training datasets.

In this paper, we propose an efficient crater image simulation method based on neural radiance field (NeRF). The method integrates image simulation technology, embedding the target 3D model into the process to simulate and render crater images with varied morphologies and illumination conditions. It implicitly captures the geometric structure of the target in the model and generates images that approximate the physical imaging process by incorporating parameters such as imaging distance, angle, and light angle. The method consists of two parts: NeRF-based image generation and image-harmonization-based target fusion. First, the 3D model is used to construct the crater data and train the NeRF model for crater image generation. Then, the imaging parameters of the remote sensing image are fed into the NeRF network to generate crater images. Next, the feature domain difference between the crater target and the background is adjusted by constructing and training an image harmonization network in combination with the NeRF algorithm. This approach compensates for fringe distortions between the target and background while simultaneously producing positional labels for the crater. The proposed method fulfills simulation requirements according to specific crater morphological types and actual imaging conditions, ensuring both scientific accuracy in crater representation and visual consistency with realistic celestial surface environments.

The simulated crater largely matches the geometric structure of actual craters, though some differences remain compared with real images (Fig. 6). The harmonized crater aligns more closely with the background’s feature domain while preserving its geometric structure. In this paper, the effectiveness of the proposed simulation method is verified by the target detection algorithm. Using our annotated lunar crater dataset, experiments show that the proposed method, leveraging small-scale real image datasets, produces effective, controllable training data for multiple crater detection approaches. As shown in Tables 1?3, adding crater images generated by our method improves detection metrics in every group, with a maximum gain of 27.2% and an average F1-score improvement of 11.3%. It is demonstrated that the crater simulation method can provide augmented data for many types of target detection algorithms, improve target detection accuracy, and remain applicable to different datasets. Moreover, our simulated images outperform those generated by three other mainstream image simulation algorithms. The results substantiate that our approach effectively supports the training of deep learning-based crater detection methods.

To address the bottleneck of insufficient training data for deep-learning-based crater detection algorithms, we propose a remote sensing image simulation method of craters with neural radiance fields. The method integrates image fusion and harmonization techniques, coupling a crater 3D model with the image simulation algorithm, allowing control over target illumination and imaging conditions, and generating crater simulation data under complex and varied scenarios. Experimental results show that the simulated images expand training datasets, improve detection accuracy, and support high-precision mapping of celestial topography. This simulation method can also be used for simulated image generation in remote sensing detection-related fields, providing diverse training data and annotations for multiple target types, meeting data volume requirements for high-precision detection, recognition, semantic segmentation, and related tasks.