Size estimation of underground targets from GPR frequency

spectra: A deep learning approach

Nairit Barkatakia,∗ , Banty Tirub and Utpal Sarmaa
a Department of Instrumentation & USIC, Gauhati University, Guwahati, India
b Department of Physics, Gauhati University, Guwahati, India

ARTICLE INFO ABSTRACT

Keywords: GPR (Ground Penetrating Radar) is a robust and effective device for identifying underground
object size estimation artefacts. Construction companies and civil engineers should be aware of the sizes of rebars
object detection and pipelines before and during construction work for various reasons. Most research efforts
FFT have typically concentrated on GPR signal analysis in the time domain, however recent studies
deep learning have increasingly focused on analysis in the frequency domain. This paper proposes an artificial
ground penetrating radar neural network (ANN) model for estimating the diameter of underground rods (solid) and pipes
(hollow). GPR data captured in the time domain domain is transformed to the frequency domain
using FFT, after which feature extraction is performed using ANN. An FPGA-based prototype
GPR system is used to collect GPR A-Scan data for a variety of targets made of aluminium,
stainless steel, rebar, and PVC. A mean absolute percentage error of 1.89% is achieved using the
proposed model. The experimental results confirm the effectiveness of the proposed approach
in extracting size-related information from GPR data.

1. Introduction
Ground Penetrating Radar (GPR) is a versatile and reliable tool that is used for detecting buried objects. Its major
strengths include its low cost compared to traditional methods, and non-destructive nature. There are numerous ap-
plications of GPR, including estimation of soil, snow and ice properties (Jol, 2008; Wu et al., 2019; Liu et al., 2021),
non-destructive inspection of archaeological sites (Miccinesi et al., 2021), examining building condition (Daniels,
2004), inspecting roads and tunnels (Persico, 2014), detecting landmines and other buried explosive devices (Daniels,
2004), borehole inspection (Solla et al., 2021), study agriculture properties of soil (Akinsunmade et al., 2019), locating
weak zones within embankments (Tomecka-Suchoń, 2019) etc.
A GPR survey often takes much lesser time than alternative technologies. When compared to other geophysical
tools such as seismic, transient electromagnetic, electrical, and magnetic techniques, GPR gives data that has higher
resolution and better accuracy (Benedetto and Pajewski, 2015).
Planning and conducting an effective GPR survey requires a lot of experience. The interpretation of GPR data
is usually not intuitive and special skills are required to convert the measurements into clear images that can help in
making critical decisions in various fields like civil engineering, archaeology, military etc.
In case of GPR-based concrete, bridge and road investigations, the most common applications appear to be the
detection of pipes (Ayala-Cabrera and Izquierdo, 2021; He et al., 2021) and reinforcement bars (Wang et al., 2020),
identifying defects in roads and bridges (Rasol et al., 2022), and to determine the thickness of pavements (Zhao and
Al-Qadi, 2017).
One area where significant improvement has been realised is in the estimation of the size of buried targets. Several
researchers have published their work on estimating rebar diameter (Mechbal and Khamlichi, 2017; Pasculli et al.,
2018; Sun et al., 2021) However, the signal processing strategies proposed by them have proven to be complex and
time intensive.
When the diameter of a pipe is smaller than the wavelength of the signal, estimating its diameter becomes chal-
lenging. In such cases, the arrival times of the echo signal cannot be co-related to the the diameter of the pipe (Jazayeri
et al., 2019). This is typical in underground pipe detection, where the pipe diameter may be 50 mm and the wavelength
∗ Corresponding author
nairitb@[Link] (N. Barkataki)
ORCID (s): 0000-0002-7988-072X (N. Barkataki); 0000-0003-1761-0983 (B. Tiru); 0000-0001-7888-1088 (U. Sarma)

