基于深度学习及广义高斯分布的时空多尺度GPR全波形反演

探地雷达(GPR)的全波形反演(FWI)是一种定量获取探测对象物性参数的高分辨成像技术。然而该技术计算量庞大，非线性、不适定性问题突出，且在实际应用中会受到源子波未知、噪音干扰等问题的制约，无法快速精确地指导生产实践。为此，申请人通过构建广义高斯分布目标函数，解决FWI的子波依赖和数据噪音问题；利用多信息约束的复合正则化方法，改善多参数反演中的串扰问题，提高反演稳定性和精度；结合深度学习的低频数据预测器和深度学习的超分辨技术构建时空多尺度反演框架，减少反演的非线性，避免陷入局部极值；最后，通过深度学习的快速正演方法和梯度预测器方法，提高反演的计算效率。可以预见，通过结合深度学习、广义高斯分布及复合正则化等技术提升GPR数据FWI的性能，能达到增强GPR反演算法高效性、可靠性和实用性的目的，加快推动实测雷达数据高精度实时反演的进程。

The full-waveform inversion (FWI) of ground penetrating radar (GPR) is a high-resolution imaging technology for quantitatively obtaining the physical parameters of the detected object. However, this technology has a large amount of calculation, the nonlinearity and ill-posedness of it are prominent, and in practical application, it will be restricted by unknown source wavelet, noise, and other problems, so it cannot guide the production implementation quickly and accurately. Therefore, the applicant first proposes to construct the generalized Gaussian probability distribution misfit function for solving the problem of wavelet dependence and noise of FWI; secondly, the multi-information constrained compound regularization method is used to eliminate the crosstalk problem in multi-parameter inversion, so as to improve the inversion stability and accuracy; then, a spatiotemporal multi-scale inversion framework is constructed by combining the deep learning (DL) based low-frequency data predictor and the DL based super-resolution technique to reduce the nonlinearity of inversion and avoid falling into local minima; and finally, the DL based fast forward modeling method and DL based gradient predictor method are used to increase the efficiency of inversion. It can be predicted that the combination strategy of deep learning, generalized Gaussian distribution and compound regularization could improve the performance of FWI, enhance the efficiency, reliability, and practicability of the inversion algorithm for GPR data, so as to accelerate the process of the real-time inversion of field radar data with high precision.