Use of GPU for time series analysis
André Andrade Rodrigues · Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT) · 2019
The Global Navigation Satellite System (GNSS) provides the position of stations with millimetre accuracy. These positions are not constant over time due to motion of the tectonic plates on which they are installed. The motion of tectonic plates is not the same all over the world, leading stress to build up in some areas, thus increasing the occurrence of earthquakes. The knowledge of such movements is of extreme importance, as is the determination of their uncertainty. The tectonic motion is a very slow process that is constant over thousands of years. As a result, the motion can be represented by a linear trend. Programs such as Hector can estimate linear trends in time series with temporal correlated noise. Correlated noise means that GNSS observations made today are similar to those of the days before, and it implies that one actually has less information than when all observations were independent. This is the reason why the real uncertainty of the estimated tectonic motion is between 5 to 11 times larger than when this temporal correlation was not considered. Unfortunately, taking this temporal correlation into account in the analysis slows down the computations considerably. With the ever growing number of GNSS stations and the increasing length of the existing time series, it is necessary to speed up these computations. This behaviour can also be found in other geodetic time series such as sea level observed at tide gauges, and surface temperature time series. For climate change studies it is important that besides estimating sea level rise or an increase in temperature, also the associated uncertainties are realistic. The Hector software package was developed with also the aim to reduce the computation time as much as possible. In this thesis we will investigate if the use of a powerful Graphics Processing Unit (GPU) can further reduce this computation time. So far, no Hector-like software uses GPUs to perform data processing. This dissertation is therefore the first attempt to go in that direction.