: Page 1 of 14
of the transmitted pulse is 75 cm (for 400MHz centre frequency antenna). A variety of solutions have been proposed
to overcome this drawback.
GPR does not directly measure the diameter of rebars, cables or pipes (Rathod et al., 2019). When used with targets
having larger diameters, GPR can produce stronger radar wave reflections. However, depending on the wavelength of
the signal, smaller objects may produce very low intensity reflections that may seem like a dot in the final B-Scan image.
Rathod et al. (2019) presented a study where they tried to estimate rebar spacing and diameters using GPR. They were
able to estimate rebar spacing with errors ranging from 2-11%. However, the percentage errors in determining rebar
diameter was beyond 100% in some cases.
In other studies, researchers have used a combination of GPR and an EM-based system to detect and estimate
the size of rebars. Zhou et al. (2018a) developed a GPR-EMI dual sensor based system that simultaneously estimated
rebar diameter and cover thickness. Barrile and Pucinotti (2005) combined GPR with a different commercial EM-based
system to detect rebars. They reported that they could estimate rebar diameter with a 12% error. This method, however,
was time-consuming as data collection using the handheld concrete pachometer was slow. A linear relationship was
found between the rebar diameter and maximum GPR amplitude by Hasan and Yazdani (2016a). However, it is not
possible to quantitatively relate the peak amplitude of the echo signal with the rebar diameter without knowing the
amplitude of the source wavelet and concrete conditions.
The quality of GPR data is usually degraded by background clutter (Kumar et al., 2021), noisy environments like
roots and bricks (Zhang et al., 2021), heterogeneous soils and mutual interactions of the waves (Lei et al., 2019).
Researchers have provided a variety of ways for analysing GPR data in recent years. Histograms of Oriented Gra-
dients (HOG) based feature extraction method has been used effectively to identify landmines (Torrione et al., 2013).
Fourier Transform and Principal Component Analysis have also been used to identify landmines (Ko et al., 2012).
Diameter estimation of underground pipelines and wires by fitting hyperbolas and localising hyperbola coordinates are
widely used methods we well (Wahab et al., 2013; Mertens et al., 2015; Zhou et al., 2018b; Jazayeri et al., 2019; Lei
et al., 2019; Liu et al., 2020).
Other techniques begin with machine learning algorithms that extract hyperbolic features, followed by a fitting
algorithm. Deep-learning-based interpretation algorithms have been the subject of extensive study in recent years
(Gamba and Lossani, 2000; Shaw et al., 2005; Ahmed et al., 2021).
Among the earliest proponents of using neural networks in GPR data interpretation, Costamagna et al. (1998)
used neural networks to search for object signatures from GPR data. They demonstrated that pipe signatures extracted
using neural networks had the same degree of precision as human operators. To determine the position of buried
items in GPR pictures, Singh and Nene (2013) used a curve fitting technique along with neural network. The buried
objects were localised with curve fitting technique. Their proposed approach could locate buried objects with a total
reported accuracy of 91%. The adoption of the deep learning neural network idea for GPR underground utilities has
demonstrated promising results (Luo et al., 2020; Tong et al., 2020; Barkataki et al., 2022).
Lei et al. (2019) utilised a R-CNN to isolate potential hyperbolic regions from gray GPR B-scan images. The
detected rectangular region was then converted to a binary image and hyperbolic signatures were identified using a
double cluster seeking estimate algorithm. Their proposed scheme could automatically extract information related
to buried objects with a precision of 96% on real GPR data. Lei et al. (2020) proposed a scheme to estimate the
diameters of buried objects. Their method involved an adaptive target region detection (ATRD) algorithm to isolate
the hyperbolic signatures in B-Scans, followed by a CNN model for feature extraction and estimation of diameter.
Their framework achieved an accuracy of 92.5% on field datasets.
The majority of research efforts have traditionally focused on the analysis of GPR signals within the time domain,
although recent studies have increasingly focused on analysis in the frequency domain. Laurens et al. (2005) analysed
the relationship between the spectral content of signals reflected by concretes of various degrees of saturation, using a
1.5 GHz antenna system. They found a difference in the central frequency of the reflected signals when the spectrum
for dry (1.4 GHz) and saturated (1.18 GHz) concrete were compared.
Studies by Rodés et al. (2015) have shown that there is a distinct correlation between the spectrum of the recorded
traces and the subsurface structures. Radar signals are scattered when targets are present in the medium. The amplitude
of the wave, A(r) at a certain distance r from the source, can be represented as a function of the wave amplitude at the
source and the losses caused by those attenuating effects:
1
𝐴(𝑟) = 𝐴0 𝑒−(𝛼+𝜇)𝑟 (1)
𝑟

: Page 2 of 14
 𝐴0 being the amplitude of reference at the source, r the distance, 𝛼 the coefficient of attenuation due to the absorption
and 𝜇 the coefficient of scattering.
The coefficient of scattering depends on the number of scatterers per unit of volume, n and on the cross section of
the scatterers, 𝐶𝑆 .
𝑛𝐶𝑠
𝜇= (2)
2
Moreover, the effect of scattering is frequency dependent, because the cross section depends on the frequency of
the incident field (Salinas Naval et al., 2018).
Estimating the size of the target object is one area where considerable progress has been made. Several research
groups have established experimental relationships between the amplitude and frequency content of the hyperbolic
returns and the diameters of embedded bars, as well as additional properties of the rebar and concrete (Kalogeropoulos
et al., 2011, 2012; Lai et al., 2013; Hasan and Yazdani, 2016b).
Bi et al. (2018) analysed the spectral characteristics of target GPR signals and proposed a method to suppress
random noise from GPR data. They used a singular value decomposition (SVD) method to filter out unwanted data
and reconstruct the GPR data. They found that the proposed approach showed improved results in suppressing noise
and clutter, compared with the conventional time-domain method.
According to the study presented by Che et al. (2021), it was easier to detect the depth and position of buried objects
when analysing the GPR data in frequency domain. Park et al. (2021) applied frequency-wavenumber migration (f-k
migration) using the Fast Fourier Transform (FFT) to the GPR data and used YOLO-v3 algorithm to detect rebars
and estimate their diameters. Three rebars were used having diameters 10 mm, 20 mm and 30 mm. Their proposed
model could estimate rebar diameters with an average accuracy of 98% and was found to be more accurate than other
conventional methods. GPR data was also analysed in the frequency domain by Kang et al. (2022) to detect and
estimate the size of a cavity beneath a concrete plate.
However, there is still scope of improvement in the accuracy, speed, and other parameters of the models proposed
by researchers. This study presents an novel artificial neural network (ANN) based model to estimate the diameter
of buried rods (solid) and pipes (hollow). GPR data collected in the time domain domain is converted to frequency
domain using FFT, after which the ANN-based estimator is used. Data is collected using a Field Programmable Gate
Array (FPGA) based prototype GPR system.

