Analysis of noise immunity of a neural network system for recognizing subsurface objects
Abstract
Background: Subsurface sensing systems based on impulse electromagnetic waves are widely used for detecting hidden objects in soil. However, the performance of such systems significantly degrades in the presence of strong interference caused by external reflections and radio-frequency sources. The task of automated object recognition from ground-penetrating radar (GPR) signals using artificial neural networks (ANNs) is particularly challenging, since the noise robustness of such systems strongly depends on the selection of modeling parameters, input data formation, and training procedures.
Objective: To investigate and optimize the parameters of an ANN-based subsurface object detection system in order to improve its noise robustness by applying a parameter grid search algorithm.
Materials and Methods: A subsurface medium containing a hidden metallic object was irradiated with a plane impulse electromagnetic wave. The electrodynamic problem of wave propagation and scattering was solved numerically using the finite-difference time-domain (FDTD) method, taking into account the interaction between the electromagnetic field, soil, and the object. The input data for the neural network were formed from the time-domain responses of the received signals, as well as from additional information obtained using a discrete tomography approach and ray tracing. A grid search was employed to identify the optimal system configuration by analyzing the influence of input data type, number of field sensors, time window parameters, data augmentation techniques, and target encoding strategies. Noise robustness was evaluated using the F1-score and the threshold Gate SNR metric.
Results: The grid search analysis identified system configurations that provide the highest noise robustness for object recognition. It was shown that the selection of the time window is a critical factor significantly affecting the Gate SNR values. Training the neural network on noisy data was found to enhance its generalization capability and improve stability under noisy conditions. Furthermore, the encoding strategy for the absence-of-object class was demonstrated to have the strongest impact on system performance, enabling lower Gate SNR thresholds compared to other investigated parameters.
Conclusions: The application of the grid search algorithm enabled systematic optimization of the parameters of an ANN-based subsurface object recognition system. The obtained results confirm the effectiveness of combining physically motivated signal processing methods with machine learning techniques to improve the noise robustness of ground-penetrating radar systems. The proposed approach can serve as a basis for further development of automated GPR data analysis methods under real experimental conditions.
Downloads
References
Gazdag J, Sguazzero P. Migration of seismic data. Proc IEEE. 1984;72(10):1302–1315. doi:10.1109/PROC.1984.13044.
Smitha N, Bharadwaj DRU, Abilash S, Sridhara SN, Singh V. Kirchhoff and F-K migration to focus ground penetrating radar images. Geo-Engineering. 2016;7:4. doi:10.1186/s40703-016-0019-6.
Plakhtii V, Dumin O, Pryshchenko O. Kirchhoff migration method for tube detection with UWB GPR. In: Proc. 2021 IEEE Int. Conf. on Direct and Inverse Problems of Electromagnetic and Acoustic Wave Theory (DIPED); 2021 Sep. p. 1–5. doi:10.1109/DIPED53165.2021.9552330.
Alani AM, et al. The use of ground penetrating radar and microwave tomography for the detection of decay and cavities in tree trunks. Remote Sens. 2019;11(18):2073. doi:10.3390/rs11182073.
Chen Y, Witten AJ. Pseudoinverse imaging for multimonostatic ground-penetrating radar data. Proc SPIE. 2001. doi:10.1117/12.445458.
Pryshchenko O, Dumin O, Plakhtii V. Discrete tomography approach for subsurface object detection by artificial neural network. In: Proc. 2022 IEEE Ukrainian Microwave Week (UkrMW); 2022 Nov. p. 1–4. doi:10.1109/UKRMW58013.2022.10037072.
Windsor CG, Capineri L, Falorni P. A data pair-labeled generalized Hough transform for radar location of buried objects. IEEE Geosci Remote Sens Lett. 2014;11(1):124–127. doi:10.1109/LGRS.2013.2248119.
Varyanitza-Roshchupkina LA, Pochanin G. Wavelet analysis of signals in short-range radiolocation problems. In: Proc. 2013 Int. Conf. on Antenna Theory and Techniques (ICATT); 2013 Sep. p. 1–4. doi:10.1109/ICATT.2013.6650794.
