Automatic classification of Martian impact craters using deep learning: a new tool to improve planetary surface dating

 

Introduction 

Impact craters are essential markers for reconstructing the geological history of planetary surfaces [1]. On Mars, where no absolute radiometric dating has yet been conducted in-situ, the density of craters remains the main chronometer used for dating surface units [2, 3]. However, this method critically depends on the correct identification of primary craters, as secondary craters (formed by ejecta from a primary impact) and ghost craters (highly degraded

or buried) must be excluded to avoid significant overestimations of surface ages [4]. As the identification of crater morphological features is still a long, repetitive, and subjective task when performed manually, the application of modern computer vision techniques has become more and more relevant. While automated crater detection has seen substantial progress in recent years thanks to deep learning and computer vision techniques [5, 6, 7], the classification of craters based on their morphology remains largely unexplored. Yet, such classification is essential to ensure both the validity of crater inventories and the robustness of derived age estimates.

 

Dataset and Preprocessing 

To train our classifier, we relied on the comprehensive work of Lagain et al. (2021) [4], which provides a manually annotated catalogue of more than 376,000 craters with a size superior at 1km in diameter into four morphological classes: Regular, Secondary, Ghost, and Layered. Image patches centered on each crater are extracted from the global CTX mosaic [8], after reprojection in local stereographic coordinates to preserve the circular geometry of craters at high latitudes. To ensure robustness, we refine the crater locations and sizes using a circle detection algorithm based on the Hough transform [9]. This preprocessing step significantly improves the alignment between craters and image content, a critical requirement for effective supervised learning. In order to train our model, we used 72,000 classified craters, divided in train (28,000 crater), validation (6,000 craters) and test (45,000 craters).

 

Methodology 

We trained a convolutional neural network classifier based on the YOLOv11 architecture, using a balanced and augmented subset of the crater database. Each image patch is resized and normalized, and we apply standard data augmentation strategies including rotations, flips, and artificial masking to simulate realistic artefacts in CTX images. The model outputs is a classification among the four crater classes describe previously. raining was conducted over 40 epochs on a high-performance multi-GPU server using a cross-entropy loss function and a cosine decayed learning rate schedule. Figure1 show the improvement of accuracy through the learning phase on the validation dataset.

Figure1: Validation accuracy with respect to learning epochs

 

Results 

The final model achieves a classification accuracy of over 80% on a geographically diverse and independent test subdataset containing over 45,000 craters. The Figure2 shows the confusion matrix which gaves us a good insight as how the classification model performed. Performance remains consistent across latitudes. Figure3 shows the classification made on 12 example craters, showing excellent classification, including robustness to illumination conditions and image condition (corrupted data). 

We also demonstrate the practical use of our classification model in the context of surface dating. By comparing cumulative crater size-frequency distributions (CSFD) before and after removing ghost and secondary craters, we show that automated filtering improves the coherence of the inferred ages with those expected from established crater chronologies.

Figure2: Confusion matrix made on the test subdataset. These results show for instance that 80% of true Ghost crater—which represent 2703 craters—where correctly classified, and 2% of them (84 instances) where misclassified as Layered. Overall, the performance is excellent. The regular crater appears slightly more difficult to classify, most probably due to human misclassification.

 

 

Figure3:  Example of 12 crater present in a test area between -100° and -92°E longitude and 0° to 8°S latitude, which were, from top to bottom, classified as Ghost, Layered and Regular.

 

Discussion and Conclusion

We present a novel, scalable, and accurate pipeline for automatic crater classification, which complements existing detection models and provides a new tool for planetary surface dating. 

This study represents the first fully automated morphological classification of Martian impact craters using deep learning. Our results demonstrate the potential of AI-based approaches to improve crater-based chronostratigraphy, especially when applied systematically to global datasets.

As a future work, we plan to extend the model to the Moon and Mercury using transfer learning, but also incorporate additional crater classes or features (e.g., central peaks, double-layer ejecta). Finally, the plan to refine existing Martian chronologies using the filtered crater populations. 

 

References

[1] W. K. Hartmann, G. Neukum, Cratering chronology and the evolution of

mars, Space Science Reviews 96 (2001) 165–194.

[2] G. Neukum, B. Ivanov, W. Hartmann, Cratering records in the inner solar system in relation to the lunar reference system (2001).

[3] B. A. Ivanov, Mars/moon cratering rate ratio estimates, Chronology and Evolution of Mars 87, 2001.

[4] A. Lagain, S. Bouley, & al., Mars crater and database: A and participative project for the classification of and the morphological characteristics of large martian and craters, The Geological, 2021.

[5] G. K. Benedix, A. Lagain, K. Chai, S. Meka, S. Anderson, C. Norman, P. A. Bland, J. Paxman, M. C. Towner, T. Tan, Deriving surface ages on mars using automated crater counting, 2020.

[6] R. La Grassa, G. Cremonese, I. Gallo, C. Re, E. Martellato, Yololens: A deep learning model based on super-resolution to enhance the crater detection of the planetary surfaces, 2023.

[7] L. Martinez, F. Andrieu, F. Schmidt, H. Talbot, M. S. Bentley, Robust automatic crater detection at all latitudes on mars with deep-learning, 2025.

[8] J. L. Dickson, B. L. Ehlmann, L. Kerber, C. I. Fassett, The global context camera (ctx) mosaic of mars: A product of information-preserving image data processing, 2024.

[9] L. Martinez, F. Andrieu, F. Schmidt, M. S. Bentley, Automatic crater classification using a deep-learning-based pipeline, JGR Machine Learning, under review, 2025.