2. Materials and Methods

2.1. Hardware Platform for the GPR System
FPGA-based embedded systems provide a rapid prototyping platform, given the flexibility of these devices for intri-
cate system designs, as well as low static and dynamic power consumption, high performance and a re-programmable
medium to implement a variety of applications.

Figure 1: Signal processing platform with test setup

: Page 3 of 14
Figure 2: Block diagram of the system

A hardware platform for the prototype GPR system has been developed which uses a Xilinx UltraScale+ MPSoC
FPGA board as the base. A 14-bit, 1.0 GSPS ADC and a 16-bit, 2.8 GSPS DAC are interfaced to the FPGA board
for both signal generation (transmission) and digitisation of the received signals. A ricker, gaussian or any custom
waveform can be generated using the DAC. This signal (waveform) is then amplified using a power amplifier before
feeding it to a transmitting antenna. Two custom designed 400 MHz coaxial-fed bow-tie antennas are used as transmit-
ting and receiving antennas (Barkataki et al., 2021). The received signal is first amplified using a low noise amplifier
(LNA) before feeding the signal to the ADC. The digitised signal is then processed by the FPGA and various GPR
signal processing algorithms are applied before a final image (A-Scan / B-Scan) can be formed (see Figure 1). A block
diagram of the hardware system is shown in Figure 2.

2.1.1. Experimental Setup

A test bed of dimension 1350 𝑚𝑚 × 750 𝑚𝑚 × 600 𝑚𝑚 (𝑋 × 𝑌 × 𝑍) is constructed for collecting the experimental
data. It is filled with sand till the height of 400 mm from the bottom of the test bed. Holes are made on one side,
through which the target objects are inserted. The antennas are placed on the surface of the sand to collect the data.
The collected data are sampled by the on-board ADC at 1 GHz which gives a single sample (data point) in 1 ns. A

(a) Front view (b) Side view

Figure 3: Diagram of the experimental setup

: Page 4 of 14
diagram of the experimental setup is shown in Figure 3. A picture of the constructed test bed used for collecting
experimental data is shown in Figure 4.

Figure 4: The test bed used for collecting experimental data

2.1.2. Validation of Hardware Platform

A verification of the hardware platform for the GPR system is necessary to ensure that it can provide reliable results
before starting the experiments for this study. Initially, multiple GPR A-Scans without any target object are collected
for 199 ns, which gives 200 data points from 0th second to 199th second for a single A-Scan. As can be seen in Figure
5, the spectra of the received wavelets are similar. The signal in the time domain is also similar. Later, GPR A-Scans
are also collected for different buried targets. Analysis of the received wavelets indicates a marked difference in the
spectrum of the signal for buried targets, as shown in Figure 6. Having verified the system’s ability to generate reliable
data, experimental data collection is started.

(a) A-Scan 1 (Time Domain) (b) A-Scan 1 (Frequency Domain)

(c) A-Scan 2 (Time Domain) (d) A-Scan 2 (Frequency Domain)

: Page 5 of 14
 (e) A-Scan 3 (Time Domain) (f) A-Scan 3 (Frequency Domain)

Figure 5: Received wavelets for no buried targets

(a) Reflection spectra for a target made of aluminium (b) Reflection spectra for a target made of rebar of 10
of 15 mm diameter mm diameter

Figure 6: Spectrum of received wavelets for buried targets

Table 1
Targets used in this study
Outer Diameter
Sl. No. Target
(mm)
1 Aluminium bar 15
2 Aluminium bar 15
3 Reinforcement bar 10
4 Reinforcement bar 15
5 Reinforcement bar 15
6 Reinforcement bar 22
7 Stainless steel pipe 10
8 Stainless steel pipe 15
9 Stainless steel pipe 18
10 Stainless steel pipe 25
11 Stainless steel pipe 30
12 Stainless steel pipe 38
13 Stainless steel pipe 50
14 PVC Pipe 22
15 PVC Pipe 29
16 PVC Pipe 29
Figure 7 17 PVC Pipe 35
18 PVC Pipe 41
The different bars and pipes used for the study

2.2. Materials
The buried objects are made of aluminium, stainless steel (SS), reinforcement bar (rebar), and polyvinyl chloride
(PVC) with diameters as mentioned in Table 1 and shown in Figure 7. The depth of the buried objects are varied from
60 mm to 140 mm from the soil surface.

: Page 6 of 14
2.2.1. Case 1: Data in Time Domain
Experimental data is collected using the prototype GPR system described earlier. A total of 1713 A-Scans are
collected having 200 data points each. Some collected A-Scans for target of different diameters are shown in Figure 8.

