ARTIFICIAL NEURAL NETWORKS PRUNING APPROACH FOR GEODETIC VELOCITY FIELD DETERMINATION

Abstract

There has been a need for geodetic network densification since the early days of
traditional surveying. In order to densify geodetic networks in a way that will
produce the most effective reference frame improvements, the crustal velocity field
must be modelled. Artificial Neural Networks (ANNs) are widely used as function
approximators in diverse fields of geoinformatics including velocity field
determination. Deciding the number of hidden neurons required for the
implementation of an arbitrary function is one of the major problems of ANN that
still deserves further exploration. Generally, the number of hidden neurons is
decided on the basis of experience. This paper attempts to quantify the significance
of pruning away hidden neurons in ANN architecture for velocity field
determination. An initial back propagation artificial neural network (BPANN) with
30 hidden neurons is educated by training data and resultant BPANN is applied on
test and validation data. The number of hidden neurons is subsequently decreased,
in pairs from 30 to 2, to achieve the best predicting model. These pruned BPANNs
are retrained and applied on the test and validation data. Some existing methods for
selecting the number of hidden neurons are also used. The results are evaluated in
terms of the root mean square error (RMSE) over a study area for optimizing the
number of hidden neurons in estimating densification point velocity by BPANN.