基于深度学习的非均质土壤条件下探地雷达反演方法

Investigation of Ground-Penetrating Radar Inversion Methods in Heterogeneous Soil Conditions

传统的地震或雷达勘探技术可以提供地下结构的一些信息。探地雷达反演方案旨在解决地球物理勘探中的地下目标检测和成像问题。本文介绍了一种基于深度学习的新型探地雷达图像反演网络,称为多特征提取网络块编码器—解码器(Encoder-Decoder with Multiple Feature Extraction Blocks,EDMFEBs)。该网络对接收到的电磁波反射信号进行分析而推断出地下结构和目标的性质、位置和几何形状。在本研究中,通过一种名为贪婪通道—空间注意力(Greedy Channel-Spatial attention,GCS attention)的新方法以提高EDMFEBs网络模型的准确性。EDMFEBs反演结果的结构相似性指数(Structural Similarity Index Measure,SSIM)达到了99.21%,平均相对误差(Mean Relative Error,MRE)降低为10.48%,均方误差(Mean Square Error,MSE)降为0.000336,平均绝对值损失(Mean Average Error,MAE)降为0.00105,相比于最新的一种基于多感受野双U形卷积神经网络(Double U-shape Convolutional Neural Networks with Multiple Receptive Fields,DMRF-UNet),SSIM和MRE分别提高了1.7%和9.27%,MSE和MAE分别下降了0.00014和0.00092。实验表明,该方法提升了反演速度,提高了反演能力,简化了模型参数,提高数据处理效率,节省时间和人力资源。

Traditional seismic or radar exploration techniques can provide some information about subsurface structures.Ground-penetrating radar inversion methods aim to address the problem of subsurface target detection and imaging in geophysical exploration.This paper introduces a novel deep learning-based ground-penetrating radar image inversion network,referred to as encoder-decoder with multiple feature extraction blocks(EDMFEBs).This network analyzes the received electromagnetic wave reflection signals to infer the properties,locations,and geometric shapes of subsurface structures and targets.In this study,a new approach called greedy channel-spatial attention(GCS attention)is employed to enhance the accuracy of the EDMFEBs network model.The structural similarity index measure(SSIM)of EDMFEBs inversion results reaches 99.21%,the mean relative error(MRE)is reduced to 10.48%,the mean square error(MSE)is reduced to 0.000336,and the mean absolute error(MAE)is reduced to 0.00105.Compared to the latest double U-shape convolutional neural networks with multiple receptive fields(DMRF-UNet),SSIM and MRE are improved by 1.7%and 9.27%,respectively,while MSE and MAE are reduced by 0.00014 and 0.00092,respectively.The experiments demonstrate that this method enhances inversion speed,improves inversion capabilities,simplifies model parameters,enhances data processing efficiency,and saves time and human resources.