Precision in planetary exploration: Crater detection with residual U-Net34/50 and matching template algorithm

Abstract

Counting craters on celestial bodies such as the Moon is essential for understanding the Solar System's evolution and its dynamic past. The conventional crater recognition methods rely heavily on human observation, often paired with standard machine learning techniques. However, these methods face obstacles, like the absence of an objective criterion and challenges in achieving accurate recognition results for overlapping or smaller craters. To mitigate these issues, our proposed solution involves implementing a convolutional neural network known as the Residual U-Net-34 and Residual U-Net-50. These models are designed to effectively identify craters within lunar Digital Elevation Model (DEM) images. The initial step of our method involves the emphasizing of crater edges and suppressing the other surfaces within lunar DEM data. Following edge detection, we employ matching template algorithm to calculate the size and positions of craters within the lunar DEM data. In the Residual U-Net-34 and Residual U-Net-50 architectures, the framework extends the U-Net model by incorporating the residual convolution block, departing from conventional convolution methods. This hybridization leverages the strengths of both U-Net and residual networks. Remarkably, Residual U-Net-34 and Residual U-Net-50 maintains the input and output image sizes, simplifying the training process due to its use of residual units. This design also enables straightforward optimization of the proposed model. The approach focuses on crater rims and demonstrates the ability to identify overlapping craters. Within the domain of lunar crater recognition, our model demonstrates elevated performance with a recall of 77.22% and precision of 83.67% when operating on DEM data. Notably, the recall and precision outperform other deep learning methods. These experimental outcomes validate the feasibility of leveraging our network for crater recognition within lunar DEM datasets.