A Study on 2D similarity transformation using multilayer perceptron neural networks and a performance comparison with conventional and robust outlier detection methods


Konakoglu B. , GÖKALP E.

Acta Montanistica Slovaca, cilt.21, ss.324-332, 2016 (SCI İndekslerine Giren Dergi) identifier identifier

  • Cilt numarası: 21 Konu: 4
  • Basım Tarihi: 2016
  • Dergi Adı: Acta Montanistica Slovaca
  • Sayfa Sayıları: ss.324-332

Özet

One of such transformation methods, the two-dimensional similarity transformation, is widely used in geodetic studies. The outliers in the measure group should be detected so that the model established during the transformation process gives accurate results. In transformation practices, conventional outlier detection test procedures based on the least squares estimation (LSE) and robust estimation methods are widely used for the detection of outliers. The aim of this study was to compare performances of the result data obtained by multilayer perceptron artificial neural networks (MLPNNs) including various activation functions and training algorithms of artificial neural networks (ANNs), which has recently begun to be widely used in scientific studies and engineering fields, and of the result data obtained using various methods to detect outliers in two-dimensional similarity transformation process between two different coordinate systems. In ANNs consisting of three layers, eight different network configurations were generated using different activation functions and training algorithms. The coordinates of the control points calculated by two-dimensional similarity and ANNs methods were compared with known coordinate values. Differences between the coordinates calculated using two-dimensional similarity transformation and eight different network configurations and the coordinates of control points were examined in terms of the root mean square error, and network configuration which uses a combination of ' tansig-purelin' activation functions and Bayesian regulation algorithm provided the best result.