(a) 10 mm (Rebar) (b) 15 mm (Rebar- S. N. 5)

(c) 22 mm (Rebar) (d) 25 mm (Stainless steel pipe)

(e) 22 mm (PVC pipe) (f) 29 mm (PVC pipe)

Figure 8: Received wavelets for various targets having different diameters

2.2.2. Case 2: Data in Frequency Spectrum

After analysing the frequency spectrum of the collected A-Scans, it is seen that the dominant spectrum for all
cases ranges from 200-400 MHz. A band-pass filter of 200-400 MHz is applied to each GPR A-Scan, before further
processing. Figure 9 shows the spectrum of A-Scans after application of the band-pass filter.
From the above figures, it is clear that there is a co-relation between the spectrum and the size of the buried objects,
as was observed by other researchers (Santos et al., 2014; Rodés et al., 2015; Salinas Naval et al., 2018; Pajewski et al.,
2019). However, it is seen that the spectra for the rebars of 15 mm (S. N. 4 and 5 in Table 1) are slightly different. This
may be due to the difference in shape between the two. Very little or no difference in the spectra is seen for targets made
of aluminium (S. N. 1 and 2) and PVC (S. N. 15 and 16) of different qualities. Data is collected using multiple targets
of the same diameter value for data augmentation. This is to ensure that small changes in the object’s orientation do not
affect the trained neural network model. Two separate datasets are created from the GPR data containing information
in time domain (TD-dataset) and frequency domain (FD-dataset).

: Page 7 of 14
 (a) 10 mm (Rebar) (b) 15 mm (Rebar- S. N. 4)

(c) 15 mm (Rebar- S. N. 5) (d) 22 mm (Rebar)

(e) 25 mm (Stainless steel pipe) (f) 22 mm (PVC pipe)

(g) 29 mm (PVC pipe)

Figure 9: Frequency spectrum of received wavelets for various targets

2.3. Proposed ANN architecture

An ANN model is used in this study, where feature extraction occurs hierarchically. It is composed of several
layers, each containing numerous neurons. Neurons in one layer are linked to every other neuron in the next layer.
ANN comprises mainly forward and backward propagation. In forward propagation, each neuron is assigned weights
and biases, as well as an activation function. The error between true and predicted values is calculated using a loss
function. To determine the optimal hyper-parameters, backward propagation uses an optimisation function to minimise
the loss in each successive iteration during training of the model.

: Page 8 of 14
Table 2
Validation Losses for different numbers of hidden layers
No of Validation Loss (MAPE)
Sl. No.
Hidden Layers
Time Frequency
Domain Domain
1 1 Layer 15.63 8.17
2 2 Layers 12.60 3.56
3 3 Layers 9.84 2.61
4 4 Layers 8.78 2.14
5 5 Layers 8.46 2.42
6 6 Layers 7.34 1.89
7 7 Layers 6.72 2.08
8 8 Layers 6.02 2.18
9 9 Layers 5.67 2.28
10 10 Layers 4.70 2.00

Out of a total of 1713 A-Scans, 1456 (85%) are used for training and 257 (15%) are used for validation of the
model. The dataset is normalised before training the model.

2.3.1. Loss Function

Mean absolute percentage error (MAPE) is used as a loss function during training of the model. It is a measure
of the absolute difference between the true and predicted values, and is expressed as a percentage. It is defined as the
mean of absolute relative errors:

1 ∑ || 𝑃𝑡 − 𝐴𝑡 ||
𝑁
𝑀𝐴𝑃 𝐸 = (3)
𝑁 𝑡=1 || 𝐴𝑡 ||

where N is the no. of samples, (𝑃𝑡 ) is the predicted value and (𝐴𝑡 ) is the true value.

2.3.2. Evaluation Metrics

MAPE is utilised as a metric to evaluate the training and validation of the ANN model. MAPE is a measure of a
regression model’s performance that can be expressed in percentages, making it easily understood by anyone.

2.3.3. Final Architecture

Initially, 200 neurons are used in the input layers and multiple hidden layers are added which have half the number
of neurons than the previous layer. Performance is evaluated for different number of hidden layers. Same procedure is
repeated for both the datasets containing GPR data in time domain and frequency domain. Table 2 shows the validation
losses (MAPE) corresponding to different numbers of hidden layers for both datasets. From the table, it is clear that
training the model on the dataset containing GPR spectrum information outperforms that one containing time domain
data.

3. Results and Discussion

Multiple training runs show that the ANN model with 6 hidden layers and the input layer containing 400 neurons
has the best validation loss of 1.89% using spectrum information from the collected GPR data. The proposed model
consists of 7 layers including the input layer. ReLU activation function is used in all the layers. To estimate the size
of the target, the final output layer with 1 unit uses ReLU as the activation function. Adam optimisation algorithm is
used with an initial learning rate of 0.00001. All the models are trained for 2000 epochs and NVIDIA RTX3090 GPU
is used during training and validation. The architecture of the best ANN model is shown in Figure 10.
257 A-Scans are used for evaluating the performance of the model. The best and the worst predicted values for
different diameters of buried objects are shown in Table 3a. Different metrics like mean squared error (MSE), mean
absolute error (MAE), mean absolute percentage error (MAPE), and mean squared logarithmic error (MSLE) are used
to analyse the model’s performance, as given in Table 3b. It is seen that the model has a MSE of 1.017, MAE of 0.457
and MSLE of 0.0017.

