In lunar sampling missions， the slender and flexible configuration of the robotic arm results in significant nonlinear characteristics and uncertainty in the end-effector’s pose. To support rapid and intelligent sampling operations， visual pose estimation technology is introduced to provide real-time and accurate decision-making support. However， the inherent depth perception limitations of monocular vision， compounded by image degradation caused by extreme lighting variations on the lunar surface， pose severe challenges to the accuracy and robustness of pose estimation. To address the aforementioned issues， this paper proposes the geometrically constrained pose estimation network （GCP-Net）. By integrating a differentiable projection module， the network constructs the reprojection error and the contour algebraic error as additional loss terms， establishing a consistency constraint between image semantics and spatial pose， thereby effectively improving depth estimation accuracy. Simultaneously， to simulate image degradation caused by extreme occlusion and lighting changes in real lunar environments， this paper introduces stochastic noise injection and feature masking enhancement strategies based on GCP-Net， proposing an enhanced geometrically constrained pose estimation network （EGCP-Net）. Simulation experiments demonstrate that the proposed method maintains good robustness even under conditions of partial feature loss. Validated by the publicly reported image data from Chang’e-6， the reprojection results of the pose estimated by the algorithm highly match the observed features， proving the reliability and feasibility of this method in aerospace engineering missions.