Optimising thresholding of bandpass filter through GPR wavelet transform in time-frequency domain

Bandpass filter is a critical GPR signal processing step for enhancing visibility of buried objects. But its setting of thresholds of upper and lower limits are subjective. This paper describes a more objective method for setting of the thresholds by Gabor Transform in time-frequency domain for determining the frequency thresholds after identifying the actual frequency responses of objects of interests, while others like direct wave and noises in other time window are not considered. The method was tested with the data extracted from a real case study with a Proceq GS8000 dual frequency Ground Penetrating Radar. Among the high-frequency data, four-A scans were selected from the same B-Scan with a deep object, a shallow pipe, reinforcement rebar and a last scan in natural soil (with no utilities or other structural elements). In the time domain, standard filtering processes were applied in both 2D and 3D spaces to identify the targets and enhance the imaging of the radagram. In the frequency domain, each target at different time of reflection lead to a different response or distribution in the frequency spectrum trough the fast Fourier transform (FFT). Depending on each element and respective material type, the frequency spectrum distribution could lead to a specific response pattern and relative amplitude over a 1D array. The Gabor Wavelet Transform could segregate the direct and reflected waves and therefore permit to interpret in a contour map (2D) the behaviour of the frequency exactly where the target is located, and allows for setting of the frequency thresholds for bandpass filter. There are observed patterns that could be useful to discriminate and categorise specific targets based on the amplitude and shape. From each (wavelet transform) spectrogram, the frequency spectrum was extracted and compared with the full spectrum delivered from the FFT. Results could enhance the resolution and improve the location of the target by determining the ideal band pass filter (with respect to low and high cut-off frequencies) for each target, instead of the traditional band pass filter applied to the whole radagram. It also alleviates the cognitive bias problem “I want to show what I want to show” which is used to be a manual and operator-dependent process.