ARTIFICIAL NEURAL NETWORKS PRUNING APPROACH FOR GEODETIC VELOCITY FIELD DETERMINATION

• Autor(a) principal: YILMAZ, MUSTAFA
• Data de Publicação: 2013
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Boletim de Ciências Geodésicas
• Texto Completo: https://revistas.ufpr.br/bcg/article/view/34873
• topic: Geociências; Geodésia; Geodésia

There has been a need for geodetic network densification since the early days oftraditional surveying. In order to densify geodetic networks in a way that willproduce the most effective reference frame improvements, the crustal velocity fieldmust be modelled. Artificial Neural Networks (ANNs) are widely used as functionapproximators in diverse fields of geoinformatics including velocity fielddetermination. Deciding the number of hidden neurons required for theimplementation of an arbitrary function is one of the major problems of ANN thatstill deserves further exploration. Generally, the number of hidden neurons isdecided on the basis of experience. This paper attempts to quantify the significanceof pruning away hidden neurons in ANN architecture for velocity fielddetermination. An initial back propagation artificial neural network (BPANN) with30 hidden neurons is educated by training data and resultant BPANN is applied ontest 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 BPANNsare retrained and applied on the test and validation data. Some existing methods forselecting the number of hidden neurons are also used. The results are evaluated interms of the root mean square error (RMSE) over a study area for optimizing thenumber of hidden neurons in estimating densification point velocity by BPANN.