Abstract: 3D information extraction from satellite imagery is a key focus in geographic information science. It involves mining implicit similarity, geometric, photometric, and temporal cues from single-view, stereo, and multi-view images for accurate surface reconstruction. This review covers four information categories: similarity, geometry, photometry, and temporal data. Emphasis is placed on similarity-based extraction, particularly stereo matching, including traditional and deep learning methods for stereo imagery, and deep learning approaches for multi-view imagery, highlighting strategies for improving reconstruction accuracy in complex terrain. For geometric information, methods based on geometric and higher-order geometric constraints are systematically discussed, focusing on robustness and application contexts. In photometric modeling, the integration of shading and silhouette cues is examined, with emphasis on neural radiance fields (NeRF) for enhancing detail and realism. Temporal information extraction analyzes dynamic scenes and change detection using multi-temporal images. By addressing current limitations and future trends, this study proposes cross-domain data integration, dual-task frameworks combining object detection, and efficient Pipeline design, offering systematic theoretical and technical insights for satellite-based 3D reconstruction.

Key words: satellite imagery, 3D information, stereo matching, deep learning, neural radiance fields

摘要： 卫星影像的三维信息提取技术是地理信息科学的重要研究方向，其核心在于挖掘单视图、双视图及多视图中隐含的相似性、几何性、光影模型信息与时序信息，以实现精准的地表三维重建。围绕四大信息类别展开综述：相似性信息、几何信息、光影模型信息以及时序信息。重点探讨的是相似性信息提取，在该技术上聚焦于立体匹配技术，涵盖双视影像的传统方法与深度学习方法和多视影像深度学习方法总结，阐明其在复杂地形中提升三维重建精度的策略。针对几何信息的提取，系统讨论了基于几何与高阶几何形式的三维重建方法，重点解析基础几何推理的鲁棒性与应用场景。在光影模型信息中，探讨光影反演与形状剪影的结合应用，重点介绍了神经辐射场（NeRF）在细节重建与真实性提升中的研究进展。时序信息提取部分结合多时序影像，剖析动态场景中三维信息的推断与变化捕捉。通过分析现有技术的瓶颈与未来发展趋势，提出了跨领域数据整合、融合目标检测的双任务协同架构及高效Pipeline设计等研究方向，为卫星影像的三维重建领域提供了系统性的理论参考和技术展望。

关键词: 卫星影像, 三维信息, 立体匹配, 深度学习, 神经辐射场