Enhancing Smartphone GNSS Positioning through Deep Learning-Based Ionospheric Prediction and Correction

The positioning accuracy of smartphone-based Global Navigation Satellite Systems (GNSS) is significantly degraded by ionospheric delay. Current correction methods primarily rely on broadcast ionospheric models, which offer limited precision. High-precision real-time ionospheric grid products, while more accurate, require a stable internet connection and incur additional costs, posing a significant constraint for mobile applications in network-free environments. To address this challenge, we propose ST-VIT-NET, a novel deep learning model based on the Vision Transformer (ViT) architecture for real-time ionospheric prediction and correction. ST-VIT-NET learns the deviation between the Klobuchar broadcast model and the high-precision final ionospheric grid product IGS-GIM, thereby enabling high-accuracy ionospheric correction for broadcast ephemeris-based models.

Experimental results demonstrate that, on a global scale, the ST-VIT-NET model achieved an average Root Mean Square (RMS) error of 4.37 TECU in predicting the Vertical Total Electron Content (VTEC) over a 131-day period from day of year 161 to 292 in 2025. This represents reductions of 64.44% and 9.34% compared to the Klobuchar model (12.29 TECU) and the IGS real-time GIM model (4.82 TECU), respectively, indicating strong temporal and spatial generalizability. In static positioning tests, Standard Point Positioning (SPP) using the ST-VIT-NET model with a Huawei P40 smartphone yielded horizontal and vertical RMS positioning errors of 1.35 m and 2.18 m. These values are 52.42% and 62.77% lower than those obtained using the Klobuchar model (2.85 m horizontal, 5.86 m vertical), and 25.42% and 13.71% lower than those using the IGS real-time GIM model (1.82 m horizontal, 2.53 m vertical). In kinematic vehicle tests, SPP using the ST-VIT-NET model with a Huawei Mate40 smartphone resulted in horizontal and vertical RMS errors of 2.37 m and 3.81 m. This corresponds to reductions of 59.62% and 62.51% compared to the Klobuchar model (5.87 m horizontal, 10.16 m vertical), and 13.41% and 33.01% compared to the IGS real-time GIM model (2.74 m horizontal, 5.69 m vertical).

Collectively, the findings confirm two key contributions of the proposed model. First, ST-VIT-NET demonstrates strong temporal and spatial generalizability, as evidenced by its sustained high-precision VTEC prediction capability over an extended 131-day period across diverse regions. Second, it provides a viable and self-contained solution for achieving real-time high-precision GNSS positioning on smartphones in network-free scenarios, as it delivers accurate ionospheric corrections using only onboard GNSS observations without any external data dependency.