A study on geoelectrical recognition model of seawater/freshwater interface based on convolutional neural network: an application in sand tank experiments

Seawater intrusion is a global environmental issue, and seawater intrusion monitoring requires a multidisciplinary approach to improve accuracy. Simplified seawater/freshwater interface models for coastal aquifers are generally divided into two types: abrupt interface models and wedge-shaped interface models. Electrical resistivity tomography (ERT) is the visualization of subsurface resistivity distributions in 2D or 3D and has been widely used in seawater intrusion monitoring. This paper presents a geoelectrical recognition model for classifying simplified seawater/freshwater interface types based on a convolutional neural network (CNN). The CNN structure is composed of three convolutional layers, three max pooling layers, two fully connected layers, and one Softmax layer. A total of 686 samples were combined for model training, and obtained 0.9581 for the average accuracy (ACU) and 1.3500 for the average cross-entropy loss (CEL). Sand tank experiments were carried out to simulate the process of seawater intrusion caused by a rise in the water level of sea water rise or a decrease in the water level of fresh water, the ERT method was used to monitor the resistivity of the aquifer during the experiments, and the fully trained CNN model was used to classify the interface types. According to the output data, the probability of observing the wedge-shaped interfaces during the experiments at 300 and 345 min were 98.85% and 99.89%, while the probability of observing the abrupt interfaces were 1.15% and 0.11%. The results showed that the ERT method offers a fast and nondestructive approach for monitoring seawater intrusion, and accurate recognition results of interface types were obtained using a well-trained recognition model in the laboratory experiments.