Periodic Reporting for period 2 - BOULDERING (A Deep Learning approach for boulder detection –The key to understand planetary surfaces evolution and their crater statistics-based ages)
Reporting period: 2023-10-01 to 2024-09-30
Boulders are one of the most abundant features on the surfaces of solid planetary bodies. Measuring their size, shape, and orientation can tell us about how they formed as well as help select landing sites that minimize hazard to spacecraft. However, mapping boulders across large areas is a labor-intensive task that often limits the scope and robustness of boulder studies. To overcome this challenge, we trained a machine-learning algorithm to automatically outline boulders on a variety of planetary surfaces using a database of over 30,000 boulders manually mapped from aerial or satellite images of Earth, the Moon, and Mars. Our algorithm, BoulderNet (https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2023JE008013(opens in new window)) performs as well as human mappers and outperforms existing automated tools. BoulderNet is made available to the community.
Why is it important for society?:
The trained boulder algorithm has the potential to assist space agencies and commercial planetary ventures greatly. For example, it can help in the (1) selection of safe landing site, (2) landing of spacecraft in real time, (3) navigation of rovers on planetary surfaces, and (4) mapping of boulders, which could represent an important resource for infrastructure constructions along with other regolith materials.
Why is it important for the science community?
Boulder mapping could help reveal how planetary surfaces evolved. Craters are very common surface features on many solid planets and moons. During an impact, rock fragments ejected from the crater cavity could be deposited elsewhere on the surface, where they could potentially form secondary craters. Boulders are the only remnants of these ejected materials. Their size and shape, as well as the terrain on which they are found, provide important insight into the ejection mechanisms. In addition, the application of our open-source boulder detection algorithm will be important in the understanding of several geological processes shaping planetary surfaces(e.g. (i) estimate the age of a crater based on its abundance of boulders, since boulders degrade over time; (ii) investigate the magnitude of seismic activity and (iii) map geological units around impact craters.
The manuscripts are accompanied by 5 open-access repositories:
- Raw drone data of the two fieldworks conducted in the Sierra Nevada (https://zenodo.org/records/14585533(opens in new window))
- Raw input and labeled boulder data collected during the project (https://zenodo.org/records/14250970(opens in new window))
- Pre-processed images and labels for use with Detectron2 and YOLOv8 (https://zenodo.org/records/14250874(opens in new window))
- Code, best trained model setups and weights for YOLOv8 (https://zenodo.org/records/14579518(opens in new window))
- Boulder populations around 82 fresh simple impact craters on the Moon and 15 fresh simple impact craters on Mars (https://zenodo.org/records/14253940(opens in new window))
And 4 github repositories:
- Manipulation of rasters (https://github.com/astroNils/rastertools(opens in new window))
- Manipulation of vector data (https://github.com/astroNils/shptools(opens in new window))
- Pre-processing, model setup and predictions with Detectron2 Mask R-CNN (https://github.com/astroNils/MLtools(opens in new window))
- Pre-processing, model setup and predictions with YOLOv8 (https://github.com/astroNils/YOLOv8-BeyondEarth(opens in new window))
More information can be found in the Tech. Report Part B.
Why is it important for society?:
The trained boulder algorithm has the potential to assist space agencies and commercial planetary ventures greatly. For example, it can help in the (1) selection of safe landing site, (2) landing of spacecraft in real time, (3) navigation of rovers on planetary surfaces, and (4) mapping of boulders, which could represent an important resource for infrastructure constructions along with other regolith materials.
Why is it important for the science community?
Boulder mapping could help reveal how planetary surfaces evolved. Craters are very common surface features on many solid planets and moons. During an impact, rock fragments ejected from the crater cavity could be deposited elsewhere on the surface, where they could potentially form secondary craters. Boulders are the only remnants of these ejected materials. Their size and shape, as well as the terrain on which they are found, provide important insight into the ejection mechanisms. In addition, the application of our open-source boulder detection algorithm will be important in the understanding of several geological processes shaping planetary surfaces(e.g. (i) estimate the age of a crater based on its abundance of boulders, since boulders degrade over time; (ii) investigate the magnitude of seismic activity and (iii) map geological units around impact craters.
The manuscripts are accompanied by 5 open-access repositories:
- Raw drone data of the two fieldworks conducted in the Sierra Nevada (https://zenodo.org/records/14585533(opens in new window))
- Raw input and labeled boulder data collected during the project (https://zenodo.org/records/14250970(opens in new window))
- Pre-processed images and labels for use with Detectron2 and YOLOv8 (https://zenodo.org/records/14250874(opens in new window))
- Code, best trained model setups and weights for YOLOv8 (https://zenodo.org/records/14579518(opens in new window))
- Boulder populations around 82 fresh simple impact craters on the Moon and 15 fresh simple impact craters on Mars (https://zenodo.org/records/14253940(opens in new window))
And 4 github repositories:
- Manipulation of rasters (https://github.com/astroNils/rastertools(opens in new window))
- Manipulation of vector data (https://github.com/astroNils/shptools(opens in new window))
- Pre-processing, model setup and predictions with Detectron2 Mask R-CNN (https://github.com/astroNils/MLtools(opens in new window))
- Pre-processing, model setup and predictions with YOLOv8 (https://github.com/astroNils/YOLOv8-BeyondEarth(opens in new window))
More information can be found in the Tech. Report Part B.
Work package 1, status: completed.
