*3.1. Conventional CNN-based Underground Object Classification*

To validate the effectiveness of UcNet, the two experimental validation results of the conventional CNN and newly developed UcNet were compared. Figure 12 shows the conventional CNN-based underground object classification results. Since Phase I of UcNet is equivalent to the conventional CNN, the processing results up to Phase I were considered as the conventional CNN one. As expected, manhole and subsoil background, which have significant features of 2D GPR grid images, are correctly classified compared with the ground truth confirmed by the portable endoscope. However, 11.74% of cavities and 33.73% of gravels are misclassified as each other due to their similar morphological features. The classification performance of the conventional CNN was evaluated by calculating statistical indices called precision and recall using the following equation:

**Figure 12.** The results of conventional convolutional neural network (CNN)-based underground object classification.

Table 1 summarizes the precision and recall values obtained from the conventional CNN results. As for the manhole and subsoil background cases, the precision and recall values are 100%, indicating that they are properly classified by the conventional CNN. On the other hand, 88.26% and 66.27% of the precision values in the cavity and gravel cases, physically meaning that false positive occurs. Similarly, the relatively low recall values of the cavity and gravel cases means the false negative alarm due to the misclassification between the cavity and gravel cases.


**Table 1.** Statistical results obtained from conventional CNN.

#### *3.2. Newly Developed UcNet*

From the Phase I results described in Figure 12, Phases II and III were subsequently carried out. Figure 13a–d shows the representative SR B-scan images which are especially misclassified in Phase I. Figure 13a,b indicates the representative cavities cases misclassified as gravels, and Figure 13c,d shows vice versa. The misclassification results show very similar geometric features to each other, but the phase information at the parabola boundaries are distinctive between the cavity and gravel cases. In particular, although the LR GPR B-scan images has ambiguous pixel-level boundary information, the SR images show that much clearer parabola boundary information, making it possible to conduct the precise phase analysis in the subsequent Phase II.

Figure 14a,b indicates the procedure of parabola boundary extraction results with SR and noise removal image corresponding to Figure 13a,b. All parabola boundaries are clearly extracted from the SR images even though unwanted noise and non-parabola features coexist. Similarly, Figure 14c,d shows that the distinctive parabola boundaries are successfully extracted, which correspond to Figure 13c,d.

(**a**)

(**b**) **Figure 13.** *Cont*.

(**d**)

**Figure 13.** Representative LR and SR B-scan images of (**a**,**b**) cavity cases misclassified as gravels and (**c**,**d**) gravel cases misclassified as cavities.

Figure 15 shows the phase analysis results corresponding to the extracted parabola boundary information in Figure 14. As shown in Figure 15a, Δθ of the radiated waves has 0.062 which is positive value. Then, the Δθ values of 0.0481 and 0.0336 shown in Figure 15b,c indicate that they can be considered as underground cavities, not gravel. Conversely, the Δθ values of the misclassified cases from gravels to cavities have −0.0803 and −0.1073 as shown in Figure 15d,e, respectively. These out-of-phase information physically imply the high permittivity of the object in comparison with the surrounding soil, meaning that they are most likely gravel in the designed category of UcNet.

(**a**) **Figure 14.** *Cont*.

(**d**)

**Figure 14.** The parabola boundary extraction results of (**a**,**b**) cavity cases misclassified as gravels and (**c**,**d**) gravel cases misclassified as cavities.

**Figure 15.** *Cont*.

**Figure 15.** Phase analysis results at the extracted boundaries of (**a**) radiated wave, (**b**,**c**) cavities, and (**d**,**e**) gravels, respectively.

Based on the phase analysis results of Figure 15, the object classification results of Figure 11 were updated as shown in Figure 16. It can be easily observed that the misclassified cavities and gravels are properly updated without false alarms. Since all of misclassification cavity and gravel cases are correctly classified, the statistical precision and recall are increased to 100%.

**Figure 16.** Updated underground object classification results using UcNet.

#### **4. Conclusions and Discussion**

This study newly proposes an underground cavity detection network (UcNet) for enhancing the cavity classification capability. Although convolutional neural networks (CNNs) utilized ground penetrating radar (GPR) triplanar images to classify underground objects, misclassification often occurs due to similar morphological features in B- and C-scan GPR images. This misclassification may lead to a substantial increase of maintenance cost and time. The proposed UcNet overcomes the existing technical hurdle through precise and reliable interpretation of GPR data without expert intervention. In particular, UcNet minimizes the misclassification between cavities and gravel chunks using the conventional CNNs. The effectiveness of the proposed UcNet was experimentally validated using in-situ GPR data obtained on real complex urban areas in Seoul, South Korea. Although the proposed UcNet works well with the validation datasets considered in this study, further investigations on other types of underground objects such as concrete dummies and underground pipes under various in-situ road and underground conditions are warranted. In particular, the authors are now creating our own deep classification network to directly handle 3D GPR data as well as constructing a GPR data library.

**Author Contributions:** M.-S.K. and Y.-K.A. conceived and designed this study. M.-S.K. and N.K. acquired experimental data and performed conventional CNN classification processing. S.B.I., J.-J.L. performed newly proposed UcNet classification processing and analysis. M.-S.K. and Y.-K.A. wrote the entire manuscript and designed the processing code. Y.-K.A. also helped to design the processing code and provided comments.

**Funding:** This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (2018R1A1A1A05078493).

**Acknowledgments:** The authors would like to thank "Development of Evaluation and Analysis Technologies for Road Sink Research Team" for providing the GPR devices.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
