Reducing Ambiguities in the Measured Vertical Electrical Sounding Data by Using Artificial Neural Networks

The methods were tested on synthetic VES data as well as on field data collected from the central part of Egypt. This work exhibits adaptability of ANN in smoothing measured apparent resistivity data involving a variety of curve types and quality. Good results are observed with the three layer type curves with very small network errors. At the same time, the network errors rise with higher order of layers as obtained in the case of five curve type. As the network is over-trained, the weights will try to adjust to the minor details of the training data set itself. Though the problem gets complex with increase in layers, it is found that a proper network design can solve it, provided a good representative database for the training. The special advantage offered by ANN for resistivity inversion is that once the network is trained, it can perform the smooth of any VES data set very rapidly. The result from a neural network can be used as a starting model for inversion to decrease the inversion time. The technique can also be developed for assisting with first hand information for the model initialization in other conventional inversion schemes.