Stepped-Frequency CW GPR: Subsurface Imaging

• SFCW GPR is a subsurface imaging technique that transmits stepped narrowband continuous wave signals to synthesize a wide frequency bandwidth.
• It achieves high range resolution through inverse Fourier transform and advanced time-frequency methods such as STFT and depth-adaptive filtering.
• The technology is effective in challenging, high-loss environments, enabling precise detection of soil strata, roots, and underground utilities.

Stepped-Frequency Continuous Wave Ground Penetrating Radar (SFCW GPR) is a frequency-domain electromagnetic subsurface imaging modality that transmits a sequence of narrowband, single-frequency continuous-wave (CW) tones whose center frequency is incrementally (“stepped”) across a wide synthetic bandwidth. Sophisticated digital sampling of the received, scattered signal at each tone—together with inverse Fourier or compressive reconstruction—yields high-resolution, depth-resolved images of dielectric discontinuities such as roots, pipes, or soil strata. SFCW GPR is distinguished by its precise frequency control, tunable energy per tone, high phase stability, immunity to front-end filter ringing, and robust performance in challenging, attenuative environments such as high-moisture, high-permittivity tropical soils.

1. Fundamental Principles and Signal Theory

The SFCW GPR system emits a sequence of monochromatic signals:

$$s(t) = \sum_{k=0}^{N-1} A \cos \left[2\pi (f_0 + k\Delta f)t \right] u_k(t)$$

where $A$ is amplitude, $\Delta f$ is the frequency step, $N$ is the number of steps, and $u_k(t)$ is a rectangular dwell window for the $k$th tone. The total synthesized bandwidth is $B = N\,\Delta f$; this determines the achievable two-way range resolution:

$\Delta R = \frac{c}{2B\sqrt{\varepsilon_r}}$

where $c$ is the speed of light and $\varepsilon_r$ is the relative permittivity of the medium (Luo et al., 2022, Luo et al., 2023, Xu et al., 2024).

At each $f_k = f_0 + k\Delta f$, the receiver synchronously captures the complex scattering coefficient, denoted $S_{21}(f_k)$. After acquiring the full frequency suite, a discrete inverse Fourier transform (IFFT) reconstructs the time-domain impulse response:

$$g[n] = \sum_{k=0}^{N-1} S_{21}(f_k) e^{j 2\pi k n/N}$$

The resulting A-scan presents the range profile; spatially stacking these yields a B-scan image across the survey path (Luo et al., 2022, Xu et al., 2024).

2. System Architectures and Key Parameters

Frequency Planning and Resolution

Typical SFCW GPR implementations operate over gigahertz bandwidths (e.g., 0.2–4.0 GHz with $N=1001$ steps and $\Delta f \approx 3.8$ MHz for tree root detection (Luo et al., 2022), or 1.3–2.9 GHz with $N=400$ steps and $\Delta f = 4$ MHz for soil analysis (Xu et al., 2024)). Range resolution scales inversely with the synthesized bandwidth and the square root of soil permittivity, e.g.,

Scenario / Soil Condition	Bandwidth (GHz)	$\varepsilon_r$	$\Delta R$ (m)
Free space	3.8	1	$\approx$ 0.04
Sandpit	3.8	4	$\approx$ 0.02
High-moisture tropical soil	3.8	13	$\approx$ 0.011

Instrument examples include Vector Network Analyzers (VNAs) with dual-polarized Vivaldi antennas (Tx/Rx separation $\sim$0.1–0.6 m, housed in absorber-lined enclosures for environmental noise immunity) (Luo et al., 2022, Xu et al., 2024). Intermediate frequency bandwidth (IFBW) settings directly affect SNR and sweep times (e.g., IFBW = 1 kHz or lower to suppress noise in high-loss soils) (Luo et al., 2020).

3. Advanced Processing: Joint Time–Frequency and Adaptive Filtering

Time–Frequency (JTFA) and STFT

To address the severe attenuation of high-frequency tones in lossy soils, SFCW GPR research leverages joint time–frequency analysis (JTFA), most commonly employing the Short-Time Fourier Transform (STFT):

$$\mathrm{STFT}\{x\}(t, f) = \int_{-\infty}^{\infty} x(\tau) w(\tau - t) e^{-j2\pi f \tau}\, d\tau$$

where $w$ is a windowing function (commonly Hamming, length $=\frac{1}{10}$ A-scan, overlap at 50%) (Luo et al., 2022, Luo et al., 2023).

Depth-Adaptive Filtering (DATFF)

A major advance is the DATFF method (Luo et al., 2023), which adapts the frequency cutoff for each depth/time index by (1) forming a 2D STFT magnitude histogram, (2) extracting a cutoff curve via weighted linear regression, and (3) applying a binary mask in the frequency-to-time conversion (inverse DFT or variant chirp Z-transform):

$$x[n] = \sum_{k=0}^{N-1} S_{21}[k]\, H[n, k]\, e^{j 2\pi k n/N}$$

where $H[n,k]=1$ for $k \leq \beta_0 + \beta_1 \tau(n)$. This fully data-driven window excludes frequencies dominated by deep-soil attenuation without requiring prior knowledge of soil permittivity, thus maximizing SNR at every depth (Luo et al., 2023).

