Research On Lunar Impact Crater Identification Based On Convolutional Neural Network

As one of the most common geomorphic structural units on the lunar surface,lunar impact craters record the history of billions of years of lunar impacts.Accurate identification of lunar impact craters can help to advance lunar exploration and build a deeper understanding of the moon for humanity.In recent years,with the accelerated development of deep space exploration technology,high-precision lunar measurement data has provided a research basis for the accurate identification of lunar impact craters.Early crater identification studies mainly used morphology,statistics,and other methods to identify them by constructing the geometric features of impact craters artificially.However,these methods have a high computational complexity,are operationally complex,and have low identification accuracy,making it difficult to meet current research needs.Convolutional Neural Networks are emerging in the field of image recognition due to powerful data representation capabilities,and the application of Convolutional Neural Networks for accurate identification of impact craters has become the mainstream research direction in this field.However,due to the large number of lunar impact craters and their different characteristics in terms of region,size,and age,conventional Convolutional Neural Networks are not effective in distinguishing them.Therefore,based on existing research,this thesis conducts an depth analysis of the challenging issues in the identification of impact craters and proposes solutions.On the one hand,by improving the existing Convolutional Neural Network architecture to enhance its learning the ability to impact craters.On the other hand,a hierarchical labeling strategy is proposed for the diameter feature of impact craters to improve the utilization of training data.In this thesis,the high-precision lunar digital elevation model provided by NASA laboratory is selected as the experimental data,which is automatically annotated by Python software based on the expert marker library of lunar impact craters,and the semantic segmentation task and object detection task are respectively used as the starting point to carry out the experiment,to verify the effectiveness of the improved network and stratification strategy The main research content of this thesis is as follows:（1）In the semantic segmentation task,UNet architecture is selected as the basic framework for the semantic segmentation experiment of impact crater,and Res UNet,DARUNet and Mobile UNet network based on UNet architecture are designed.Res Net network with stronger feature extraction ability and Mobile Net network with more efficient operation are used to replace the coding layer of the original UNet network to enhance the feature extraction ability and operation efficiency of the impact crater.At the same time,the Dense ASPP module is designed to enhance the multi-scale feature capture capability of the network,and the DARUNet network is proposed on the basis of Res UNet.Finally,the segmentation results of each network are combined,a post-processing module is designed to improve the overall segmentation effect of impact crater.（2）In the object detection task,the instance segmentation model Mask RCNN network is reconstructed based on the deep learning framework MMdetection,and the training process of the model is optimized.Increase data enhancement operations and enhance model generalization.According to the diameter range of lunar impact craters,a stratification strategy is proposed,and a multi-scale lunar impact craters data set is generated to improve the utilization of training data.The experimental results show,the₁score of Res UNet network constructed in this thesis reaches68.42%,the₁score of DARUNet network reaches 68.73%,and the₁score of Mobile UNet network reaches 68.45%.After the post-processing module,the₂score reached 76.23%,which effectively improved the prediction accuracy.In the lunar impact crater target detection mission,the₁score of the Mask RCNN model reached 83.4%in the prediction results of the original dataset,and that of the Mask RCNN model reached 84.3%after using the data stratification strategy.