3D terrain reconstruction of extraterrestrial bodies is a core element of deep space exploration, providing essential spatial information for landing site selection, rover navigation, and resource exploration. Traditional techniques—such as photogrammetry, photoclinometry, and laser altimetry interpolation—have been extensively applied to the Moon, Mars, and asteroids, achieving significant progress in building high-precision terrain models, interpreting geomorphological features, and supporting resource prospecting. However, these methods remain constrained by limited imaging conditions, the absence of reliable control references, and the complexity of terrain and illumination, often resulting in issues such as low data quality, difficult feature matching, missing observations, and limited automation. In recent years, artificial intelligence (AI) techniques—including convolutional neural networks (CNNs), generative adversarial networks (GANs), attention-based models (Transformers), and neural radiance fields (NeRF)—have shown growing potential in extraterrestrial 3D reconstruction. This review synthesizes three major AI-driven approaches: ①Feature extraction and image matching. ②Depth estimation from single-view images. ③Radiance field modeling from multi-view observations. We further compare their underlying mechanisms, representative applications, applicable scenarios, and performance characteristics. Finally, we outline key technical challenges and discuss future directions in multi-source data fusion, self- and weakly supervised learning, foundation models, and real-time processing, aiming to foster broader applications of AI in extraterrestrial 3D terrain reconstruction.