STFT Filtering Workflow

The general pipeline is:

1. Pre-process A-scans (zero-offset, time-zero, SVD-based background removal)
2. Compute STFT
3. Identify maximal frequency at each time index
4. Extract frequency cutoff curve (e.g., via weighted linear regression)
5. Apply adaptive mask/thresholding in the frequency domain
6. Inverse transform to synthesize enhanced B-scans

This approach exposes hyperbolic root reflections buried beneath clutter in full-band images and provides substantial SNR gains (e.g., +6–7 dB for deep roots (Luo et al., 2022)).

4. Compressed Sensing and Sampling Optimization

Sparse Recovery in SFCW GPR

SFCW GPR naturally aligns with sparse recovery paradigms. The linear measurement model $y = \Phi h + n$ exploits the physical sparsity of subsurface reflectivity, enabling robust recovery in presence of missing or jammed frequency steps (Hu et al., 2010). Convex $\ell_1$-minimization or greedy algorithms (e.g., OMP) can reconstruct the extended-range profile across multiple range bins, offering higher resilience to data loss and lower sidelobes compared to classical stretch processing (Hu et al., 2010, 0904.1910).

Sampling Strategies

Optimal acquisition speed and reconstruction fidelity depends on the distribution of the signal's spectral energy. Uniform (frequency-equipartition) sampling, random frequency selection, and energy-equipartition sampling (EES) are compared empirically (0904.1910):

Method	PSNR (M = 20, single monocycle)	Implementation Notes
FES (Uniform)	–18.2 dB	Ignores spectral energy variations
Random	–13.5 dB	Typically better than FES but may waste power
EES	23.7 dB	Exploits prior $S(f)$, yields fastest acquisition

EES delivers the best reconstruction for a given number of frequency steps, accelerating surveys and reducing front-end power demand (0904.1910).

5. SFCW GPR in Challenging Environments: Empirical Validation

High-Permittivity, High-Loss Soils

In tropical and high-moisture soils ($\varepsilon_r\sim 13-27$), SFCW GPR performance is dictated by attenuation, frequency-dependent SNR, and bandwidth choice. Experimental results demonstrate:

• In sand ($\varepsilon_r\approx4$), two-way resolution is $\sim$2 cm, and metal objects are detected to 30 cm depth.
• In field tests, roots identified with STFT-enhanced SFCW GPR are resolved at twice the maximum depth versus unfiltered full-band processing (40 cm vs. 20 cm) with lateral accuracy $<$5 cm and depth error $\pm$3 cm (Luo et al., 2022).
• For high-contrast but lossy targets (e.g., tree roots), narrowing IFBW from 500 Hz to 50 Hz can yield $\sim$10 dB SNR improvement, dramatically clarifying root hyperbolae (Luo et al., 2020).

STFT-based and depth-adaptive time-frequency filtering methods consistently improve SNR and detection rate, with reported root detection rates increasing from 70% to 95% in repeated scans and SCR improvements (e.g., baseline SCR $-0.30$ dB to $+0.10$ dB with DATFF) (Luo et al., 2022, Luo et al., 2023).

6. Data-Driven and Machine Learning Enhancements

Machine learning advances the interpretation of SFCW GPR data, particularly for soil parameter inversion:

• Large-scale field surveys with air-coupled SFCW GPR and electromagnetic induction (EMI) mapping enable regression of apparent electrical conductivity from the GPR frequency response (Xu et al., 2024).
• Features (e.g., 400-dimensional GPR sweeps) are fed to linear, random forest, or $k$–nearest neighbor regressors.
• Performance metrics (MAE, MSE, Pearson’s $r$) reach $r$ up to 0.43 for intra-field prediction; spatial coherence assessed by nugget-to-sill ratio in variogram fits aligns with error metrics.

This suggests SFCW GPR can supplement or partially replace EMI for agricultural mapping tasks requiring high lateral or vertical resolution. Limitations are observed in reduced dynamic range and generalization under domain shift, motivating future sensor fusion and expanded training sets (Xu et al., 2024).

7. Limitations, Practical Guidelines, and Future Directions

Principal limitations of SFCW GPR relate to frequency-dependent attenuation (ultimate penetration depth $\sim$50 cm in water-rich soils), computational demands of real-time JTFA, and parameterspecific site adaptation. Practical implementation requires careful antenna engineering (absorption-lined enclosures, polarization control) and real-time data pipelines optimized for GPU/FPGA acceleration (Luo et al., 2022, Luo et al., 2020).

Future enhancements encompass:

• Adaptive windowing and wavelet JTFA to further improve discrimination of weak/deep targets
• Automated STFT-feature classification by machine learning for clutter rejection
• Multi-static and polarimetric array configurations for three-dimensional root zone mapping and permittivity estimation
• Full integration of SFCW GPR with cloud-based ML model updates for agricultural or arboricultural operations (Luo et al., 2023, Xu et al., 2024)

SFCW GPR, particularly in combination with advanced time-frequency and adaptive filtering, establishes a robust framework for high-resolution, noise-tolerant subsurface imaging in diverse and challenging environments (Luo et al., 2022, Luo et al., 2023, Xu et al., 2024, Luo et al., 2020, Hu et al., 2010, 0904.1910).