Plakhtii V, Dumin O, Pryshchenko O, Shyrokorad D, Pochanin G. Influence of noise reduction on object location classification by artificial neural networks for UWB subsurface radiolocation. In: Proc. 2019 IEEE Int. Seminar on Direct and Inverse Problems of Electromagnetic and Acoustic Wave Theory (DIPED); 2019 Sep. p. 1–4. doi:10.1109/DIPED.2019.8882590.
Wang S, Han L, Gong X, Zhang S, Huang X, Zhang P. MCMC method of inverse problems using a neural network: application in GPR crosshole full waveform inversion. Remote Sens. 2022;14(6):1320. doi:10.3390/rs14061320.
Persanov I, Dumin O, Plakhtii V, Shyrokorad D. Subsurface object recognition in soil using UWB irradiation by butterfly antenna. In: Proc. 2019 IEEE DIPED; 2019. p. 160–163. doi:10.1109/DIPED.2019.8882577.
Pryshchenko OA, et al. Implementation of an artificial intelligence approach to GPR systems for landmine detection. Remote Sens. 2022;14(17):4421. doi:10.3390/rs14174421.
Capineri L, Falorni P, Borgioli G, Bossi L, Pochanin G, Ruban V, et al. Background removal for processing scans acquired with the “UGO-1st” landmine detection platform. In: Proc. 2019 PhotonIcs & Electromagnetics Research Symposium—Spring (PIERS); Rome, Italy; 2019 Jun. p. 3965–3973.
Dumin O, Plakhtii V, Pryshchenko O, Pochanin G. Comparison of ANN and cross-correlation approaches for ultra-short pulse subsurface survey. In: Proc. 2020 IEEE TCSET; 2020. p. 1–6. doi:10.1109/TCSET49122.2020.235459.
Feng K, Zhao Y, Wu J, Ge S. Cross-correlation attribute analysis of GPR data for tunnel engineering. In: Proc. 2014 Int. Conf. on Ground Penetrating Radar (ICGPR); 2014 Jun. p. 1–4. doi:10.1109/ICGPR.2014.6970461.
Feng K, Zhao Y, Wu J, Ge S. Cross-correlation attribute analysis of GPR data for tunnel engineering. In: Proc. 2014 Int. Conf. on Ground Penetrating Radar (ICGPR); 2014 Jun. p. 1–4. doi:10.1109/ICGPR.2014.6970461.
Mint Mohamed Mostapha A, Faize A, Alsharahi G, Louzazni M, Driouach A. Effect of external noise on ground penetrating radar ability to detect objects. Int J Microw Opt Technol. 2019;14:124–131.
Agapiou A, Sarris A. Working with Gaussian random noise for multi-sensor archaeological prospection: fusion of GPR depth slices and ground spectral signatures. Remote Sens. 2019;11(16):1895. doi:10.3390/rs11161895.
Lee Y. Image classification with artificial intelligence: cats vs dogs. IEEE Xplore; 2021 Jan 1. doi: 10.1109/CDS52072.2021.00081
Pryshchenko O, Dumin O, Plakhtii V, Pochanin G. Subsurface object detection in randomly inhomogeneous medium model. In: Proc. 2021 IEEE UKRCON; 2021 Aug. p. 1–4. doi:10.1109/UKRCON53503.2021.9575688.
Pryshchenko O, Dumin O, Plakhtii V, Pochanin G. Classification of objects buried in inhomogeneous medium by artificial neural network using impulse GPR with 1Tx + 4Rx antenna system. In: Proc. 2021 Int. Workshop on Advanced Ground Penetrating Radar (IWAGPR); 2021 Dec. p. 1–4. doi:10.1109/IWAGPR50767.2021.9843169.
Dumin O, Plakhtii V, Shyrokorad D, Pryshchenko O, Pochanin G. UWB subsurface radiolocation for object location classification by artificial neural networks based on discrete tomography approach. In: Proc. 2019 IEEE UKRCON; 2019 Jul. p. 1–4. doi:10.1109/UKRCON.2019.8879827.
Avola D, et al. A shape comparison reinforcement method based on feature extractors and F1-score. IEEE Access. 2019;7:1–10. doi: 10.1109/SMC.2019.8914601.
This work is licensed under a Creative Commons Attribution 4.0 International License.
