Deep Learning‐Based Automated Segmentation of Layered Crater Ejecta on Mars

Layered crater ejecta (LCE) with fluidization on Mars is generally believed to be related to the near‐surface ice‐bearing deposits. The improvement in the resolution of planetary image data has led to the discovery of more and more LCE, providing support for further research on the formation and geological significance of layered ejecta craters (LECs). Concurrently, there is a requirement for a method that can efficiently segment LCE from imagery. This study presents an Efficient Multi‐Head Multi‐Scale Attention Efficient Up‐Convolution UNet (EMHSEU UNet) model, which is capable of identifying the extent of LCE of craters on Mars that have already been identified as having LCE. This model enhances segmentation performance by introducing the efficient multiscale attention (EMA) module and a efficient up‐Convolution block, and transforming EMA into a multihead attention mechanism. Through ablation experiments and comparative experiments on the data from NASA′s Planetary Data System, we have demonstrated that EMHSEU UNet outperforms existing models and is effective in segmenting LCE, which can facilitate geological analysis and understanding of Mars' geologic processes. By applying our proposed model, we supplemented the commonly used parameters (ejecta mobility and lobateness (Г)) for 2,937 single‐layered ejecta craters (SLECs) in the Robbins Mars impact crater database, which were not counted due to the image clarity at the time. These data can provide reference value for future studies on the distribution of water on Mars. Future work will focus on enhancing this approach to better identify other classes of LEC on Mars.