Artificial neural network inversion of magnetotelluric data in terms of three-dimensional earth macroparameters

Details Artificial neural network inversion of magnetotelluric data in terms of three-dimensional earth macroparameters Artificial neural network inversion of magnetotelluric data in terms of three-dimensional earth macroparameters

Jul 2000

adsabs.harvard.edu/cgi-bin/nph-data_query?bibcode=2000geoji.142...15s&link_type=abstract

Geophysical Journal International, Volume 142, Issue 1, pp. 15-26.

Astronomy and Astrophysics

Astronomy

16

3-D Inversion, Dyke, Magnetotellurics, Neural Network

Scientific paper

The possibility of solving the three-dimensional (3-D) inverse problem of geoelectrics using the artificial neural network (ANN) approach is investigated. The properties of a supervised ANN based on the back-propagation scheme with three layers of neurons are studied, and the ANN architecture is adjusted. A model class consisting of a dipping dyke in the basement of a two-layer earth with the dyke in contact with the overburden is used for numerical experiments. Six macroparameters of the 3-D model, namely the thickness of the top layer, which coincides with the depth of the dyke (D), the conductivity ratio between the first and second layers (C1/C2), the conductivity contrast of the dyke (C/C2), and the width (W), length (L) and dip angle of the dyke (A), are used. Various groups of magnetotelluric field components and their transformations are studied in order to estimate the effect of the data type used on the ANN recognition ability. It is found that use of only the xy- and yx-components of impedance phases results in reasonable recognition errors for all unknown parameters (D: 0.02 per cent, C1/C2: 8.4 per cent, C/C2: 26.8 per cent, W: 0.02 per cent, L: 0.02 per cent, A: 0.24 per cent). The influence of the size and shape of the training data pool (including the `gaps in education' and `no target' effects) on the recognition properties is studied. Results from numerous ANN tests demonstrate that the ANN possesses good enough interpolation and extrapolation abilities if the training data pool contains a sufficient number of representative data sets. The effect of noise is estimated by means of mixing the synthetic data with 30, 50 and 100 per cent Gaussian noise. The unusual behaviour of the recognition errors for some of the model parameters when the data become more noisy (in particular, the fact that an increase in error is followed by a decrease) indicates that the use of standard techniques of noise reduction may give an opposite result, so the development of a special noise treatment methodology is required. Thus, it is shown that ANN-based recognition can be successfully used for inversion if the data correspond to the model class familiar to the ANN. No initial guess regarding the parameters of the 3-D target or 1-D layering is required. The ability of the ANN to teach itself using real geophysical (not only electromagnetic) data measured at a given location over a sufficiently long period means that there is the potential to use this approach for interpreting monitoring data.

Affiliated with

Also associated with

No associations

Rating

Artificial neural network inversion of magnetotelluric data in terms of three-dimensional earth macroparameters does not yet have a rating. At this time, there are no reviews or comments for this scientific paper.

If you have personal experience with Artificial neural network inversion of magnetotelluric data in terms of three-dimensional earth macroparameters, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Artificial neural network inversion of magnetotelluric data in terms of three-dimensional earth macroparameters will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFWR-SCP-O-1719333