Sub-theme
4. Supporting Aquatic Ecosystem Health and Functions
Topic
4-7. Groundwater and ecosystem
Body
The interface simplified models of seawater intrusion in coastal aquifers are generally divided into two types: abrupt interface model and transition zone interface (or wedge interface) model. Electrical resistivity tomography (ERT) is the visualization of the subsurface resistivity distribution in 2D or 3D, which is a mainstream method in monitoring the seawater intrusion. A geoelectrical recognition model to classify the two simplified seawater/freshwater interface types in the process of seawater intrusion based on Convolutional Neural Network (CNN) was established. This model was applied in the seawater intrusion laboratory experiments which were carried out to simulate the process of seawater intrusion caused by seawater level rise or freshwater level drop. The main body of the experimental sand tank is designed to be 1.80m long, 0.10m wide and 0.60m high, and the main body of the tank is filled with fine sand with the thickness of 0.55m. The geoelectrical recognition model was designed as a CNN structure consisting of three convolutional layers, three maximum pooling layers, two fully connected layers and one Softmax layer, and the size of convolutional cores is 3×3 and the maximum pooling size is 2×2. The preliminary test was designed to conduct the resistivity values of fine sand aquifer under different Cl- concentrations. The training sample database of seawater/freshwater interface geoelectrical models with different resistivity, wedge dip angles, morphology and width of transition zone, etc., was established with 0.2 noise added, and Wenner-α device was selected as the measuring method, and finite element method was used in resistivity forward simulation. A total of 686 training samples were participated in the training process, and the CNN training epochs, dropout learning rate is set, and batch size of training were set depending on the parameter optimization. The training results showed that the average training accuracy was 0.9581 and cross entropy loss was 1.3500, and the average training time cost was 860.23s. The ERT monitoring method were carried out to analyze the apparent resistivity of the aquifer during the seawater intrusion process, and the fully trained CNN recognition model was introduced to classify the interface types which presented accurate results that the interface type can be accurately identified. In the future research, this geoelectrical recognition model can be applied in solving the field monitoring problems of the seawater intrusion in both homogeneous and heterogeneous aquifers.