Accurate real-time ANN classification of cavities and utilities using 3D GPR data in the MFCC domain

Abstract

Real-time identification of serviceable utilities and cavities is critical for the safety of urban settlements because delay has been reportedly costly and destructive. For utilities, maintaining their integrity is the goal, while for cavity imaging, prevention or mitigation becomes the focus. Several existing research has recommended some methodology for classification of these utilities and cavities, but their approaches are either geometrically restrictive or cumbersome. We therefore present a practical methodology that provides accurate real-time classification of cavity, box, manhole, patch, and pipe. 3D GPR data is reshaped into 1D supertrace and a 2D mel-frequency cepstrum coefficients (MFFC) feature are computed, serving as an input pattern, paired up with a utility class. The ANN is subsequently trained and tested during training to avoid overfitting of data. 99% model testing accuracy was achieved after 100 epochs, translating to only 3 misclassifications out of 1000 3D GPR data inputs used for computing a confusion matrix.