A system for geographically context-related crater detection

Abstract

A system for geographically context-related crater detection, consisting of: a camera configured to capture raw images of the descent route; one or more integrated sensors configured to provide terrain context data, including altimeter, inertial measurement, or geological mapping information; a context-based super-resolution module (CISR) configured to produce a super-resolution image based on terrain context data; a crater detection network configured to detect one or more craters from the high-resolution image; a graph-based relational matching module configured to match the detected craters with a pre-calculated crater database; and an adaptive state estimation module configured to generate a spacecraft state output based on the matched crater pattern

Description

[0001] The invention relates to the field of imaging of planetary surfaces and automated crater detection, in particular systems that integrate geological contexts with deep learning methods for navigation and terrain analysis.

[0002] Crater detection is a critical component of planetary exploration, particularly for lunar and Martian landers, which rely on surface features for safe landing, navigation, and scientific mapping. Existing optical navigation pipelines primarily rely on visual cues extracted from landing or orbital images. Traditional methods use hand-crafted features such as edges, corners, and template matching to identify crater-like structures. While computationally efficient, these techniques produce poor results when images are acquired under challenging conditions such as poor lighting, high solar angle, motion blur, dust interference, or low resolution at the start of descent. Small craters, essential for precise localization, are often not rendered clearly enough, disrupting continuous navigation updates.Deep learning-based crater detectors, including YOLO-based and region proposal networks, have improved detection accuracy. However, these models treat crater detection as a purely visual classification task, without considering the planetary geological context such as terrain type, surface composition, or expected landmark distributions. As a result, they frequently misidentify shadows, pits, boulders, or sensor artifacts as craters, especially in terrains where similar morphological features are common.

[0003] Their performance deteriorates further when the available images are low-resolution, limiting their ability to reliably detect small or partially obscured craters. Attempts to improve performance by applying super-resolution prior to detection have shown potential, but current super-resolution approaches are optimized for general image enhancement rather than crater-specific feature reconstruction. These methods can generate artificial textures that existing detectors may misinterpret as craters. Furthermore, landmark-matching or constellation-based navigation algorithms, inspired by star-matching techniques, require consistent and accurate crater detection; their performance collapses if any portion of the crater detections is missing or erroneous.Overall, current technology lacks an integrated system capable of simultaneously addressing low-resolution imagery, geological plausibility, and robust crater matching performance. None of the existing technologies combine terrain-based image enhancement, context-aware crater detection, and occlusion-tolerant landmark association in a single pipeline. This leads to significant navigational uncertainties under real-world mission conditions, resulting in increased landing error margins and reduced scientific reliability.

[0004] Therefore, there is a need for an improved system that directly incorporates the geographical context into the detection process, improves images taking craters into account, reduces false alarms, improves the detection of small craters, and provides reliable mapping of craters for navigation and exploration of planets.

[0005] To solve this problem, the present invention offers a cloud-based AI system for seating information in auditoriums.

[0006] The system improves the accuracy of crater detection by incorporating geological and terrain-specific information into the detection process.

[0007] The system is able to function reliably under difficult imaging conditions, including low resolution, extreme lighting, partial occlusion, and noise.

[0008] The system incorporates terrain-based image enhancement or super-resolution to reconstruct fine crater features that are normally lost in low-quality planetary images.

[0009] The system reduces false alarms by filtering out visually similar but geologically implausible features based on context-related and terrain-specific constraints.

[0010] The system improves the detection of small craters, especially those that occupy only a very small area of ​​the image or are only faintly visible in low-contrast areas.

[0011] The system combines detection with graph-based landmark mapping to generate robust crater constellations for navigation and localization.

[0012] The system improves the accuracy of planetary landing and surface navigation by providing reliable, context-verified crater detections to guidance, navigation and control (GNC) systems.

[0013] The system offers a scalable, software-controlled system that can be deployed on orbital spacecraft, landers, autonomous rovers, and remote sensing platforms.

[0014] One embodiment of the present invention is to provide a cloud-based AI auditorium seating information system. The system comprises a context-based image enhancement or super-resolution module that reconstructs high-resolution crater features from low-resolution or degraded input images. Terrain-class embedding or geological metadata is used to condition the enhancement process, enabling the system to restore crater rims, central peaks, and fine morphological structures that are otherwise unresolvable in conventional images. The enhanced images are processed by a deep-learning-based crater detector equipped with a self-calibrating attention mechanism configured to highlight crater-relevant features under varying lighting and terrain conditions.The detector uses a context consistency module that assesses the geological plausibility of each detected crater, thereby reducing false alarms caused by shadows, pits, rocks, or sensor artifacts. The detected craters are then compiled into crater constellation diagrams and compared to reference landmark maps using an occlusion-tolerant relational matching module. This enables robust identification even when some landmarks are only partially visible or obscured due to lighting conditions or environmental factors. In certain embodiments, the system outputs calibrated detection confidence scores that can be integrated into navigation filters or other downstream decision modules to aid in precise landing, localization, and scientific mapping.

[0015] In one embodiment, the invention provides an end-to-end system for geographically contextualized crater detection that generates precise crater-based navigation outputs through the sequential processing of raw data from descent images and contextual sensor data. Raw images acquired by an onboard descent or orbital camera, along with additional inputs such as elevation readings, inertial measurement unit (IMU) data, and terrain classification maps, are forwarded to a context-infused super-resolution (CISR) module. The CISR module reconstructs a high-resolution, terrain-based image by incorporating geological context in the form of terrain class embeddings, thereby highlighting crater rims, shadows, and morphological features that are not normally discernible in low-resolution images. The high-resolution image is then forwarded to a geographically contextualized crater detection network.In this embodiment, the network includes a backbone feature extractor equipped with a self-calibrating attention mechanism (SCAM) that combines rich visual cues with terrain-derived contextual features. A multi-scale neck aggregates feature maps across different spatial resolutions, and an uncertainty-aware detection head outputs crater bounding frames, crater class information, and associated uncertainty scores. During training, the network employs a contextual consistency loss that penalizes detections inconsistent with the local geological context, thereby improving detection robustness. The set of detected craters is subsequently transformed into a crater constellation diagram, where nodes represent individual craters and edges encode spatial relationships.A Graph Attention Network (GAT) generates a relational embedding of the observed graph, which is compared to a pre-computed library of crater graph embeddings stored in an integrated database. The best-matching crater pattern is identified, and if multiple candidates exist, an optional Joint Probabilistic Data Association (JPDA) procedure resolves ambiguities to determine the most probable association. The matching crater pattern and its corresponding weighted measurement are passed to an adaptive extended Kalman filter (EKF). The EKF integrates the crater-based measurements with inertial sensor data and dynamically adjusts the measurement weights according to the detection uncertainty. The filter outputs highly accurate estimates of the spacecraft's position and velocity, suitable for real-time landing guidance, localization, and mission planning.