月球表面精细三维重建对月球探测任务和科学研究具有重要意义。过去几十年的研究主要侧重于激光测高与摄影测量方法的研究，但受限于硬件和方法原理，其三维重建DEM (Digital Elevation Model)的最高分辨率只能达到几十米或图像分辨率的若干倍，无法为地势复杂的月球南极探测提供有效支撑。影像光度法因其可以充分利用图像中每个像素的明暗信息，从而重建像素级分辨率的DEM受到了广泛关注。但月球南极太阳高度角接近0度，阴影效应严重，使用单张图像无法覆盖所有区域。其次，图像之间光照存在较大差异，直接通过多张图像建立约束方程容易造成冲突，使得优化计算不收敛，无法有效重建三维地形。因此，本文提出了一种基于多张单目影像光度法的像素级三维重建方法，方法的输入为多张高分辨率影像和相应的现有低分辨DEM，通过构建多分辨率金字塔尺度空间，通过影像光度法逐级优化DEM，将原始的低分辨率DEM提升至像素级。在损失函数方面，本文充分考虑到月球南极图像之间的阴影差异，提出了相似性约束，从而保证在引入多光照方向约束的同时，也可有效利用每张图像特有的信息。在损失优化方面，本文引入了自适应学习率优化器，从而更好地保证损失函数的收敛。而优化过程中的梯度更新也为影像光度法的重建质量提供了参考。本文选取了月球南极着陆探测的重点区域—沙克尔顿撞击坑周边的2个局部区域来验证算法的有效性。与公开发布的基于摄影测量方法和激光点云生成的DEM相比，本文所提方法得到的DEM获得了更精细的像素级地形细节，高程精度优于一个像素。基于影像光度法DEM生成的光照模拟影像与遥感影像基本一致，证明实验区地形得到了较好的恢复。本文提出的新算法可以有效重建月球南极大范围高精度、像素级分辨率的DEM，为后续的月球南极探测任务和科学研究提供重要的数据支撑。

Three-dimensional mapping of the lunar surface is of great importance for lunar exploration missions and scientific research. Research in the past decades has mainly focused on laser altimetry and photogrammetry methods to generate precise DEM based on rigorous mathematics calculation. Due to the limitation of the hardware and algorithms, these methods cannot meet the requirements for exploration of the lunar south pole, which requires meter-level resolution to guarantee safe descending and traversing. Photoclinometry has thus received much attention because it can make full use of the images to reconstruct pixel-wise-resolution DEMs from monocular images. However, the lunar south pole suffers from solar elevation angles of nearly 0 degree, and the shadow effect is severe that it is not possible to effectively recover all the terrain details using only a single image. Moreover, the loss functions established directly from multiple images in the photoclinometry algorithm may be inconsistent due to the large differences among the images caused by the varying illuminations. This may cause difficulties in converging the loss and unable to recover the terrain. The paper presents a multiple image photoclinometry method for pixel-wise 3D reconstruction at the lunar south pole. The inputs of the method are multiple high-resolution images and corresponding low-resolution DEM, and a multi-resolution pyramid is used to optimize the DEM step by step. In terms of the loss function, this paper proposes a similarity constraint by fully considering the shadow differences between the lunar south pole images. Rather than optimizing a single DEM, the approach initializes multiple DEMs using the low-resolution DEM associated with each image. While the conventional photoclinometry constraints established on each DEM, an encourage mechanism is then introduced to ensure the pixels cover the same region exhibit consistent elevation. With respect to the loss optimization, an adaptive learning rate optimizer is introduced to guarantee the convergence of the loss function. The gradient update information in the optimization stage could also provide a reference for the photometric quality of the image, which becomes an important index for the following DEM fusion. Two regions are selected around the Shackleton Crater at the lunar south pole for experimental analysis. Regarding the LDEM and photogrammetric DEM as the reference, the geometric accuracy of the DEMs generated from the proposed approach is about one pixel, and the terrain details (e.g., small craters, ridges) presented on the images are fully reconstructed. Rendering simulation is also conducted based on the reconstructed photoclinometry DEM for in-depth evaluation. By analyzing the DEM under various illumination conditions, the simulated images closely resemble the original NAC images, thereby demonstrating the accuracy and precision of the photoclinometry-derived DEM. The method presented in this paper can effectively reconstruct high-resolution and high-precision DEMs from monocular images at the lunar south pole, providing important 3D data support for future exploration missions to the lunar south pole and related scientific research.