Multisource data integration with neural networks: optimal selection of net variables for lithologic classification

Different types of images generated from gravity, magnetic, gamma ray spectrometry and remote sensing images such as Landsat Thematic Mapper, radar and SPOT are available in Melville Peninsula, N.W.T. to delineate geological patterns as an aid to geological field mapping in Arctic regions in Canada. Feedforward neural networks were trained to formulate mapping classifiers to predict the lithologic units. Through the analysis of classification accuracy with increased number of iterations, the authors demonstrated that the optimal choice of input layers is the most sensitive factor in achieving better accuracy result. The classification accuracy may be maximized by choosing an optimal combination of input data layers. The complexity of the training task which include´s the selection of the training samples, the number of training samples, are critical for a satisfactory classification. The classification accuracy is inversely proportional to the number of output classes. The overall average accuracy of classification gets better by increasing the number of iterations to a certain degree, however, at the expense of some individual classification accuracy. The variance in the individual classification accuracy were found to be significant which has led to some criterion on the selection of net variables. For lithologic mapping, the network should be structured in accordance with the importance of each individual class