基于卷积神经网络的月球南极-艾特肯盆地撞击坑自动识别及中型撞击坑绝对模式年龄估算

Automated crater detection on the South Pole-Aitken basin of the Moon and absolute model ages of mid-sized craters based on convolution neural network

月球南极-艾特肯盆地是太阳系最大的撞击盆地之一,也是月球上最大、最古老的撞击盆地.南极-艾特肯盆地是研究早期大型撞击事件的重要窗口,而小型撞击坑的识别与计数定年是研究南极-艾特肯盆地演化史的基础.由于撞击坑直径和数量符合幂次定律,数量众多的小型撞击坑难以单纯依靠人力进行识别.近年来,计算机算力的提升使得训练复杂的卷积神经网络成为可能.采用已有的专家标注训练神经网络,进而实现图像特征的自动提取,能够在保证准确率的同时极大地提高识别效率.采用基于卷积神经网络算法的You Only Look Once Version5(YOLO V5)目标探测系统来自动识别月球南极-艾特肯盆地直径为2~15 km的小型撞击坑.在训练神经网络时,使用融合了SELENE和LRO数据的数字高程模型SLDEM2015和最新的专家标记撞击坑数据库.训练好的网络在测试集上的结果与专家标记的撞击坑数据库相比,识别结果的准确率(Precision)为0.96,召回率(Recall)为0.95,F1值为0.95.通过对与专家标注不符的识别结果进行可视化,识别出至少十个专家误标记的撞击坑,证明撞击坑自动识别方法可以用于检验专家标注的可靠性.基于南极-艾特肯盆地的撞击坑自动识别结果,确定了南极-艾特肯盆地四个典型中型撞击坑的绝对模式年龄,并与已有的定年结果对比,进一步验证了自动识别结果的可靠性,也显示了提出的方法在利用自动识别的撞击坑进行中型撞击坑定年方面的潜力.提出的撞击坑自动识别方法有望进一步拓展到更小撞击坑的识别,并迁移到月球其他地质单元乃至其他行星的研究中.

The South Pole-Aitken Basin(SPA) is one of the largest impact basins in the solar system,and the largest and oldest impact basin on the Moon. The SPA basin provides a critical example for investigating giant impact events on the Moon at its early evolution stage. Because the diameter and number of impact craters follow a power law,it is difficult to identify numerous small impact craters merely by human labor. In recent years,the improvement of computer computing power makes it possible to train complex convolutional neural networks. Automatic identification of craters can be realize by a trained neural network,simultaneously improving the efficiency and ensuring the accuracy of crater identification. In this study,we applied the You Only Look Once Version 5(YOLO V5) target detection system based on the convolutional neural network algorithm to automatically identify small impact craters with a diameter range of 2~15 km in the SPA basin. For the model training,we used SELENE-LRO merged digital elevation model SLDEM2015 and the latest expert-labeled crater catalog. Compared with the expert-labeled crater catalog,the trained network gets a precision of 0.96,a recall of 0.95,and an F1-score of 0.95 on the test set. By verifying our identified craters inconsistent with the expert labeling,we find more than 10 impact craters mislabeled by the expert. This proves that the automatic crater identification method can be used to verify the reliability of expert labeling. Based on the automatically identified craters,we also determine the absolute model ages for four typical mid-sized craters,providing a useful application of the crater identification results. Our estimated absolute model ages are consistent with existing dating results. We expect that the automatic crater identification of this study will be extended to the identification of smaller craters,and be transferred to other geological units on the Moon and even other terrestrial planets and rocky satellites.