Detection of macropore-sacle soil heterogeneities related to water preferential flows at meter depth by using Electrical Resistivity Tomography (ERT) and machine learning 2D inversion

Electrical resistivity tomography (ERT) is a non-destructive geophysical method that provides images of the soil electrical resistivity. In particular, this study focuses on small-scale soil subsurface areas (up to 1 m deep). Soil resistivity depends on several factors, particularly soil water content. Consequently, ERT has proven to be effective in detecting water flow pathways during infiltration tests. ERT has already been used for hydroecological and hydrogeological applications, but its use still requires further investigation and new developments, particularly for specific processes such as preferential flows. Indeed, water infiltration into the soil is all but homogeneous with the occurrence of preferential flows due to several factors (lithological heterogeneity, macropores, cracks, macrofauna galleries, and root channels). The presence of these preferential flows promotes the downward movement of water and solutes to deeper soil horizons. Thus, the detection of these flows using geophysical methods, such as ERT, should provide a better understanding of the dynamics of preferential flows. However, conventional ERT inversion methods have proven unable to provide insight into the processes at adequate scales (e.g., the macropore scale) and have shown difficulties in detecting sharp variations because of smoothing constraints. Information from apparent resistivity may be lost because of these constraints.

To overcome these limitations, an inversion method based on Convolutional Neural Networks (CNNs) is proposed to detect small-scale resistivity heterogeneities. For the training step, we designed a specific generator to produce synthetic resistivity data for various 2D random electrical resistivity distributions mimicking different types of soil heterogeneity (earthworm and root induced macropore, layering, etc.) and for typical protocols for ERT acquisition. For each case, our database associates the true resistivity field with the apparent resistivity. This database contains a large number of training pairs that allow machine learning.

The preliminary results show that while some predictions were able to predict properly the soil heterogeneities (shape and size), the values of the estimated true resistivity were far from the target values. To avoid unrealistic estimates, we added physical constraints during the training process. An additional forward calculation was performed based on the true resistivities predicted by the neural network, then the apparent resistivities corresponding to the predicted resistivities were compared to the apparent resistivities corresponding to the targeted true resistivities and a supplementary loss function was added to the initial loss function. At this stage, the neural networks were trained on heterogeneous resistivity distributions in the soil with neither time evolution nor link to hydrological processes. In the future, the resistivity generator will be coupled with hydrological models to simulate water infiltration dynamics and changes in soil water content over time. This perspective is essential for applying the proposed method to detect flow pathways during infiltration tests.