Development of ANN optimized affine-6 2D coordinate transformation model

Abstract

Coordinate transformation facilitates the integration of geodetic coordinates of points obtained from different sources into a common geodetic reference frame. In existing studies, mathematical transformation models such as Bursa-Wolf, Molodensky-Badekas, Veis, the affine transformation models and others have been applied. These models can lead to low accuracy, due to various factors, such as lack of understanding of the distortions and inconsistencies of the local datum and geodetic network distribution. Recently, Artificial Neural Networks (ANN) techniques for coordinate transformation have been evaluated in several countries and have been found to achieve better results compared to similarity models. In Kenya, there is little literature on the evaluation of these techniques for improving coordinate transformation. Therefore, this study aims to optimise the affine six-parameter 2-dimension coordinate transformation using ANN techniques. The methodology involves acquisition and processing of geodetic control datasets with common points in two coordinate systems: UTM and Cassini Arc 1960 for part of the Nyeri-Kirinyaga geodetic network, in Central region of Kenya. The Affine-6 transformation parameters are determined, applied for coordinate transformation and the distortions modelled. The transformation resulted in relatively low accuracy, possibly due to the limited ability of the model to map nonlinear patterns in the datum. This study proposed application of nonlinear ANN models; Multi-Layer Perceptron (MLP), and Radial Basis Functions Neural Network (RBFNN) to map the non-linear patterns and adjust the transformed coordinates, hence optimizing the Affine-6 model. A comparative evaluation was performed to determine the improvement in performance and compare the models. It was found that the ANN techniques improved the Affine-6 transformation by 92.55% and 92.27% in RMSE and 99.35%, 98.06% in horizontal error for MLP and RBFNN respectively.