Pavement Moisture Content Prediction: A Deep Residual Neural Network Approach for Analyzing Ground Penetrating Radar

The non-destructive detection and monitoring of moisture content in asphalt concrete (AC) pavement is important, as moisture may cause adhesion failures between aggregates and asphalt, resulting in AC stripping and raveling. This article introduces a novel deep residual network-Mois-ResNets-to predict the internal moisture content of AC pavement from ground-penetrating radar (GPR) measurements. A GPR signal database was established from field tests and numerical simulations. A developed heterogeneous numerical model was used to generate synthetic GPR signals by simulating asphalt pavements at various configurations, moisture contents, volumetrics, and dielectric properties. The Mois-ResNets model, which contains a short-time Fourier transform followed by a deep residual network, was trained to minimize the error between predicted moisture content level and ground-truth data. Testing results show that Mois-ResNets can achieve a classification accuracy of 91% on testing datasets, outperforming conventional machine learning methods. The proposed Mois ResNets has the potential for using GPR measurements and deep learning methods for pavement internal moisture content prediction.