Pointing estimation for spacecraft using Inverse Synthetic Aperture Radar (ISAR) images plays a significant role in space situational awareness and surveillance. However, feature extraction and cross-range scaling of ISAR images create bottlenecks that limit performances of current estimation methods. Especially, the emergence of staring imaging satellites, characterized by complex kinematic behaviors, presents a novel challenge to this task. To address these issues, this article proposes a pointing estimation method based on Convolutional Neural Networks (CNNs) and a numerical optimization algorithm. A satellite’s main axis, which is extracted from ISAR images by a proposed Semantic Axis Region Regression Net (SARRN), is chosen for investigation in this article due to its unique structure. Specifically, considering the kinematic characteristic of the staring satellite, an ISAR imaging model is established to bridge the target pointing and the extracted axes. Based on the imaging model, pointing estimation and cross-range scaling can be described as a maximum likelihood estimation problem, and an iterative optimization algorithm modified by using the strategy of random sampling-consistency check and weighted least squares is proposed to solve this problem. Finally, the pointing of targets and the cross-range scaling factors of ISAR images are obtained. Simulation experiments based on actual satellite orbital parameters verify the effectiveness of the proposed method. This work can improve the performance of satellite reconnaissance warning, while accurate cross-range scaling can provide a basis for subsequent data processes such as 3D reconstruction and attitude estimation.