Ground Penetration Radar (GPR) as a nondestructive method for identifying underground objects has been successfully applied to different fields of science such as geotechnical investigations, oil and gas exploration, geology, pipe detection and archeology investigations. Metallic and nonmetallic objects can be identified by this method. The depth of penetration is dictated by the GPR antennas. Low frequency antennas (from 25-200 MHz) explore materials from deeper depths in the low cost resolutions. High-frequency antenna (>200 MHz) obtains reflections from shallow depths with higher resolutions. Ground penetrating radar is considered as the most suitable approach to detect shallow buried objects. Transmitter and receiver antenna are closely spaced together and can detect changes in the electromagnetic properties of an object. Electromagnetic waves are transmitted through an antenna and the reflected waves form various buried objects or contacts between different materials are received and stored in digital control unit. Antenna shielding is performed to eliminate interferences from other intruder sources. Electromagnetic waves are emitted by the transmitting antenna and distorted by the soil conductivity variation, dielectric permittivity, and magnetic permeability. The reflected waves are recorded by the receiving antenna in nanoseconds. The shape of GPR radargrams, vertical map of the radar reflection returned from subsurface objects, of cylindrical objects is similar to a hyperbola. Interpretation of acquired GPR data needs an expert geoscientist with a lot of knowledge and time.
The classical Hough transform is a common method for identification of buried objects (Capineri et al. 1998, Simi et al. 2008). However, this method is time-consuming and computationally expensive. Alternatively, artificial neural network used by some authors (Al-Nuaimy et al. (2000), Gamba and Lossani (2000)), but these methods also need many training data to gain high accuracy and producing such data is difficult. Genetic optimization algorithm has been applied by Pasoli (2009) to detect the hyperbolic objects in GPR images. Local search of original genetic algorithm is poor. Chen and Cohn (2010) have presented a method for detecting hyperbola shapes based on probabilistic mixture model. However, the method is computationally expensive and is not robust with respect to noise.
In the current study, a modified genetic optimization algorithm has been applied to GPR sectional images for identifying hyperbola signatures of small-buried objects (mainly pipes and channels). The performance of genetic algorithm highly depends on genetic operators. Arithmetic crossover is used to improve the local search ability and point-wise crossover is applied to explore new regions. The hyperbolas are searched through the edge image resulting from an image pre-processing step. Hyperbola detection is achieved with sub-pixel accuracy. After identifying each hyperbolic object, the object is removed from the image and algorithm searches for new possible hyperbolas in the GPR image.
The performance of proposed method is evaluated using synthetic and real data. The synthetic data were generated with GprMax 2D, a computer program that generates GPR images using an electromagnetic simulator, based on the finite-difference time-domain (FDTD) method in 2D, and real data surveyed in campus of Isfahan University of Technology (IUT). Some preprocessing steps including dewow filtering and removing DC bias, background removal, manual gain function, and image thresholding were applied to the data, before employing the proposed method. Then, hyperbola parameters extracted using a modified genetic algorithm. Depth and radius of the buried object were estimated by hyperbola parameters. The results show that the proposed method gains high accuracy in estimating depth and radius of buried objects.