NOISE ATTENUATION FROM 3D GPR DATA USING ARTIFICIAL NEURAL NETWORK
Sid‐Ali Ouadfeul, Leila Aliouane · Symposium on the Application of Geophysics to Engineering and Environmental Problems 2014 · 2014
In this paper, a tentative of noise attenuation from the 3D Ground Penetrating Radar data (GPR) data using the Multilayer Perceptron neural network model is implanted. GPR data recorded in Algerian Sahara are filtered firstly using the discrete wavelet transform, after that a MLP machine with three layers is trained in a supervised learning mode, the input is an extracted profile from the raw 3D GPR data and the output consists of the GPR data of the same profile but after filtering. The estimated weights of connection are used to propagate the remaining non-filtered data through the implanted machine, the calculated output consists to the filtered GPR data. Comparison between the filtered data using the MLP machine and the continuous wavelet transform shows that the neural network machine can be used for S/N ratio improvement of the noisy GPR data. Introduction Ground Penetrating Radar (GPR) data processing using the wavelet transform has becoming a very interesting subject of research. Ouadfeul and Aliouane (2010) have published a paper that use the wavelet transform for identification of obstacles direction by the 3D GPR data using the wavelet transform. The analyzing wavelet is the Mexican Hat. The proposed method shows its robustness and useful in the 3D seismic design. Ouadfeul and Aliouane (2012) have shown the sensitivity of the wavelet transform to random noise and they have proposed to apply a filter to the 2D wavelet coefficients for small scales. In this paper, we propose another method to filter GPR data from random noise, it is based on the use of the neural network and discrete wavelet transform. Discrete wavelet transform and signal denoising . The function ) (t is said to be a wavelet if and only when the following condition is satisfied (Ouadfeul et al, 2012): 0 ) ( dt t The wavelet transform of a function ) 2 ( 2 ) ( R L t is defined by (Ouadfeul et al, 2012, Ouadfeul and Aliouane, 2013b) : ) ( * ) ( ) ( t a t f t f a