Multiscale Crater Extraction On Moon Via Deep Learning

With the development of space industry,deep space exploration activities are carried out by countries all over the world.As the earth of natural satellite,the moon has scarce materials and special places,hence the moon has become one of the main research objects of deep space exploration.Craters are regarded as natural landmarks of stars,and play a key role in the study of stars,and they are natural landmarks of stars.And the study of the size and location of craters has important significance for the landing site selection and operation obstacle avoidance of the probe.However,the large number and wide scale of craters on the moon make it difficult to compile a complete crater catalogue.In order to improve the effect of multi-scale lunar crater extraction,the following studies are carried out in this paper:The radius range of craters extracted by the existing crater extraction algorithms based on deep learning is often small,and the multi-scale craters affect the extraction effect,resulting in the recall rate to be improved.In view of the above problems,the extraction effect is improved by improving relevant methods.On the improvement of the method,a pyramidal image segmentation method based on U-Net is proposed.The image pyramid strategy is introduced to ensure that the study area was covered with a small number of images,and each crater existed in multiple images with different resolutions.Firstly,the crater pyramidal images are constructed;then,randomly cropped images are used to train U-Net model,and weight file is saved;thirdly,the craters rim of the pyramidal images are segmented by the U-Net model loaded with weight file;fourthly,crater coordinates and radius are obtained by the crater the location and size extraction;finally,the multi-scale craters in each layer of the image pyramid are fused layer by layer.In the experiment,based on Lunar Digital Elevation Map(DEM)data,the effectiveness of the image pyramid strategy added,and superiority of the proposed method are verified by comparing with other segmentation models with or without the image pyramid strategy,and the method before improvement and other segmentation models.The results show that image pyramid can effectively improve the effect of crater,and this method improve effectively the detection performance of large-scale crater,93.1% of the craters are recovered from the manual annotation dataset,and the maximum radius of extracted craters was extended to 241.6 km,it further realized the extraction of multi-scale crater.Owing to the DEM only contains the height characteristics of crater,it is not good to extract the craters with no obvious edge height.Digital Orthophoto Map(DOM)contains richer edge semantic features,but it is easily affected by lighting conditions and shadows,resulting in many missed craters.Therefore,in this paper,DEM and DOM data are used for crater extraction,and the research is carried out about how to improve the extraction effect of craters in DOM data,so as to get a more complete craters catalogs of testing area.In view of the pyramidal image segmentation method based on U-Net,a framework for crater extraction based on DEM and DOM pyramidal hybrid images is proposed,and the edge height features of DEM are used to facilitate crater extraction from DOM data.DOM data,DEM and DOM hybrid data are used for extract crater respectively,and DEM and DOM hybrid data are used to train the model to increase the extraction effect of single DOM data,and DEM and DOM hybrid extraction catalogs are generated to improve crater extraction in the testing area.The experiments show that the extraction recall of pyramidal image segmentation method based on U-Net for DOM data,DEM and DOM hybrid data is improved by 31.0% and 44.7% at most,respectively,compared with the original method;and the extraction recall of DOM data using DEM and DOM hybrid training model is improved by 4.9% at most,compared with the single DOM training model;compared with DEM extraction catalog,DOM extraction catalog,DEM and DOM hybrid testing extraction catalog,DEM and DOM hybrid extraction catalogs has obtained the highest recall rate of 96.7%.Through experiments,it is verified that the pyramidal image segmentation method based on U-Net has good generalization performance,and the mixed DEM and DOM training model can improve the crater extraction effect of DOM data.Moreover,DEM and DOM hybrid extraction catalogs can recover more craters.