Imaging Subsurface Hydraulic Properties From Spectral Electrical Tomographic Images: Artificial Neural Network Approach

地球地下的地质环境、水文地质和地球物理研究涉及含水层特征、污染物的运输和污染物修复策略的设计，这些研究要求对水力特性的空间变异性进行准确可靠的预测。本研究的主要目标是开发一种综合方法，该方法结合了基于物理化学的模型开发、实验室测量、使用人工神经网络的解释和成像工具，通过频率相关电阻率图像预测地球地下的水力导电性和孔隙度图像。将开发基于物理化学原理的描述频率相关土壤电阻率的模型，其中包括水力和岩石物理参数作为主要变量。实验室实验将设计用于测量来自不同地质环境的100多个样品的频率相关电阻率以及岩石物理和水力特性，旨在获得土壤质地和结构特性的足够变异性。这些测量将允许全面研究土壤的谱电响应如何受到其岩石物理和水力特性的影响。土壤样品将在实验室中通过测量其水力导率、孔隙度、密度、含水量和粒度分布来充分表征。将探索人工神经网络（ANN）适应、推广和识别变量间关系非线性的能力，以确定土壤的谱电响应与其水力特性之间的函数关系。电阻率和相位在特定频率下的现场测量将在已知水力特性信息的选定地质地点进行。提出了一种正则牛顿法反演方法，并将其用于获得地下电性的频率扫描层析成像。利用基于人工神经网络函数关系开发的转换，将从现场测量的电特性层析成像中预测水力特性的地下图像。本文将评估所提出的预测方法中由于噪声和模型参数错误表征以及训练方法效率低下而产生的不确定性。该方法从现场测量的水力特性输出将与现有的地面真实信息进行比较。

Geo-environmental, hydrogeological and geophysical studies of the earth's subsurface involving aquifer characterization, transport of contaminants and design of contaminant remediation strategies demand accurate and reliable prediction of the spatial variablity of hydraulic properties. The primary goal of this research is to develop an integrated approach, which incorporates physico-chemically based model developments, laboratory measurements, interpretational and imaging tools using artificial neural networks, to predict hydraulic conductivity and porosity images of the earth's subsurface from frequency dependent resistivity images. Models describing the frequency dependent resistivity of soils based on physico-chemical principles will be developed that will include hydraulic and petrophysical parameters as primary variables. Laboratory experiments will be designed to measure frequency dependent resistivity as well as the petrophysical and hydraulic properties of over 100 sil samples fromdifferent geological environments aiming to obtain adequate variablity in soil textural andstructural properties. These measurements will allow a comprehensive investigation of how spectral electrical responses of soils are influenced by their petrophysical and hydraulic properties. The soil samples will be fully characterized by measuring their hydraulic conductivity, porosity, density, moisture content and particle size distribution in the laboratory. The capabilities of artificial neural networks (ANN) to adapt, generalize and identify non-linearities in relations among variables, will be explored to determine the functional relationships between the spectral electrical response of soils and their hydraulic properties. Fieldmeasurements of resistivity and phase at specified frequencies will be conducted at selected geologic sites with known information on the hydraulic properties. An inversion procedure using a regularized Newton's method is proposed and will be utilized to obtain frequency - scanned tomographic images of electrical properties of the subsurface. Subsurface images of hydraulic properties will be predicted from the field measured tomographic images of electrical propertiesusing transformations to be developed based on ANN functional relationships. Uncertainties in the proposed prediction method due to noise and model parameter misrepresentations and inefficiencies in the training method will be assessed. The outputs of the method from field measured hydraulic properties will be compared to that of existing ground truth information.