Study On Artificial Neural Networks And Their Applications In Geoscience

There are a lot of methods used in the geoscience information process. They arechosen by the characters of the variable and data. When there are many variables, thatis, the relation between the phenomenon and essence of geoscience is morecomplicated, if the variables are quantitative, we use the multivariable statisticalanalysis and so on, and if the variables are qualitative, we use the theory ofquantification and otherwise.When there are many variables, and if the variables have not any statisticaldistribution, and have some non-linear characters, then we had better use thenon-linear analysis methods, such as the artificial neural networks and so on.Because the origin, evolvement and development of the earth is a unrepeatableprocess, and because the science and technology development history of the mankind,the geoscience data have the characters of multi-dimension, multi-time, multi-accuracy,multi-scale and multi-solve. This results in that the essential relation between theobservation data and the research objects has the non-linear characteristic. Therefore,the geoscience data should be studied by the non-linear methods which are developedin the recent years.We should ensure that the information is not distorted in the information process.If we use the linear mathematical methods to handle the non-linear data, we wouldcover up the non-linear characters of the information. Because the linear is thedegradation of the non-linear, it is reliable to process the information by the non-linearmethods.In the non-linear methods used in the geoscience information process currently,we choose some artificial neural networks to carry out the non-linear geoscienceinformation process. The thesis studies the mathematical models, algorithms andprograms of the artificial neural networks to provide the technique support for thenon-linear geoscience information process.The thesis studies the principles, algorithms and applications of the artificialneural networks. For meeting the request of the geoscience applications, we write thealgorithms, program the practical programs and apply the programs combining thegeoscience examples. We develop the practical algorithms and programs of thebackpropagation network, the radial basis function network, the Hopfield network andthe self-organizing feature map network which are used in geoscience and provide theexamples for the geoscience applications.The thesis discusses the research development history and the foundation theoriesof the artificial neural networks and discusses the foundation theories demanded tostudy the backpropagation network emphatically.The thesis discusses the theory of the basic backpropagation algorithm in detail,especially makes a succinct and strict mathematical deduction for it and discusses itslimitations. However, the basic backpropagation algorithm is too slow for mostpractical applications, so we discusse some variations of it, which provide significantspeedup and make the algorithm more practical. We discusse the methods of thebackpropagation network design. Then we discusse some modifications of the basicbackpropagation algorithm and develop a program on the base of the improvedbackpropagation algorithms. Next, on the base of the widely study of the improvedbackpropagation algorithms, we select the resilient backpropagation (RPROP)algorithm to develop the "Resilient Backpropagation Network Program". We describethe resilient backpropagation network program in detail, including the theory of theresilient backpropagation algorithm, its detailed algorithm, its program and itsapplications.We choose the resilient backpropagation network as the quantitative predictiontool of the mineral resources. We take the Au deposit concentrated regions in GuizhouProvince as the model units, and the quantitative reserve prognosis of the Au anomalyconcentrated regions is carried out by the resilient backpropagation network. It has acertain actual value.We use the resilient backpropagation network to carry on the igneous rockclassification, and the result is good. In this application, we select 247 examples' dataas the training set. There is 361 iterations used in a training which takes 9 seconds andthe accuracy is 10-7. If needed, the accuracy can reach 10-9. So the resilientbackpropagation network is rapid and practical.The thesis verifies the occurrence of the El Nino and the La Nina event by theresilient backpropagation network. According to the times of the yearly global abovelevel-7 earthquake, the conditions of the solar eclipses and the average data of the seatemperature matrix in 1973-1986, applying the resilient backpropagation network, weverify the occurrences of the El Nino and the La Nina event during 1987-2000. Theaccurate rate is 71.4%.The thesis discusses another kind multilayer feedforward neural network, which isthe radial basis function (RBF) network. It is designed to solve the question of thecurve surface approach in a high dimension space. The thesis discusses the principle ofthe RBF network and develops its program. Taking the geochemical exploration datain a place of Jilin Province as an example, the author experiments the interpolationeffect by the RBF network in the different data absent degree conditions. The RBFnetwork is fit for the reconstruction of the geological curve surface and has not specialrequest in the distribution of the original data and the boundary condition of thegeological curve surface. So it is fit for the reconstruction of the geological curvesurface which is destroyed or has less exploring data.The Hopfield network is a kind of recurrent neural network what is studied bestand has the most extensive application. The thesis discusses the principle of theHopfield network, works out its detailed algorithm, develops the Hopfield networkprogram and gives its applications. The Hopfield network is applied in the associativememory and the optimization calculation.The self-organizing feature map (SOFM) network is a competitive network whatcan carry on the unsupervised self-organizing learning in training. The thesis discussesthe principle and the algorithm of the SOFM network and develops its program. Thethesis introduces a logging lithological identification technology based on the SOFMnetwork. An example is used to show how to build up an SOFM network model of thelogging lithological identification and to show its application to the logginglithological identification. The result indicates that the accuracy of identification ishigh and it can be used in the lithological identification of the logging data. The SOFMnetwork has the self-organizational ability and self-adaptability, and it canautomatically cluster the learning samples. It has a stronger permitting wrong abilityand has high identification accuracy to the unknown samples.Finally, the thesis narrates the artificial neural network software developed by usand discusses the characteristics of the artificial neural networks and their appliedcharacteristics in geoscience. The thesis introduces the other type artificial neuralnetworks, the present researches and applications of the artificial neural networks andpredicts their future.The thesis studies on applying the artificial neural network methods to process thenon-linear geoscience information. For meeting the request of the geoscienceapplication, we develop the algorithms and programs of the artificial neural networks,study the efficient methods of the artificial neural networks used in geoscience,summarize their applied characteristics and offer their applied examples.