Classifying Electrical Resistivity Tomography Profiles of Underground Utilities using Convolutional Neural Network

Utilities such as pipelines are vital for the urban community. The most used material for pipelines is metal and plastic that may have different size and shape depending on its use. Due to stress, heat, and pressure overtime, underground pipelines may encounter breakage that may lead to problems such as road cracks and pipe leakage. Subsurface monitoring such as ERT can be used to detect subsurface artifacts such as underground utilities to conduct maintenance and prevent damage caused by subsurface artifacts. ERT measurement utilizes geophysical software and instruments that relies heavily on the resistivity of the subsurface that will result to the subsurface profile. The ERT profile will result to a contoured image indicating different subsurface artifacts or anomalies in the region of interest. The development of deep learning techniques paved the way for emerging studies concerning AI being applied to ERT. In this study, CNN using pretrained models such as InceptionV3, ResNet101, NasNetLarge, and MobileNetV2 was applied to homogenous ERT profiles containing pipes to classify the profile into metallic and plastic pipe. The generated synthetic profiles are pre-classified to contain either metallic pipe or plastic pipe. The performance of pretrained models will be evaluated by their confusion matrix. The model that performed best is the ResNet101 model, producing the highest accuracy of 83% compared to other models. The reconfigured pre trained model can be integrated to geophysical software to provide more information with the profile and may lead to minimized amount of effort on inversion process.