基于深度卷积神经网络的场地污染非线性反演方法

提出了将Sobel边缘检测算子与深度卷积神经网络（CNN）算法结合的方法（E-ConvNet），用于污染场地的ERT反演过程.通过Sobel算子提取污染区域视电阻率数据的边缘特征作为CNN的先验信息，提高E-ConvNet的计算效率及识别精度.在5种理论模型（单异常体、双异常体及含双异常体的层状结构）和现场实例上测试了E-ConvNet算法的性能，并与最小二乘算法（LS）比较.测试结果表明：E-ConvNet能够准确识别污染处的面积、位置及阻值，其识别精度和计算效率均优于LS.E-ConvNet的单异常识别准确率为81.8%~84.9%，而LS则仅为9.6%~36.2%；多异常识别准确率为68.6%~84.4%，LS仅为2.8%~27.6%；E-ConvNet用时约为112~190ms，LS耗时为6000~7000ms.因此，在污染场地调查工作中，E-ConvNet能够准确高效地反演出污染区域的位置及范围，为开展后续评估/修复工作提供技术支持.

A site pollution nonlinear inversion method based on deep convolutional neural network

A novel nonlinear method named E-ConvNet was proposed for ERT inversion of contaminated sites, which combined Sobel edge detection operator with deep convolutional neural network algorithm (CNN). The edge features of apparent resistivity data in contaminated sites were extracted by the Sobel operator as the prior information feed into CNN, which improved computational efficiency and identification accuracy of E-ConvNet. The performance of E-ConvNet algorithm were testedon five theoretical model data (single anomaly, double anomalies, and layered structures with double anomalies) and field data, and then compared with results from the traditional Least Squares (LS) algorithm. The results showed that E-ConvNet can accurately identify the area, location and resistance of pollution, and its accuracy and computing efficiency were better than those of LS. The single anomaly recognition accuracy of E-ConvNet and LS were 81.8%~84.9% and 9.6%~36.2%, respectively; the multiple anomalies recognition accuracy of E-ConvNet and LS were 68.6%~84.4% and 2.8%~27.6%, respectively; the computing time of E-ConvNetis about 112~190ms,and the computing time of LS was 6000~7000ms. Therefore, E-ConvNet proposed in this study can accurately and efficiently inversed the polluted areas in the investigation of contaminated sites and provide technical support for the follow-up assessment/restoration work.