Automatic Detection Of Basal Units Of Ice Sheet With Deep Learning

In recent decades,the accelerating melting of the Greenland and Antarctic ice sheets has led to rising sea level,which has become an important issue in the scientific,environmental,and political fields.Studying the characteristics of subglacial targets is particularly important for reliably analyzing the future evolution trends of ice sheets.Basal Units are structures that exist at the ice sheet-base interface and are markedly different from normal ice sheet.They can provide rich information on subglacial processes and ice sheets evolution,while also exerting a significant influence on the flow of surrounding ice,making them o ne of the key features for polar researchers.This paper proposes a deep learning-based method which can achieve full automatic detection and recognition of Basal Units in radar gram,as opposed to traditional machine learning methods based on manual featur e extraction.A Basal Units dataset is created using radar data collected by the Polar Geophysics Group(PGG)at the Earth Institute of Columbia University.Two different deep learning algorithms are designed for precise identification and real-time identification on mobile devices,respectively.The experiment selects the neural network composed of ResNet and ASPP module to accurately recognize Basal Units.ResNet networks can extract deeper features while achieving fast convergence by increasing the number of network layers.ASPP module can adjust dilation rate to change the receptive field of feature map,achieving recognition of features and radargrams of different sizes.The experimental results demonstrate that this algorithm also has high accurac y on small batch datasets,making it suitable for retrieving rare subglacial features.In addition,Real-ESRGAN super resolution technology is used to preprocess the dataset to improve recognition accuracy.Adapting to small size,light load,and fast movement speed of domestic radar data acquisition device,the paper proposes a more efficient deep learning method,which uses lightweight network MoblieNet V3 instead of ResNet network as the backbone of network structure.MobileNet networks can greatly reduc e model parameters and computational complexity while reducing accuracy slightly,achieving a balance between time consumption and performance,which are suitable for real-time detection scenarios on mobile devices.Compared with the recognition algorithm based on ResNet network,this algorithm can save 81.9% of memory and 92.2% of inferring time,while also having high recognition accuracy,providing a possibility for real-time recognition of radargram features on mobile devices.