GSI-UNet: Extracting Geomagnetic and Solar Indices Features for Real-Time Global Ionospheric Map Enhancement

Real-time Global Ionospheric Map (RT-GIM) products, provided by the International GNSS Service (IGS), are designed to support time-sensitive applications by offering ionospheric information with only a few minutes of latency. However, their accuracy falls far short of the high-precision Final GIM, particularly during periods of intense solar activity. Previous state-of-the-art approaches, such as the Convolutional Neural Network-Enhance (CNN-Enhance) method, have attempted to narrow this gap, but they processed RT-GIM data and solar parameters jointly without distinguishing their heterogeneous features, thereby limiting the ability to capture underlying physical relationships. To overcome these limitations, this paper proposes GSI-UNet (Geomagnetic and Solar Indices-guided U-Net), a deep learning method that introduces dedicated encoders for RT-GIM data and geomagnetic/solar indices, followed by feature fusion to generate an enhanced GIM. In the final GIM evaluation, GSI-UNet further reduced the mean absolute error of RT-GIM by 7.8% compared to the CNN-Enhance method, improved accuracy across all latitude ranges, and achieved 10%–20% higher precision during geomagnetic storms. In single-point positioning tests, the method maintained polar station accuracy within 4.8 m and further improved ocean-based station accuracy by 1–2 m, confirming its effectiveness for real-time ionospheric modeling under severe space weather conditions.