: Page 9 of 14
Figure 10: Architecture of the proposed ANN

Table 3
Performance of the ANN model on the validation set
(a) Predicted values (b) Performance metrics

Predicted values Metrics Value

True values Median
Sl. No. (mm) MSE 1.017
(mm) (mm)
Worst Best MAE 0.457
MAPE 1.887
1 10.0 10.84 9.99 9.99 MSLE 0.0017
2 15.0 14.17 15.00 14.97
3 18.0 18.70 17.98 18.11
4 22.0 14.77 22.05 22.10
5 25.0 26.81 25.05 25.05
6 29.0 24.40 29.00 28.83
7 30.0 31.06 30.02 29.94
8 35.0 36.86 35.03 34.88
9 41.0 41.96 41.04 41.04
10 38.0 39.43 37.95 37.95
11 50.0 46.82 50.04 49.86

3.1. Testing on unseen data

Repeatability is a major concern for any algorithm or instrument. Hence, it is essential to test a ANN model on
new and unseen data to be able to properly evaluate its performance.
The experimental setup is used to collect 191 additional A-Scans for the final performance evaluation of the pro-
posed model. When taking measurements, targets of different diameters made of aluminium, rebar, stainless steel, and
PVC are picked at random. Table 4a shows the best and worst expected values for objects of various sizes, along with
different metrics MSE, MAE, MSLE, and MAPE as given in Table 4b.
From Table 4, it is seen that the ANN model performs exceedingly well on the new data, achieving a MAPE of
1.57%. It achieves a MSE of 0.565 and MAE of 0.383. This time, it achieves a 0% error on couple of occasions when

: Page 10 of 14
Table 4
Performance of the ANN model on the new data
(a) Predicted values (b) Performance metrics

Predicted values Metrics Value

True values Median
Sl. No. (mm)
(mm) (mm) MSE 0.565
Worst Best MAE 0.383
1 10.0 09.72 10.00 09.98 MAPE 1.567
2 15.0 16.06 14.99 14.96 MSLE 0.001
3 18.0 17.33 17.98 17.89
4 22.0 23.58 21.96 22.13
5 25.0 24.51 24.98 24.92
6 29.0 24.97 28.86 27.60
7 30.0 28.66 30.00 29.89
8 35.0 36.13 34.96 34.89
9 38.0 37.38 38.01 38.10
10 41.0 40.18 41.14 41.14
11 50.0 47.71 49.97 50.05

Table 5
Performance of the ANN model on for metal targets
(a) Predicted values (b) Performance metrics

Predicted values Metrics Value

True values Median
Sl. No. (mm) MSE 0.149
(mm) (mm)
Worst Best MAE 0.234
MAPE 1.179
1 10.0 09.72 10.00 09.98 MSLE 0.0003
2 15.0 16.06 14.99 14.96
3 18.0 17.33 17.98 17.89
4 22.0 23.58 21.96 22.13
5 25.0 24.51 24.98 24.92
6 30.0 28.66 30.00 29.89
7 38.0 37.38 38.01 38.10
8 50.0 47.71 49.97 50.05

estimating the size of objects having diameters 10 mm and 30 mm.

3.1.1. Metallic targets

Of the collected 191 A-Scans, the data is segregated into metallic and non-metallic objects and the model’s perfor-
mance is evaluated again. The predictions made by the ANN model for metallic objects are shown in Table 5a. From
Table 5b, it is seen that the model is able to achieve a MAPE of 1.18% on metallic targets.

3.1.2. Non-metallic targets

Performance is also evaluated for PVC targets. The true and predicted values by the model for non-metallic targets
are given in Table 6a. Its performance is slightly poorer than the metallic targets, achieving a MAPE of 3.03%. The
signals scattered from non-metal pipes are smaller. The dielectric contrast between the target (pipe) and medium (soil)
also affects the GPR signal. The lower the contrast, the weaker the reflected signal. Moreover, this phenomenon is
frequency-dependent (Shaari et al., 2010) due to which the model has different MAPE values for metallic and non-
metallic targets.

: Page 11 of 14
Table 6
Performance of the ANN model on for non-metal targets
(a) Predicted values (b) Performance metrics

Predicted values Metrics Value

True values Median
Sl. No. (mm)
(mm) (mm) MSE 2.138
Worst Best MAE 0.945
1 29.0 24.97 28.86 27.6 MAPE 3.032
2 35.0 36.13 34.96 34.89 MSLE 0.003
3 41.0 40.18 41.14 41.14

