The article addresses the problem of automatic detection, segmentation, and classification of spacecraft using depth maps. The study was based on synthesized semi-natural data, which made it possible to simulate various orbital imaging conditions and to provide accurate ground-truth annotations for the quantitative evaluation of algorithms. Due to the instability of texture and color features under real space conditions, the geometric structure of objects is proposed as a key invariant feature. To solve the detection problem, a multi-stage algorithm has been developed, which includes adaptive histogram equalization, local threshold segmentation, morphological processing, and filtering of false contours based on geometric criteria. To classify spacecraft, deep neural networks of the ResNet family were used in two configurations: based on optical images alone and based on combined data (image + depth map). Experiments on a synthetic dataset of 19440 images showed that the proposed algorithm significantly outperforms the SAM 3 model in terms of detection accuracy (Precision 0.917 vs 0.171), ensuring a low level of false positives. The multimodal approach to classification achieved an accuracy of 99.9% (ResNet-152 model), confirming the effectiveness of using depth maps for reliable recognition of spacecraft under near-real space imaging conditions.