In June 2024, when Chang’E-6 returned samples from the lunar farside, the mission marked a technological milestone in human lunar exploration. In such farside sampling missions, trajectory and landing location are not merely engineering parameters. They serve as essential prerequisites for mission verification, navigation assessment, and ultimately the geological interpretation of returned samples. Yet for a mission like Chang’E-6, which faced sparse data due to communication constraints and required highly time-sensitive and autonomous processing, the limitations in ground tracking made high-precision reconstruction and remote geological analysis extremely challenging.
Our team, aerospace scientists from Beihang University and mineralogists from Peking University, joined forces to tackle this challenge(Shu et al., 2025). We realized that the descent video, containing only a few minutes of grayscale images, was the only visual record of the lander in motion. Could the imagery itself “speak” and reveal the lander’s trajectory, precisely localize the landing site, and provide insights into the geological characteristics of the region?
Automation and generalization, not mission-specific “black magic”
Two major obstacles confronted us at the outset: 1) The descent imagery contains limited geometric cues and significant degradation. 2) Available data is sparse, making task-specific modeling or reconstruction infeasible.
We therefore focused on creating a system that relies primarily on imagery, with an emphasis on automation and adaptability across missions This meant the system needed to operate under feature-sparse lunar images, avoid any mission-specific priors, and still achieve meter-level accuracy at engineering-ready standards. This challenge was daunting, yet exhilarating.
Chang’E-6 provided a unique opportunity for real-mission validation
Our core method was to establish a vision-based landing site localization pipeline. This system does not require digital elevation models of the landing area but instead treats the descent video and remote sensing images of the landing region as input sequences. By leveraging a deep learning-based matching algorithm to automatically detect and track reliable visual features across consecutive frames, and tightly integrating these measurements with known camera parameters and lander dynamic models, the algorithm is able to estimate the position and orientation of the lander in each frame. This continuous stream of poses constitutes the reconstructed trajectory. Ultimately, through the matching of descent images with remote sensing imagery of the landing zone, the landing site is projected into the lunar coordinate system. This entire process is fully constrained and optimized by the intrinsic geometric properties of the imagery and the dynamic characteristics of the landing phase.
To validate the robustness and generalizability of our approach, we first evaluated the system on Chang’E-3, -4, and -5. The consistent performance across missions gave us confidence. When the framework successfully reconstructed a high-fidelity trajectory directly from the Chang’E-6 descent images and localized the touchdown point with striking accuracy, we realized that the images themselves contained all the necessary geometric information, and the key was to enable the system to read it intelligently. This not only offered a new way to interpret the Chang’E-6 landing process but also demonstrated that high-precision temporal-spatial information can be extracted from imagery alone, without heavy reliance on mission-specific datasets.
From engineering validation to scientific discovery
Subsequently, we found that the framework also enabled significant geological discovery. Once the descent images were accurately anchored to the lunar surface through our reconstructed trajectory and camera geometry, the sequence transformed into a unique low-altitude, high-resolution geological survey data which is impossible to obtain from any orbital instrument.
With physically corrected camera geometry, the descent images provided a distortion-free basis for rigorous crater morphometry. This was essential because in the raw images, severe perspective distortion compressed circular craters into ellipses, rendering crater size-frequency measurements unusable. After correction, however, we could reliably restore true crater shapes, enabling meaningful CSFD analyses directly from descent imaging. This allowed us to determine an absolute model age of ~ 855 Ma for a local 1.24 km2 patch near the Chang’E-6 landing site which is an age dramatically younger than the canonical 2.4-3.3 Ga for the regional Imbrian basalts. These results, together with the ~454 Ma age we obtained using physically corrected Chang’E-5 descent imagery, highlight how indispensable our correction pipeline is for reliable localized geological dating.
More importantly, the young model ages support the interpretation that the farside experienced resurfacing events far later than previously thought. When placed alongside the recently discovered ~123 Ma volcanic glass beads from Chang’E-5 samples (Wang et al., 2024), our findings suggest that both crater populations and lunar soil components preserve records of late-stage volcanism and localized resurfacing that global orbital datasets had missed, largely because earlier studies relied on coarse-resolution images or kilometer-scale craters that obscure small-scale resurfacing signatures. The descent-based observations therefore reopen the question of how long farside volcanism persisted and challenge the long-held view that farside magmatism waned much earlier than nearside activity.
Beyond chronology, the corrected descent images revealed mineralogical and lithological clues that point to deeper crustal and possibly upper mantle materials exposed near the landing site. The lander touched down beside a ~3.5 m crater whose ejecta included angular, high-reflectance clasts measuring 1.2 to 5.7 cm across. Their distribution along a clear ejecta trajectory, combined with their bright photometric properties, suggests origins in plagioclase-rich lithologies with distinct Ca-Al signatures. At the particle scale, the descent imagery captured regolith fragments up to ~0.86 mm which significantly coarser than lunar soil distributions measured in Chang’E-5 samples. Such coarse grains are indicative of more resistant minerals like pyroxene and olivine, consistent with minerals derived from the lower crust or even the upper mantle. When integrated with our geomorphological mapping, these observations support the hypothesis that the Apollo basin impact may have penetrated the farside’s thinner crust, excavating deeper materials that remain exposed in the Chang’E-6 landing area.
These geological insights, derived from the corrected imagery, demonstrate that a descent sequence can serve as a unique scientific dataset. Quantitative crater morphometry, refined surface dating, evidence of late resurfacing, and preliminary mineralogical interpretation, all derived entirely from just a few seconds of corrected descent imaging, made it clear to us that a descent sequence is not merely telemetry. It is a scientific observation in motion, capturing a close‑range perspective and spatial resolution that no orbital dataset or landed panorama can replace.
Looking back
This work demonstrates, to our knowledge, a novel integration of automated, vision-based trajectory reconstruction with subsequent geological discovery on a planetary surface. What was once treated as auxiliary navigation data can now serve as a primary source for trajectory reconstruction, an independent means of landing-site localization, and a powerful tool for high-resolution geological analysis of landing zones.
We hope this study not only sheds new light on the Chang’E missions but also provides a generalizable and automated vision framework for future landings on the Moon, asteroids, and Mars. With the increasing frequency of planetary missions, the ability to rapidly and autonomously identify and select landing sites will become even more crucial. Such an automated pipeline can be extended to analyze multispectral or hyperspectral descent imagery to extract additional mineralogical and geological information, providing a richer picture of planetary surfaces without solely relying on costly in situ sampling.
Furthermore, this paradigm can be applied to missions targeting planets or moons where direct landing or sample return is currently impractical, such as Venus or Icy Moons, as well as to early-stage automated exploration of asteroids for mineral resource assessment. By demonstrating that descent imagery can simultaneously support trajectory reconstruction, landing-site localization, and geological discovery, this work points toward a promising avenue for more efficient, automated, and scientifically rich planetary exploration.
References
Shu S, Lin L, Hou B, et al. Intelligent vision-guided trajectory reconstruction enables rapid localization and characterization of the Chang’E-6 landing site. Communications Earth & Environment, 2025.
Wang, B. W., Zhang, Q. W., Chen, Y., Zhao, W., Liu, Y., Tang, G. Q., ... & Li, Q. L. Returned samples indicate volcanism on the Moon 120 million years ago. Science, 2024, 385(6713), 1077-1080.