NEURONAV is an advanced AI-augmented positioning system developed for maritime navigation, designed to enhance the accuracy of Global Navigation Satellite System (GNSS) data in challenging conditions.
Developed by Romanian InSpace Engineering (RISE) in collaboration with the Maritime Hydrographic Directorate (MHD), and funded by the European Space Agency NAVISP Element 2, the system addresses key maritime navigation issues such as multipath errors, interference, and limited ground truth accuracy.
The concept builds on expertise gained from previous ESA projects, including extensive data collection campaigns in the Black Sea, the Aegean Sea, and along the Danube River. These operations involved both piggyback and dedicated vessel deployments, with GNSS data collected and analyzed for performance assessment, interference detection, and multipath characterization.
At a recent ESA-hosted event, Sergiu-Stefan Mihai and Ileana Mihu of RISE presented the final results of the project. At the core of NEURONAV is a machine learning model, Mihai said. Specifically, a convolutional neural network (CNN) is trained to predict and correct position errors caused by multipath.
Unique data processing
The system first gathers GNSS-derived parameters such as satellite azimuth, elevation, carrier-to-noise density ratio (C/N₀), and pseudorange residuals. These inputs are transformed into a Cartesian error space representation, which serves as the CNN’s input matrix. The network processes this data through convolutional layers, max pooling, and fully connected regression layers to predict position corrections in X, Y, and Z coordinates.
NEURONAV hardware features an integrated Septentrio mosaic-X5 multi-constellation, multi-frequency GNSS receiver, with a Jetson Nano single-board computer for AI computation. Data acquisition, processing, and storage are automated, enabling continuous model training and validation without human intervention.
Testing campaigns using MHD’s hydrographic vessels in the Black Sea demonstrated consistent results across multiple trials. Even under heavy radio frequency interference, in line with the observed regional uptick in GNSS jamming and spoofing since the start of the conflict in Ukraine, the trial results demonstrated measurable accuracy gains. For example, in one validation run, mean position error dropped from 1.507 m with the raw receiver output to 1.268 m with NEURONAV corrections. The 95th percentile error improved by over 25 percent compared to the uncorrected GNSS data.
With its ability to augment existing GNSS receivers and process data locally for security and resilience, NEURONAV offers potential benefits for users in both professional and recreational maritime sectors. Mihai said future work will focus on improving performance under severe interference, expanding datasets, and integrating additional navigation sensors for enhanced robustness.
