- Copyright: © 2006 This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media for any purpose, provided this notice is preserved along with the article's original URL.
The problem of fresh water availability in coastal aquifers is a reality. For in-situ and dynamic characterization of seawater encroachment into coastal aquifers, electrical geophysical methods are better suited. Vertical Electrical sounding (VES) when combined with induced polarization soundings (IPS) can resolve saline sands from moist clays. Our feed forward back-propagation neural network (BPNN) based approach automates the analysis of combined vertical electrical and induced polarization soundings to suit practical needs.
Our method is initially tested on synthetic data computed from available geo-electric sections and geological information concerning coastal aquifers of the East Coast of India.
The synthetic data comprised 18 combined Schlumberger IPS and VES soundings (504 apparent resistivity and chargeability samples) spread over five profiles in the study region. Fictitious apparent resistivity (product of apparent resistivity and apparent chargeability) soundings are derived from them. We used 118 carefully selected discrete fictitious apparent resistivity values from 210 sample sets gathered from 15 (420 samples) combined soundings to train the BPNN, while 33 samples from 3 separate combined soundings, and 26 random samples from 92 unused training samples of 15 soundings were used for testing. Our trained BPNN involved one input node and one bias-unit at the input layer stage, one node in the output layer, and 18 nodes and one bias-unit in hidden layer. The trained neural net showed an overall success rate of 83% in testing phase for distinguishing clays from saline sands in the synthetic example.
Our method is also tested on real data concerning a shaly groundwater aquifer in Bahia, Brazil yielding an overall accuracy of 85%, quite comparable to that of synthetic case. Thus, both synthetic and field data analysis validate our neural network based algorithm.