Location of objects in a heterogeneous soil using pulse georadar and artificial neural networks
DOI:
https://doi.org/10.26565/2311-0872-2022-37-06Keywords:
ground-penetrating radar, Gaussian impulse, artificial neural network, heterogeneous medium, humanitarian demining, ultrawideband signalAbstract
Background: Detection of hidden objects in the soil layer is a problem that is important in construction, archeology, humanitarian demining, non-destructive testing of road surfaces, flaw detection etc. Studying the peculiarities of electromagnetic filed behavior in heterogeneous media provides a possibility to create subsurface survey systems that can work effectively in real conditions.
Objectives: To recognize an object hidden in a heterogeneous medium using ultrawideband ground penetrating radar (GPR) and artificial neural network (ANN), to evaluate the performance of a neural network that is trained only on homogeneous medium, to investigate the stability of recognition results in the presence of noise of different levels in the received time dependencies, to compare the results with a network which is trained on heterogeneous media. Check the ability of the ANN to correctly identify typical objects that were not involved in training.
Materials and methods: Modeling of the electrodynamic problem of electromagnetic field propagation is carried out using the finite difference in time domain (FDTD) method. The classification of hidden objects is carried out using the approach of artificial neural networks.
Results: An effective algorithm for detecting objects in a heterogeneous soil model was developed using ground-penetrating radar and ANN. Probability distributions of the classification of hidden objects in presence of additive Gaussian noise in time dependencies were obtained.
Conclusion: The use of ANNs has shown successful results in the classification of objects located in a heterogeneous ground model. The ability to detect objects containing only a few metal parts was demonstrated. The developed algorithm has a high level of noise immunity even at high signal-to-noise levels. The results of detection and recognition of typical objects, which were not involved in ANN training, showed the effectiveness of this approach.
