Deep Learning is Making Your GPS Way More Accurate (And We Helped!)

Hey there! Ever wonder how your phone or car knows *exactly* where you are? It’s pretty amazing, right? We rely on GNSS (Global Navigation Satellite Systems) like GPS for so much these days – from navigating unfamiliar streets to coordinating complex logistics. But there’s a sneaky layer way up in our atmosphere that loves to mess with those signals: the ionosphere.

Why the Ionosphere Matters (And Causes Headaches)

Think of the ionosphere as a dynamic, electrically charged layer. It’s full of free electrons, and when radio waves from satellites pass through it, they get bent and slowed down. This bending and slowing isn’t uniform; it depends on the electron density, which changes constantly based on things like time of day, season, and especially solar activity. This causes a delay in the signal reaching your receiver on the ground, and that delay translates directly into positioning errors. For single-frequency GNSS receivers (like most in consumer devices), this ionospheric delay is the biggest source of error! Dual-frequency receivers can mostly cancel it out, but they’re more expensive.

So, to get really accurate positioning, especially with single-frequency devices, you need to know the state of the ionosphere. That’s where Global Ionospheric Maps (GIMs) come in. These maps provide information about the Total Electron Content (TEC) – basically, how many electrons are in a column above a certain point – across the globe.

The Old Way vs. Our New Idea

For years, the International GNSS Service (IGS) has been providing high-quality GIMs. They get solutions from various analysis centers around the world (like CODE, ESA, JPL, and UPC) and combine them using a pretty standard method: a weighted average. It works, and it’s widely used, especially for post-processing data (meaning you get the map *after* you’ve collected your signal data).

But you know how artificial intelligence, especially deep learning, has been a game-changer in so many fields? We started thinking, “Could we use deep learning to combine these ionospheric maps in a smarter way? Could we capture those complex, non-linear changes in the ionosphere better than a simple average?” That was the core idea behind our research.

Building Our Deep Learning Model

Our objective was to build a deep learning model, specifically using recurrent neural networks (RNNs). RNNs are great for handling data that changes over time, which is exactly what the ionosphere does! We wanted to feed our model all the information available from the different IGS analysis centers’ maps, but also give it context: where the measurement was taken (latitude, longitude, even solar zenith angle), when it was taken (UTC time, converted to sine and cosine to handle that tricky date line!), and what the space weather was like (using solar activity indices like F10.7 and geomagnetic activity indices like Kp).

The challenge with combining maps from different centers is that they have slightly different ways of modeling the ionosphere. Plus, the ionosphere is incredibly dynamic, especially during geomagnetic storms caused by stuff flying off the sun. Modeling those disturbed periods is tough because they don’t happen *that* often, meaning less data for training a model.

For our model to learn, it needed something to aim for – a reference. We used TEC data derived from Jason altimetry missions (Jason-1, -2, -3). These satellites measure the height of the ocean surface using radar, but they also use GNSS and DORIS, allowing us to calculate the TEC between the satellite and the ocean. We collected 16 years of this data (2005-2021).

Filling the Gaps (Literally)

Now, there was a catch with the Jason data: it only gives you TEC up to the satellite’s orbit, which is about 1336 km. GNSS satellites orbit much higher, over 20,000 km! So, we were missing the TEC contribution from the plasmasphere above Jason’s orbit. This was a crucial piece of the puzzle. We tested several ionospheric/plasmaspheric models to see which one best complemented the Jason data. Turns out, the Neustrelitz Electron Density Model (NEDM) was the winner. Adding the correction from NEDM made a huge difference in our model’s performance.

We structured our RNN model with several recurrent and dense layers, carefully tuning it to avoid overfitting while still learning the complex patterns. We trained it on the Jason data, using the interpolated TEC values from the CODE, ESA, JPL, and UPC maps, plus all the positional, temporal, and space weather parameters as inputs. The Jason TEC was our target value.

Putting the Model to the Test (Calm Days)

Once our model (we called it the GUT model, after Gdańsk University of Technology where the work was done) was trained, we needed to see if it actually *worked*. The ultimate test? Single-frequency GNSS positioning. We generated GIMs from our GUT model and compared the positioning accuracy to the results obtained using the standard IGS final maps.

We tested this on data from over 300 IGS ground stations worldwide, covering both quiet ionospheric periods and stormy ones. For the quiet periods (we picked weeks in 2016, 2022, and 2023), our GUT model showed a noticeable improvement in positioning accuracy. On average, for each quiet day studied, over 70% of the stations saw better positioning results compared to using the IGS maps.

The biggest gains were in the Up (altitude) component of the position, with improvements ranging from 5% to a whopping 15% on average for different quiet periods. For the horizontal (2D) positioning, our results were mostly comparable to IGS, sometimes slightly better, sometimes slightly worse, depending on the specific period.

One really cool finding was how well the model performed for stations located far inland or in the polar regions, areas where the Jason reference data (being ocean-based and limited in latitude) wasn’t available. This showed that our deep learning model is really good at *extrapolating* – predicting values outside the range of its training data. For stations in the northern polar region, we saw average improvements of 11% for calm days, and for the southern polar region, an impressive 22%!

Handling the Storms

Ionospheric storms are the ultimate test. During these events, the electron density can change rapidly and dramatically, causing large positioning errors. Could our model handle this? We tested it on data from intense geomagnetic storms in 2015, 2021, and 2023. Importantly, the 2021 and 2023 storms were *outside* our 2005-2021 training data range, again testing the extrapolation ability.

Guess what? Even during these extreme events, our GUT model held its own and often outperformed the standard IGS maps. We saw average improvements in 3D positioning of 8.5% to 12.9% depending on the storm. Again, the improvement was most significant in the Up component, while 2D results were comparable to IGS.

For the strongest storm we tested (March 17, 2015), 78% of the stations saw improved positioning. For the April 2023 storm, it was 86%! Even the November 2021 storm, which was completely outside the training period, saw improvements for 69% of stations. This confirms that the model learned the underlying physics and patterns well enough to predict ionospheric states it hadn’t explicitly seen during training, even during turbulent conditions.

The Plasmasphere Puzzle

We mentioned earlier how important the plasmaspheric correction above Jason’s orbit was. Our research really highlighted this. When we tried training the model *without* any correction for the plasmasphere, the positioning results were significantly worse than IGS, sometimes degrading by over 15% during storms! Using other models like IRI2016 or IRIPLAS helped, but NEDM was clearly the best fit, leading to improvements for over 70% of stations and a 7.9% average reduction in 3D RMSE compared to IGS.

What We Learned

So, what’s the takeaway? Our work showed that using deep learning, specifically RNNs, is a really effective way to combine global ionospheric maps from different sources. By feeding the model not just the map data but also location, time, and space weather info, and using high-quality Jason data as a reference (complemented by the right plasmaspheric model like NEDM), we could create GIMs that lead to better single-frequency GNSS positioning.

The GUT model improved 3D positioning accuracy for over 70% of the stations we tested, both during calm periods and intense geomagnetic storms. The biggest win was in the vertical accuracy, which is often the weakest component in GNSS positioning. The model also proved surprisingly good at extrapolating to areas outside its training data coverage, like inland and polar regions.

This isn’t just about making maps; it’s about making your GPS more reliable and accurate, even with standard equipment, and even when the sun is throwing a tantrum. It shows the huge potential of data-driven approaches like deep learning to enhance critical technologies like satellite navigation. And hey, it’s pretty cool to think that AI is helping us navigate the world a little bit better!

Source: Springer