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

Impact craters are crucial morphological features in planetary exploration, providing insights into celestial body composition, structure, geological history, and resource potential. They are used in age estimation through crater size-frequency distributions (CSFD) and chronostratigraphy. Manual crater mapping is slow, unscalable, and prone to human error. While semi-automatic and automatic crater detection algorithms (CDA) have been developed, existing machine learning (ML) approaches often rely on data-specific training (e.g., Lunar Reconnaissance Orbiter Camera (LROC) data), limiting their universality. This necessitates a universal CDA capable of handling various data types, sources, and celestial bodies without retraining. This research addresses this need by proposing a novel universal crater detection approach.

Existing CDAs employ diverse methodologies, including convolutional neural networks (CNNs) with Canny edge detection, hybrid supervised-unsupervised machine learning, and Adaboost with support vector machines. U-nets, a type of deep learning architecture, have shown promise in crater detection and size estimation using labeled data from various missions (Mars Express, Chang'E). However, these U-net based methods often lack generalizability, performing reliably only on data similar to their training data. A major limitation of existing ML-based CDAs is their data-specificity and lack of universal applicability across different planetary bodies and data types. This paper aims to address this limitation.

The proposed universal crater detection scheme utilizes the Segment Anything Model (SAM), a foundation model from Meta AI for image segmentation. The methodology involves three steps: 1. **Segmentation:** SAM segments the input image into masks irrespective of data type, celestial body, or resolution. 2. **Shape Filtering:** Shape indexes (circularity and ellipticity) are applied to filter out non-circular/elliptical masks, focusing on crater-like features. Circularity is assessed by comparing the ratio of the calculated circumference to the measured perimeter of the mask. Ellipticity is evaluated by comparing the area of a fitted ellipse to the actual area of the mask. Thresholds for these indexes are used to distinguish between circular and elliptical shapes. 3. **Ellipse Fitting:** A Canny filter is applied to the remaining masks to detect edges, which are then fitted with ellipses. The resulting parameters (axes, center coordinates) define the detected craters. Finally, a post-processing step removes potential duplicates and false positives. SAM provides various outputs including segmentation masks, areas, bounding boxes, and quality scores, all useful for crater detection and size estimation.

The SAM-based CDA is demonstrated across diverse case studies: * **Lunar Data:** The algorithm effectively identified and sized craters in Lunar DEM and LROC images from various angles, even in high-oblique images, suggesting real-time detection potential for rover and lander cameras. Resolution limitations were identified as the main factor in missing the smaller craters. * **Mars Data:** Successful crater detection and measurement were achieved using THEMIS infrared images, HiRISE orthoimages, and HRSC mosaics. * **Phobos Data:** The algorithm performed well on a false-color image from MRO, highlighting SAM's adaptability to diverse data types and planetary bodies. Across all cases, the algorithm demonstrates a remarkable ability to identify and measure craters with minimal false positives. The results show the efficacy of the proposed approach and the adaptability of SAM for universal crater detection across diverse datasets and celestial bodies. The algorithm's performance is primarily determined by the input image resolution. High-resolution images provide better results, while resolution limitations may lead to missing smaller craters. Zooming into specific areas can often improve crater detection in low-resolution images, indicating the method’s potential to detect smaller features in high resolution images.

The SAM-based CDA effectively addresses the need for a universal crater detection algorithm by leveraging the zero-shot generalization capabilities of SAM. This eliminates the need for dataset-specific training, reducing computational costs and development time. The ability to process diverse data types (images, DEMs) and handle varied celestial body characteristics makes it a powerful tool for planetary science. While the algorithm shows great promise, its primary focus on circular/elliptical shapes is both a strength and a limitation, creating potential for both false positives (e.g., misclassifying central crater peaks, shadows) and false negatives (e.g., elongated craters, craters with indistinct features). The need for threshold tuning depending on data characteristics is another factor affecting the accuracy. Despite these limitations, the demonstrated adaptability and generalizability significantly improve upon existing CDA approaches.

This study demonstrates SAM's potential as a universal crater detection algorithm and a pattern recognition tool in planetary science. The method achieves effective crater detection and sizing across diverse datasets and celestial bodies without requiring additional training. Future work could involve fine-tuning SAM via transfer learning using a larger, more diverse dataset to improve accuracy and address the identified limitations, particularly focusing on reducing false positives and negatives by incorporating additional shape or contextual information. This will lead to a more robust universal CDA.

The SAM-based CDA primarily identifies circular/elliptical shapes, which can lead to false positives (e.g., central crater peaks, shadows) and false negatives (e.g., elongated craters). The algorithm's performance is sensitive to image resolution, with lower resolution leading to missed smaller craters. Threshold tuning for the geometrical indexes is dataset dependent, requiring careful calibration for optimal performance. Addressing these limitations through improved shape recognition, incorporation of contextual information, and adaptive threshold determination is crucial for future development.