As it is known, Machine Learning encompasses numerous algorithms that can be trained using a reference (training) data set in order to build a data processing model, applicable to the same data type used for training. In the GNSS case, the data processing model is usually based on the physic models of the different aspects that constitute the GNSS signal (i.e. geometric range, ionospheric and tropospheric delay, orbits, clocks, ...) and positioning filters such as Kalman filter, that estimate a state (position, velocity and time or PVT) based on a set of observations. However, after the estimation has been obtained, little attention is paid to the so-called post-fit residuals (in summary, the values that indicate how well the input data "agrees" with the estimated state). This project attempts to make use of this post-fit residuals along with the positioning errors, as proposed in [1], to develop a processing stage in Rokubun's Jason cloud GNSS service to improve the accuracy of position estimates, specially in challenging scenarios where multipath is pervasive (urban canyons). Being a non-invasive technology (only the products of the Kalman filter are required), this technique can be added as a post-facto processing module. In fact, the input required by this processing module could be later exploited commercially to third parties in the form of licensing or pay-per-use. For the development, training and testing stage of the project, Rokubun will use the vast amount of datasets gathered by Jason (Rokubun GNSS processing service in the cloud). These datasets cover numerous geographic locations (low latitudes, mid-latitudes, ...), diverse scenarios (urban canyons, open sky, semi-residential), rover dynamics (static or moving receivers) and platforms (mass-market GNSS receivers, smartphones, geodetic grade receivers). Additional datasets gathered by Rokubun in previous campaigns may be used as well for this purpose.
