Feeding Neural Network Models with GPS Observations: A Challenging Task

Abstract

Much has been done in terms of functional and stochastic modelling of observations in space geodesy, aiming at the development of adequate adjustment models. One of the techniques, which has been the focus of more attention in the last years, is the Neural Network model. Although not trivial to be used, this kind of model provides an extreme adaptation capability, which can be an issue of fundamental importance for certain applications. In this paper we discuss the use of GPS observations in Neural Networks models, providing a brief description how a neural model works and what are its restrictions, as well as how to treat the GPS observations in order to satisfy them.

A Neural Network is an information processing system formed by a big number of simple processing elements, called artificial neurons. Typically the input values must be normalized, with typical range [0,1], or alternatively [−1,1]. After processed, the signal can be transformed back to its original origin and amplitude. When dealing with GPS observations, namely ranges and range rates, the absolute numerical values are usually pretty large (e.g. order of 20 millions of meters for ranges) coupled with precisions in the order of mm for carrier-phase and meter for pseudoranges. The observations need to be modified to avoid degrading their precision during the normalization, in order to make the application of neural models suitable for GPS data.

In this work methods to make the use of GPS data possible in neural models are discussed and showed with real examples. The analysis is made for both pseudoranges and carrier-phases. It is demonstrated that with the adequate treatment the use of those observables can be made without degradation of precision.