Automated Remote Sensing of Underground Features

| |


While many advancements have  been made this last decade in automated classification of above surface features using remote sensing data, progress for detecting underground features has lagged in this area. Technologies for detecting features, including ground penetrating radar, electrical resistivity, and magnetometry exist, but methods for feature extraction and identification mostly depend on the experience of instrument user.

One problem has been creating approaches that can deal with complex signals. Ground penetrating radar (GPR), for instance, often produces ambiguous signals that can have a lot different noise interference relative to the feature one wants to identify. One approach has been to apply approximation polynomials to classify given signals that are then inputs for an applied neural networks model using derived coefficients. This technique can help reduce noise and differentiate signals that follow clear patterns that vary from larger background signals. Differentiation of signals based on minimized coefficients are one way to simplify and better differentiate data signals.[1] Another approach is to use multilayer perceptron that has a nonlinear activation function which transforms the data. This is effectively a similar technique but uses different transform functions than other neural network models. Applications of this approach include being able to differentiate thickness of underground structures from surrounding sediments and soil.[2]

GPR (Ground Penetrating Radar) system. Figure: Szymczyk & Szymczyk, 2019.
GPR (Ground Penetrating Radar) system. Figure: Szymczyk & Szymczyk, 2019.

Other methods have been developed to determine the best location to place source and receivers that can capture relevant data. In seismic research, the use of convolutional neural networks (CNNs) has been applied to determine better positioning of sensors so that better data quality can be achieved. This has resulted in very high precision and recall rates at over 0.99. Using a series of filtered layers, signals can be assessed for their data quality with that of manually placed instruments. The quality of the placement can also be compared to other locations to see if the overall signal capture improves. Thus, rather than focusing on mainly signal processing, this method also focuses on signal placement and capture that compares to other placements to optimize data capture locations.[3] One problem in geophysical data is inversion, where data points are interpreted to be the opposite of what they are due to a reflective signal that may hid the nature of the true data. Techniques using CNNs have also been developed whereby the patterning of data signals around a given inversion can be filtered and assessed using activation functions. Multiple layers that transform and reduce data to specific signals helps to identify where patterns of data suggest an inversion is likely, while checking if this follows patterns from other data using Bayesian learning techniques.[4]

Increasingly, mapping underground features is becoming a safety issue in communities. Buried utilities are not always known or have been forgotten. One approach integrates sensor data and old map data to estimate likely areas in order to create a Bayesian derived map of underground utilities. In this approach, image segmentation was used from map data, while Bayesian statistics are used to classify given areas. This then helps to create a derived map and classification system of underground features that can be visualized for likely areas where old utilities might be located.[5]

Blue lines represent simulated utility record based on manhole and sensor readings. Red lines represent segment-manhole connections established by the model.
Blue lines represent simulated utility record based on manhole and sensor readings. Red lines represent segment-manhole connections established by the model. Source: Bilal et al., 2018.

Increasingly, mapping underground features is not only an issue of potentially safety but geophysics and other sensor-derived data have now begun to hit the era of ‘big data.’ Similar to above ground remote sensing, below ground mapping can now better utilize these remotely sensed data to create more accurate maps and automating the understanding of features, while also removing complex noise signals so common with geophysical data. These techniques have not been applied at a mass scale but this may just be a matter of time before the geophysical community catches up with other researchers in the remote sensing community.


[1]    For more on the use of neural networks to differentiate and assess input coefficients from GPR signals, see:  Szymczyk M and Szymczyk P (2019) Neural networks based method for automatic classification of GPR data. In: COMPUTATIONAL TECHNOLOGIES IN ENGINEERING (TKI’2018): Proceedings of the 15th Conference on Computational Technologies in Engineering, Jora Wielka, Poland, 2019, p. 020014. DOI: 10.1063/1.5092017.

[2]    For more on multilayer perceptron neural network approach, see:  Sukhobok YA, Verkhovtsev LR and Ponomarchuk YV (2019) Automatic Evaluation of Pavement Thickness in GPR Data with Artificial Neural Networks. IOP Conference Series: Earth and Environmental Science 272: 022202. DOI: 10.1088/1755-1315/272/2/022202.


[3]    For more on location choice for relevant data capture points using CNN, see:  Jiang W, Zhang J and Bell L (2019) 3D seismic geometry quality control and corrections by applying machine learning. GEOPHYSICS 84(6): P87–P96. DOI: 10.1190/geo2018-0617.1.

[4]    For more on inversion differentiation, see:  Liu Bin, Guo Q, Li S, et al. (2019) Deep Learning Inversion of Electrical Resistivity Data. arXiv:1904.05265 [physics]. Available at: (accessed 10 December 2019).

[5]    For more information on using Bayesian learning for mapping and classification of underground utilities, see:  Bilal M, Khan W, Muggleton J, et al. (2018) Inferring the most probable maps of underground utilities using Bayesian mapping model. Journal of Applied Geophysics 150: 52–66. DOI: 10.1016/j.jappgeo.2018.01.006.



Open Repository for Geospatial Training Data Released

What is a GeoTIFF?