Table 7
Comparison of present work with past studies

Sl. Accuracy /
Author Technique used Application
No. Error
1 Rathod et al. GPR and Profoscope data Rebar detection and di- 2-11% error.
(2019) ameter estimation 100% error for
small diameters
2 Jazayeri et al. full-waveform inversion of GPR data Reinforced concrete map- 11% error
(2019) ping and rebar diameter
estimation
3 Lei et al. (2020) Hyperbolic feature extraction CNN- Diameter identification of 92.5% accuracy
LSTM Framework cylindrical objects in GPR
B-Scans
4 Giannakis et al. Autoencoder as a pre processing step fol- Estimate the diameter of ±6 mm accuracy
(2021) lowed by ANN based regressor reinforcement bars in con-
crete
5 Barkataki et al. Hyperbolic feature detection followed by Estimation of radius/ di- 7% error
(2022) deep CNN ameter of buried objects
6 Present work Feature extraction using ANN on spectral Estimation of diameter of 1.89% error
data of GPR A-Scans buried objects

4. Summary
A new approach to estimate the size of buried objects from GPR A-Scans was proposed in this study. The proposed
ANN model is able to extract important features through several layers of neurons to finally estimate the diameter of
buried objects with a high degree of accuracy. The overall performance of the ANN model is given in Tables 3 and 4.
It was seen that the dataset containing spectrum information of GPR A-Scans outperformed that containing data in the
time domain. The proposed model’s MAPE of 1.89% on the validation set, and 1.57% on new unseen data demonstrates
its high reliability and good repeatability. It is to be noted that the highest errors were found for non-metallic targets
for smaller diameter values.
A comparison with previously reported techniques is shown in Table 7 from which it can be concluded that the
use of spectrum information along with the application of ANN in this problem has demonstrated an improved per-
formance. The current work can be extended to estimating sizes of other anomalies of the subsurface like voids in
concrete, tunnels etc.

Acknowledgements
The authors extend their gratitude to several individuals for their invaluable contributions to this study. Firstly, Dr.
Manoj Kumar Phukan (Principal Scientist, Geo Sciences and Technology Division, CSIR-NEIST, Jorhat) provided
insightful guidance on GPR data interpretation. Ms. Sharmistha Mazumdar’s assistance in visualizing the results was
also greatly appreciated. Additionally, Mr. Anil Kumar Rajak, Mr. Hafizuddin Ahmed, Mr. Hriday Jyoti Saikia, and
Mr. Tilak Chandra Deka provided valuable support during the construction of the experimental setup. The authors

: Page 12 of 14
would also like to acknowledge Mr. Ankur Jyoti Kalita for his contributions to the database creation process. Lastly,
the authors express their gratitude to Ms. Sweta Kumar Gaur for her suggestions that helped to polish the manuscript
linguistically.

Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author
on reasonable request.

CRediT authorship contribution statement

Nairit Barkataki: Conceptualization, Methodology, Formal analysis and investigation, Writing - Original Draft.
Banty Tiru: Supervision, Writing - Review & Editing. Utpal Sarma: Conceptualisation, Resources, Supervision,
Writing - Review & Editing.

