LCDNet: An Innovative Neural Network for Enhanced Lunar Crater Detection Using DEM Data

Dingruibo Miao, Jianguo Yan, Zhigang Tu, Jean-Pierre BarriotIEEE JSTARS 2024 · DOI

Introduction

Mapping lunar craters is a fundamental task for planetary science, providing insights into the solar system's impact history and enabling autonomous spacecraft navigation. However, automated detection systems face two primary bottlenecks: the "Incomplete Teacher" problem, where human-curated catalogs are missing real craters, and the "Overlap Dilemma," where standard post-processing algorithms accidentally delete real, overlapping craters. LCDNet addresses these challenges by integrating novel sampling and suppression techniques into a deep-learning framework.

Methodology

LCDNet is built upon a Faster R-CNN backbone with a Feature Pyramid Network (FPN) to effectively detect craters of varying sizes. The core innovations lie in two specific modules:

  • Reserved Negative Sampling (RNS): Standard supervised learning penalizes a model for predicting a crater that isn't in the training catalog. RNS identifies high-confidence predictions that have no overlap with known craters and "reserves" them rather than labeling them as mistakes. This allows the model to act as an explorer, learning from unmapped features.
  • Adaptive Non-Maximum Suppression (A-NMS): Instead of using NMS as a rigid post-processing step, LCDNet integrates it into the training phase. It utilizes a Pull Loss to encourage multiple detections of the same crater to converge and a Push Loss to ensure that detections of distinct, overlapping craters remain separate.
The LCDNet architecture integrates Reserved Negative Sampling (RNS) and Adaptive NMS (A-NMS) within the Region Proposal Network (RPN) to improve detection accuracy and handle overlapping craters.

Data Preparation

The study utilized high-resolution Lunar Digital Elevation Models (DEM) from the Lunar Reconnaissance Orbiter (LOLA) and Kaguya missions. DEM data is preferred over optical imagery because it is invariant to changing lighting conditions and shadows. To prepare the data, the global DEM was resampled and partitioned into a multi-scale grid, ensuring the network could recognize craters across a wide range of diameters.

The global lunar DEM is partitioned into training, validation, and test regions, then resampled into multi-scale grids to create a robust dataset.

Key Results and Insights

LCDNet significantly outperformed state-of-the-art models like DeepMoon and YOLO-v5, achieving a 95% recall rate on test sets. The model's ability to "learn from the unknown" led to the discovery of 30,028 potential new craters. Upon cross-referencing with high-resolution secondary catalogs, 24,504 of these were confirmed as credible, previously unmapped craters. In complex, high-density scenes, the A-NMS module allowed the model to maintain a recall of over 90%, whereas traditional models saw significant performance drops due to over-cleansing of overlapping detections.

Limitations and Future Work

Despite its success, LCDNet faces challenges with very small craters (1-2 km) due to the resolution limits of the underlying DEM. Additionally, detecting very small craters nested entirely within much larger ones remains a spatial challenge. Future research will explore multi-data fusion—combining optical images with DEM data—and transfer learning to apply the LCDNet framework to other planetary bodies like Mars and Mercury.

Summary updated

BibTeX
@article{miao2024lcdnet,
  author = {Dingruibo Miao and Jianguo Yan and Zhigang Tu and Jean-Pierre Barriot},
  title = {{LCDNet: An Innovative Neural Network for Enhanced Lunar Crater Detection Using DEM Data}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2024},
  doi = {10.1109/jstars.2024.3405580},
}
Figure