Xun Geng

Conditional diffusion model with residual decomposition generation strategy for SAR to optical image translation

Construction of Control Network for Multi-temporal LRO NAC Images Based on Matching of Lunar Impact Craters

Abstract. The Lunar South Pole (LSP) has extreme illumination, extensive shadows and weak surface texture, which greatly challenge high-precision mapping and render conventional image matching algorithms ineffective for control network construction. To solve this problem, this paper proposes a multi-temporal LRO NAC image control network construction method based on impact crater matching, leveraging their morphological stability and spatial consistency. A manually annotated 1 m/pixel LRO NAC dataset was used to train YOLOv8 for accurate crater parameter extraction. A virtual feature point matching algorithm was developed, which builds crater descriptors by fusing geometric attributes and neighborhood topology, and achieves reliable matching via multi-attribute similarity and bidirectional verification. Secondary NCC refinement was applied to large craters to improve tie point accuracy. Finally, tie points were mapped back to original images for control network construction and bundle adjustment. Comparative experiments on 94 LRO NAC images near 88.5°S (2298 pairs) with SIFT and SuperPoint showed that the proposed method generated ~10,000 stable tie points per pair, with only 45 pairs lacking valid tie points—far fewer than the comparison algorithms. A large-area control network built from 1,847 images yielded 2,219,604 tie points. After bundle adjustment, the unit weight standard error was 0.67, average image RMS 0.83 pixels, maximum below 2 pixels. The resulting DOM and DEM products had excellent quality. This method effectively solves the LSP tie point matching problem and supports high-precision lunar polar topographic mapping.

Refinement of Asteroid Rotation Parameters via Stereo Intersection Angle Optimization and Masked Feature Matching

Abstract. Determining accurate asteroid rotation parameters is essential for establishing a body-fixed coordinate system during deep-space proximity operations. However, early mission phases often lack prior exterior orientation data. When combined with edge effects from the deep-space background, this deficiency leads to high mismatch rates and ill-conditioned geometries. In this paper, we propose an integrated photogrammetric pipeline to refine these parameters using adaptive feature masking and stereo intersection angle optimization. Rather than using conventional matching, our approach isolates the asteroid target via an adaptive grayscale-threshold mask and morphological refinement, which restricts feature extraction to valid surface textures and significantly reduces outliers. We then introduce a geometric filter to discard stereo pairs with intersection angles below 5°, effectively preventing error propagation along the line of sight. Ultimately, the precise Right Ascension (RA) and Declination (Dec) are determined through an iterative coarse-to-fine grid search that minimizes spatial intersection residuals. Testing our method on 127 images of asteroid (162173) Ryugu from the Hayabusa2 ONC-T camera yielded strong results. The masking strategy successfully cut the number of mismatches in half (from 14,949 to 7,369). Within four iterations, the rotational parameters converged to RA = 96.5° and Dec = -66.4°. These refined results offer a reliable foundation for subsequent 3D reconstruction and high-precision planetary mapping.