References
Ahmed, A.M., Deo, R.C., Ghahramani, A., Raj, N., Feng, Q., Yin, Z., Yang, L., 2021. Lstm integrated with boruta-random forest optimiser for soil
moisture estimation under rcp4. 5 and rcp8. 5 global warming scenarios. Stochastic Environmental Research and Risk Assessment , 1–31.
Akinsunmade, A., Tomecka-Suchoń, S., Pysz, P., 2019. Correlation between agrotechnical properties of selected soil types and corresponding gpr
response. Acta Geophysica 67, 1913–1919.
Ayala-Cabrera, D., Izquierdo, J., 2021. Gpr image interpretation advancement for smarter technical management of water leakage in urban water
infrastructures, in: Earth Resources and Environmental Remote Sensing/GIS Applications XII, SPIE. pp. 220–227.
Barkataki, N., Borah, P., Sarma, U., Tiru, B., 2021. Design of a 400 mhz cavity backed cpw fed bow-tie antenna for gpr applications, in: 2021
International Conference on Industrial Electronics Research and Applications (ICIERA), IEEE. pp. 1–6.
Barkataki, N., Tiru, B., Sarma, U., 2022. A cnn model for predicting size of buried objects from gpr b-scans. Journal of Applied Geophysics 200,
104620.
Barrile, V., Pucinotti, R., 2005. Application of radar technology to reinforced concrete structures: a case study. NDT & e International 38, 596–604.
Benedetto, A., Pajewski, L., 2015. Civil engineering applications of ground penetrating radar. Springer.
Bi, W., Zhao, Y., An, C., Hu, S., 2018. Clutter elimination and random-noise denoising of gpr signals using an svd method based on the hankel
matrix in the local frequency domain. Sensors 18, 3422.
Che, C.K.N.A.H., Melor, K., Joret, A., Razali, M., Ponniran, A., Sulong, M.S., Omar, R., 2021. Frequency based signal processing technique for
pulse modulation ground penetrating radar system. International Journal of Electrical and Computer Engineering 11, 4104.
Costamagna, E., Gamba, P., Lossani, S., 1998. A neural network approach to the interpretation of ground penetrating radar data, in: IGARSS’98.
Sensing and Managing the Environment. 1998 IEEE International Geoscience and Remote Sensing. Symposium Proceedings.(Cat. No.
98CH36174), IEEE. pp. 412–414.
Daniels, D.J., 2004. Ground penetrating radar. The Institution of Electrical Engineers.
Gamba, P., Lossani, S., 2000. Neural detection of pipe signatures in ground penetrating radar images. IEEE Transactions on Geoscience and Remote
Sensing 38, 790–797.
Giannakis, I., Giannopoulos, A., Warren, C., 2021. A machine learning scheme for estimating the diameter of reinforcing bars using ground
penetrating radar. IEEE Geoscience and Remote Sensing Letters 18, 461–465.
Hasan, M.I., Yazdani, N., 2016a. An experimental and numerical study on embedded rebar diameter in concrete using ground penetrating radar.
Chin. J. Eng 2016, 1–7.
Hasan, M.I., Yazdani, N., 2016b. An experimental study for quantitative estimation of rebar corrosion in concrete using ground penetrating radar.
Journal of Engineering 2016.
He, W., Hao, T., Ke, H., 2021. Frequency 3d slice image visualization for gpr applications, in: IOP Conference Series: Earth and Environmental
Science, IOP Publishing. p. 012019.
Jazayeri, S., Kruse, S., Hasan, I., Yazdani, N., 2019. Reinforced concrete mapping using full-waveform inversion of gpr data. Construction and
Building Materials 229, 117102.
Jol, H.M., 2008. Ground penetrating radar theory and applications. Elsevier.
Kalogeropoulos, A., Hugenschmidt, J., van der Kruk, J., Bikowski, J., Brühwiler, E., 2012. Gpr full-waveform inversion of chloride gradients in
concrete, in: 2012 14th International Conference on Ground Penetrating Radar (GPR), IEEE. pp. 320–323.
Kalogeropoulos, A., Van der Kruk, J., Hugenschmidt, J., Busch, S., Merz, K., 2011. Chlorides and moisture assessment in concrete by gpr full
waveform inversion. Near Surface Geophysics 9, 277–286.
Kang, S., Yu, J.D., Han, W., Lee, J.S., 2022. Nondestructive detection of cavities beneath concrete plates using ground penetrating radar and
microphone. NDT & E International , 102663.
Ko, K.H., Jang, G., Park, K., Kim, K., 2012. Gpr-based landmine detection and identification using multiple features. International journal of
antennas and propagation 2012.
Kumar, B.S., Sahoo, A.K., Maiti, S., 2021. Removal of clutter and random noise for gpr images, in: 2021 IEEE 18th India Council International
Conference (INDICON), IEEE. pp. 1–6.
Lai, W.L., Kind, T., Stoppel, M., Wiggenhauser, H., 2013. Measurement of accelerated steel corrosion in concrete using ground-penetrating radar
and a modified half-cell potential method. Journal of Infrastructure Systems 19, 205–220.
Laurens, S., Balayssac, J., Rhazi, J., Klysz, G., Arliguie, G., 2005. Non-destructive evaluation of concrete moisture by gpr: experimental study and
direct modeling. Materials and structures 38, 827–832.