The first WP focused mainly on collecting very high-resolution imagery of dry unvegetated rocky surfaces in California with a drone to provide us with (1) data to train the automatic boulder detection algorithm, (2) ground truth boulder data and (3) investigate the influence of the environment and illumination condition during image acquisition on the detected number of boulders. Two locations with significant boulder density were selected in the Sierra Nevada (California). The fieldwork campaign was conducted at the Obsidian and South Deadman volcanic domes in June 2022, and at the Courtright reservoir in September 2022. At each location, drone images were acquired multiple times during the day (under different illumination conditions), and we manually measured between 20 and 30 boulders. This provided us with very high-resolution boulder data and ground truth data. The latter was significant to investigate the model's performances in the first manuscript.
Work package 2, status: completed.
A very long project period was used to collect additional boulder data from satellite imagery of the surface of the Moon and Mars. This process is very time-consuming (this is why we want to automatize this process!) as there are many boulders. Unfortunately, we had to go through this process as deep learning algorithms require many images to learn detection patterns automatically. A big chunk of 2022 and early 2023 was therefore used in the digitization of boulders, familiarization with the model architecture, conduction of pre-processing steps, coding of scripts to manipulate vector and raster data, and training and fine-tuning of the deep learning model. The current version of our neural network, BoulderNet, was trained from a data set of >33,000 boulders in >750 image tiles from Earth, the Moon, and Mars. During early and mid-2023, the time was used to write the first manuscript and continuously improve the models.
Work package 3, status: 50-60% completed.
The boulder detection algorithm successfully identified more than 2,000,000 boulders around 82 young and fresh simple impact craters on the Moon and 15 fresh simple impact craters on Mars. The resulting database provides valuable support for both deep learning applications in boulder detection and detailed analyses of ejection patterns. This is the first large-scale database focused specifically on boulders, addressing a significant gap in planetary science compared to the well-established impact crater databases. While the addition of new boulder populations is temporarily on hold to prioritize the preparation of the third manuscript of the BOULDERING project, significant progress is being made in analyzing the existing dataset. The manuscript, aimed at refining our understanding of ejection mechanisms during crater formation, has its methods and results sections completed. Work on the remaining sections is underway, with the goal of submitting the article by March 1, 2025.
The first WP focused mainly on collecting very high-resolution imagery of dry unvegetated rocky surfaces in California with a drone to provide us with (1) data to train the automatic boulder detection algorithm, (2) ground truth boulder data and (3) investigate the influence of the environment and illumination condition during image acquisition on the detected number of boulders. Two locations with significant boulder density were selected in the Sierra Nevada (California). The fieldwork campaign was conducted at the Obsidian and South Deadman volcanic domes in June 2022, and at the Courtright reservoir in September 2022. At each location, drone images were acquired multiple times during the day (under different illumination conditions), and we manually measured between 20 and 30 boulders. This provided us with very high-resolution boulder data and ground truth data. The latter was significant to investigate the model's performances in the first manuscript.
Work package 2, status: completed.
A very long project period was used to collect additional boulder data from satellite imagery of the surface of the Moon and Mars. This process is very time-consuming (this is why we want to automatize this process!) as there are many boulders. Unfortunately, we had to go through this process as deep learning algorithms require many images to learn detection patterns automatically. A big chunk of 2022 and early 2023 was therefore used in the digitization of boulders, familiarization with the model architecture, conduction of pre-processing steps, coding of scripts to manipulate vector and raster data, and training and fine-tuning of the deep learning model. The current version of our neural network, BoulderNet, was trained from a data set of >33,000 boulders in >750 image tiles from Earth, the Moon, and Mars. During early and mid-2023, the time was used to write the first manuscript and continuously improve the models.
Work package 3, status: 50-60% completed.
The boulder detection algorithm successfully identified more than 2,000,000 boulders around 82 young and fresh simple impact craters on the Moon and 15 fresh simple impact craters on Mars. The resulting database provides valuable support for both deep learning applications in boulder detection and detailed analyses of ejection patterns. This is the first large-scale database focused specifically on boulders, addressing a significant gap in planetary science compared to the well-established impact crater databases. While the addition of new boulder populations is temporarily on hold to prioritize the preparation of the third manuscript of the BOULDERING project, significant progress is being made in analyzing the existing dataset. The manuscript, aimed at refining our understanding of ejection mechanisms during crater formation, has its methods and results sections completed. Work on the remaining sections is underway, with the goal of submitting the article by March 1, 2025.
During the first two years, a big focus was on completing the two first work packages (fieldwork and training of a robust boulder detection algorithm). The boulder detection algorithm gives us now the possibility to investigate the two main primary objectives of the BOULDERING project:
1. To determine the relationship between boulders and fragments that produce secondary craters, and to assess how both boulders and secondary craters influence the age derived by the cratering statistics method.
2. To evaluate the influence of impactor and terrain properties on the ejection cratering mechanisms.
Two additional manuscripts (one per primary objective) will be written in the third year of the project (outside of this reporting period). Investigating the boulder populations around the freshest impact structures on the lunar surface will help us answer the second primary objective. As a second step, we will study the spatial distribution of boulders compared to secondary craters, answering primary objective (1).
Please refer to the first question for the potential impacts of the project on society and the science community.
1. To determine the relationship between boulders and fragments that produce secondary craters, and to assess how both boulders and secondary craters influence the age derived by the cratering statistics method.
2. To evaluate the influence of impactor and terrain properties on the ejection cratering mechanisms.
Two additional manuscripts (one per primary objective) will be written in the third year of the project (outside of this reporting period). Investigating the boulder populations around the freshest impact structures on the lunar surface will help us answer the second primary objective. As a second step, we will study the spatial distribution of boulders compared to secondary craters, answering primary objective (1).
Please refer to the first question for the potential impacts of the project on society and the science community.