DEEP LEARNING UNIVERSAL CRATER DETECTION USING SEGMENT ANYTHING MODEL (SAM)

Impact craters—circular to elliptical depressions formed by meteorite, asteroid, or comet impacts—are key morphological features in planetary science used to infer subsurface composition and structure, study stratigraphy, estimate surface ages via crater size-frequency distributions, and assist navigation and landing site selection. Manual crater mapping is labor-intensive, time-consuming, unscalable, and subject to human bias. Prior automated crater detection algorithms (CDAs) span edge detection and classical ML to modern deep learning (e.g., U-Nets), often trained for specific data types, planetary bodies, and sensor configurations. Consequently, their performance degrades outside the training domain. This work proposes a universal CDA leveraging the Segment Anything Model (SAM), a foundation model for segmentation with strong zero-shot generalization, coupled with geometric shape filtering to robustly identify circular/elliptical crater-like features across diverse planetary datasets without additional training.

Crater detection has evolved from classical image processing to modern deep learning. Early and classical approaches include edge-based and feature methods (e.g., Canny-based pipelines), Adaboost and SVMs, convex grouping, and hybrid supervised–unsupervised schemes. Comprehensive surveys categorize CDAs as image-based versus DEM-based. With deep learning, U-Nets and variants have been trained for crater detection and sizing using Mars Express images, Martian topography, Lunar DEMs, and LROC images; Chang'E missions data have also enabled crater identification and age estimation. However, models trained on specific bodies (e.g., Mars) or sensors often fail to generalize. In segmentation more broadly, deep learning (e.g., DeepLab, encoder–decoder networks, GAN-based segmentation, deconvolutional networks) has become state-of-the-art, and foundation/interactive segmentation models trained on massive datasets (e.g., COCO, LVIS) have emerged. SAM is a recent foundation model trained on SA-1B with billions of masks, demonstrating strong zero-shot generalization across image domains, motivating its use for universal crater detection.

The proposed pipeline comprises three stages: (1) segmentation via SAM, (2) geometric filtering of masks to retain circular/elliptical shapes, and (3) post-processing with edge extraction and ellipse fitting plus de-duplication. 1) Segmentation with SAM: Any imagery (photos, DEMs, false-color products, various resolutions, capture angles, and planetary bodies) is input to SAM. SAM produces multiple outputs per segment: masks, mask areas (in pixels), bounding boxes, mask quality (0–1), input prompts, stability scores (0–1), and image crops. SAM’s default hyperparameters are generally recommended. The segmentation isolates dominant morphological features, enabling downstream shape-based filtering. 2) Circular–elliptical shape indexes and filtering: Because craters are typically circular/elliptical, masks are evaluated using circularity and ellipticity metrics. - Circularity via area–perimeter consistency: For a circular mask with measured area A (pixels), the inferred radius is r = sqrt(A/π), the ideal circumference is d = 2πr = 2π sqrt(A/π), and the actual perimeter P is measured from the mask boundary. The index n = d / P should be close to 1 for a true circle. To mitigate confusion with low-eccentricity ellipses, an ellipse is also fitted to the mask; if the fitted ellipse has major/minor axes a and b, an additional circle-consistency index m based on the fitted ellipse is evaluated, and the mask is classified as a circle if both n and m ≈ 1 ± T, where T is a tunable threshold (default ~0.1–0.5). - Ellipticity via area consistency: An ellipse is fitted to the mask boundary to obtain a and b; the ellipse area is w = πab. The ratio q = w / A should satisfy q ≈ 1 ± T for an ellipse, with T ~0.1–0.5. If a mask is not classified as a circle, ellipticity is then assessed using q. Additional constraints on eccentricity or axis ratio a/b can filter elongated non-crater ellipses. Masks failing circularity/ellipticity checks are discarded. For retained masks, a Canny edge detector is applied to extract boundaries, then circle or ellipse fitting is performed accordingly. The resulting parameters (center coordinates, axes a and b) are stored per detected crater. 3) Post-processing: Filters remove duplicates (e.g., masks sharing similar centers), artifacts (e.g., shadow-induced ellipses), and obvious false positives (e.g., overly elongated shapes). Thresholds for geometric indexes are tuned per dataset characteristics (noise, resolution, clutter).

Across diverse datasets and planetary bodies, the SAM-based CDA effectively detects and sizes craters without additional training: - Mars (HRSC natural color, HiRISE orthoimages, THEMIS IR): The majority of craters are correctly identified and mapped. Small craters missed at coarse scale can be recovered by zooming into subregions, improving detection of finer features. - Moon (LROC images from different angles, including high-oblique; Lunar DEM): The approach detects craters robustly across data types and capture geometries, suggesting potential for real-time use with rover/lander cameras. - Phobos (MRO false-color imagery): Despite the unusual data type and lack of task-specific training, the method identifies most craters with minimal false positives/negatives, demonstrating universality. General observations: SAM segments dominant features reliably in zero-shot fashion; subsequent geometric filtering retains crater-like masks and ellipse fitting yields locations and size/geometry. Resolution limits govern small-crater detection at global scales; focusing on subareas mitigates this. Some false positives arise from circular non-crater features (e.g., central peaks, circular elevated topography) and from shadows; eccentricity and center-based filters reduce, but do not eliminate, such errors.

The research question—whether a universal, data-agnostic crater detector can be built without task-specific training—is addressed by pairing SAM’s zero-shot segmentation with geometric shape filtering. Results across Lunar, Martian, and Phobos datasets, spanning images and DEMs and varied viewing geometries, indicate that the approach generalizes well and can both detect and parameterize crater geometry. This reduces dependence on body- and sensor-specific training data and labeling, a key barrier in existing CDAs. The significance lies in enabling a single pipeline for rapid mapping, potentially supporting terrain-relative navigation and mission planning across bodies. However, the reliance on shape properties introduces characteristic failure modes: circular non-crater features and shadows can be misclassified, while elongated true craters may be excluded if eccentricity thresholds are restrictive. Threshold tuning per dataset is necessary, and occasional unexplained false negatives indicate algorithmic instability requiring further study. Incorporating domain knowledge and transfer learning to adapt SAM to planetary surfaces could reduce false positives and improve sensitivity to diverse crater morphologies.

The study demonstrates a universal crater detection approach that leverages SAM’s zero-shot segmentation and geometric filtering to identify and characterize craters across multiple planetary bodies and data types without additional training. The method successfully maps crater locations and sizes and shows robustness to differing capture geometries, laying groundwork for a single CDA applicable across planetary science. Future work should (i) integrate transfer learning to adapt SAM using diverse, well-labeled planetary datasets; (ii) incorporate additional domain-informed constraints (e.g., photometric cues, topographic context) to reduce false positives from non-crater circular features and shadows; (iii) improve small-crater detection via multiscale or super-resolution strategies; and (iv) systematically assess performance and stability across broader datasets with quantitative metrics.

Primary limitations include: (1) shape-based ambiguity causing false positives from circular/elliptical non-crater features (e.g., central peaks, circular elevated topography, shadowed regions); (2) dependence on tunable geometric thresholds (circularity/ellipticity, eccentricity) that must be adjusted per dataset, with tighter thresholds risking missed elongated true craters and looser thresholds increasing false positives; (3) resolution constraints that hinder small-crater detection at full-scene scales (mitigated by zooming, but not eliminated); (4) occasional unexplained false negatives indicating instability; and (5) potential duplicate detections requiring center-based post-processing that may inadvertently remove true overlapping craters.