: Page 13 of 14
Lei, W., Hou, F., Xi, J., Tan, Q., Xu, M., Jiang, X., Liu, G., Gu, Q., 2019. Automatic hyperbola detection and fitting in gpr b-scan image. Automation
in Construction 106, 102839.
Lei, W., Luo, J., Hou, F., Xu, L., Wang, R., Jiang, X., 2020. Underground cylindrical objects detection and diameter identification in gpr b-scans
via the cnn-lstm framework. Electronics 9, 1804.
Liu, H., Lin, C., Cui, J., Fan, L., Xie, X., Spencer, B.F., 2020. Detection and localization of rebar in concrete by deep learning using ground
penetrating radar. Automation in construction 118, 103279.
Liu, S., Li, Z., Zhao, G., 2021. Attenuation characteristics of ground penetrating radar electromagnetic wave in aeration zone. Earth Science
Informatics 14, 259–266.
Luo, X., Bai, X., Guo, S., Wang, L., Mi, H., Chen, H., 2020. Machine learning methods applied in detection of buried targets for ground penetrating
radar, in: IET International Radar Conference (IET IRC 2020), IET. pp. 450–454.
Mechbal, Z., Khamlichi, A., 2017. Determination of concrete rebars characteristics by enhanced post-processing of gpr scan raw data. NDT & E
International 89, 30–39.
Mertens, L., Persico, R., Matera, L., Lambot, S., 2015. Automated detection of reflection hyperbolas in complex gpr images with no a priori
knowledge on the medium. IEEE Transactions on Geoscience and Remote Sensing 54, 580–596.
Miccinesi, L., Beni, A., Monchetti, S., Betti, M., Borri, C., Pieraccini, M., 2021. Ground penetrating radar survey of the floor of the accademia
gallery (florence, italy). Remote Sensing 13, 1273.
Pajewski, L., Fontul, S., Solla, M., 2019. Ground-penetrating radar for the evaluation and monitoring of transport infrastructures, in: Innovation in
Near-Surface Geophysics. Elsevier, pp. 341–398.
Park, S., Kim, J., Jeon, K., Kim, J., Park, S., 2021. Improvement of gpr-based rebar diameter estimation using yolo-v3. Remote Sensing 13, 2011.
Pasculli, D., Natali, A., Salvatore, W., Morelli, F., Morandi, D., 2018. Investigation of reinforced concrete bridges by using a dual-polarized
high-frequency gpr, in: 2018 17th International Conference on Ground Penetrating Radar (GPR), IEEE. pp. 1–5.
Persico, R., 2014. Introduction to ground penetrating radar: inverse scattering and data processing. John Wiley & Sons.
Rasol, M., Pais, J.C., Pérez-Gracia, V., Solla, M., Fernandes, F.M., Fontul, S., Ayala-Cabrera, D., Schmidt, F., Assadollahi, H., 2022. Gpr monitoring
for road transport infrastructure: A systematic review and machine learning insights. Construction and Building Materials 324, 126686.
Rathod, H., Debeck, S., Gupta, R., Chow, B., 2019. Applicability of gpr and a rebar detector to obtain rebar information of existing concrete
structures. Case Studies in Construction Materials 11, e00240.
Rodés, J.P., Pérez-Gracia, V., Martínez-Reguero, A., 2015. Evaluation of the gpr frequency spectra in asphalt pavement assessment. Construction
and Building Materials 96, 181–188.
Salinas Naval, V., Santos-Assunçao, S., Pérez-Gracia, V., 2018. Gpr clutter amplitude processing to detect shallow geological targets. Remote
Sensing 10, 88.
Santos, V.R.N.d., Al-Nuaimy, W., Porsani, J.L., Hirata, N.S.T., Alzubi, H.S., 2014. Spectral analysis of ground penetrating radar signals in concrete,
metallic and plastic targets. Journal of Applied Geophysics 100, 32–43.
Shaari, A., Ahmad, R., Chew, T., 2010. Effects of antenna-target polarization and target-medium dielectric contrast on gpr signal from non-metal
pipes using fdtd simulation. NDT & E International 43, 403–408.
Shaw, M., Millard, S., Molyneaux, T., Taylor, M., Bungey, J., 2005. Location of steel reinforcement in concrete using ground penetrating radar and
neural networks. Ndt & E International 38, 203–212.
Singh, N.P., Nene, M.J., 2013. Buried object detection and analysis of gpr images: Using neural network and curve fitting, in: 2013 Annual
International Conference on Emerging Research Areas and 2013 International Conference on Microelectronics, Communications and Renewable
Energy, IEEE. pp. 1–6.
Solla, M., Pérez-Gracia, V., Fontul, S., 2021. A review of gpr application on transport infrastructures: Troubleshooting and best practices. Remote
Sensing 13, 672.
Sun, H.H., Lee, Y.H., Luo, W., Ow, L.F., Yusof, M.L.M., Yucel, A.C., 2021. Compact dual-polarized vivaldi antenna with high gain and high
polarization purity for gpr applications. Sensors 21, 503.
Tomecka-Suchoń, S., 2019. Ground penetrating radar use in flood prevention. Acta Geophysica 67, 1955–1965.
Tong, Z., Gao, J., Yuan, D., 2020. Advances of deep learning applications in ground-penetrating radar: A survey. Construction and Building
Materials 258, 120371.
Torrione, P.A., Morton, K.D., Sakaguchi, R., Collins, L.M., 2013. Histograms of oriented gradients for landmine detection in ground-penetrating
radar data. IEEE transactions on geoscience and remote sensing 52, 1539–1550.
Wahab, W.A., Jaafar, J., Yassin, I.M., Ibrahim, M.R., 2013. Interpretation of ground penetrating radar (gpr) image for detecting and estimating
buried pipes and cables, in: 2013 IEEE International Conference on Control System, Computing and Engineering, IEEE. pp. 361–364.
Wang, Y., Cui, G., Xu, J., 2020. Semi-automatic detection of buried rebar in gpr data using a genetic algorithm. Automation in Construction 114,
103186.
Wu, K., Rodriguez, G.A., Zajc, M., Jacquemin, E., Clément, M., De Coster, A., Lambot, S., 2019. A new drone-borne gpr for soil moisture mapping.
Remote Sensing of Environment 235, 111456.
Zhang, X., Xue, F., Wang, Z., Wen, J., Guan, C., Wang, F., Han, L., Ying, N., 2021. A novel method of hyperbola recognition in ground penetrating
radar (gpr) b-scan image for tree roots detection. Forests 12, 1019.
Zhao, S., Al-Qadi, I., 2017. Pavement drainage pipe condition assessment by gpr image reconstruction using fdtd modeling. Construction and
Building Materials 154, 1283–1293.
Zhou, F., Chen, Z., Liu, H., Cui, J., Spencer, B.F., Fang, G., 2018a. Simultaneous estimation of rebar diameter and cover thickness by a gpr-emi
dual sensor. Sensors 18, 2969.
Zhou, X., Chen, H., Li, J., 2018b. An automatic gpr b-scan image interpreting model. IEEE Transactions on Geoscience and Remote Sensing 56,
3398–3412.

: Page 14 of 14