Fulltext

Two Supervised Neural Networks for Classification ofSedimentary Organic Matter Images fromPalynological Preparations

Weller, Andrew ; Harris, Anthony ; Ware, J.

In: Mathematical Geology, 2007, vol. 39, no. 7, p. 657-671

Add to personal list
    Title
    • Two Supervised Neural Networks for Classification ofSedimentary Organic Matter Images fromPalynological Preparations
    Author
    • Weller, Andrew. Geological Institute, Department of Earth Sciences, ETH Zurich, 8092, Zurich, Switzerland
    • Harris, Anthony. School of Applied Sciences, University of Glamorgan, CF37 1DL, Pontypridd, UK
    • Ware, J.. School of Computing, University of Glamorgan, CF37 1DL, Pontypridd, UK
    Document Type
    • Postprint
    OAI-PMH Identifier
    • oai:doc.rero.ch:320786
    Summary
    An improvement in the supervised artificial neural network classification of sedimentary organic matter images from palynological preparations is presented. Sedimentary organic matter encompasses the entire acid-resistant organic micro-particles (typically with a diameter of 5-500μm) recovered from a sediment or sedimentary rock. Supervised neural networks are trained to recognize patterns within databases for which the correct classifications are already known. Once trained, they are verified on pre-classified samples not seen by the network, and then used for classification of samples whose class is not known. Such networks have an input, hidden and output layer. Typically, these networks determine what the output class is by adjusting weights associated with the layer interconnects, and by modifying the signals that propagate through the hidden layer by a non-linear transfer function. In this example, the inputs in each network are the salient features selected from an available set of 194, while the outputs are the sedimentary organic matter classifications which were formerly developed with the rationalization of descriptive terms from previous classification schemes. The author's past work tested the supervised back propagation neural network for the classification of sedimentary organic matter images. This gave an overall correct classification rate of 87%. However, because the back propagation network underperformed on two of the four classes, the radial basis function neural network was tested on the same databases initially used in an attempt to improve the recognition rate of these two classes. The difference between the back propagation and radial basis function networks lies in the non-linear transfer function applied in the hidden layer, which was modified by a Gaussian function in the latter. In the best-case scenario, this improved the recognition rate by 4% to just over 91%. This has also determined that a series of different supervised neural networks may be better for classification of sedimentary organic matter images. These results are encouraging enough to prompt further research that may result in a commercially viable system