Technical Article

One-hour-ahead forecasting of ionogram morphology and spread-F signatures using a spatial group-wise enhanced ConvTransformer

¹ State Key Laboratory of Media Convergence and Communication, Communication University of China, Beijing, PR China
² Key Laboratory of Media Audio & Video (Communication University of China), Ministry of Education, Beijing, PR China
³ State Key Laboratory of Solar Activity and Space Weather, National Space Science Center, Chinese Academy of Sciences, Beijing, PR China
⁴ Hainan National Field Science Observation and Research Observatory for Space Weather, Danzhou, Hainan Province, PR China
⁵ Planetary Environmental and Astrobiological Research Laboratory (PEARL), School of Atmospheric Sciences, Sun Yat-sen University, Zhuhai, PR China
⁶ University of Chinese Academy of Sciences, Beijing, PR China

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 12 August 2025
Accepted: 16 April 2026

Abstract

The Digisonde Portable Sounder (DPS) ionosonde at Hainan Station (19.5°N, 109.1°E; magnetic latitude: 11°N) has been monitoring ionospheric conditions since 2002, routinely recording ionospheric plasma profiles, sporadic E layers, and Spread-F structures through ionograms. A Spatial Group-wise Enhanced ConvTransformer (SGE-ConvTransformer) is proposed in this study for spatiotemporal ionospheric prediction at the Hainan station, with emphasis on Spread-F, enabling a one-hour lead time with a 15-minute sampling resolution. The SGE module optimizes semantic feature extraction from the global spatial context, dynamically recalibrating attention to prioritize information-rich regions, such as F-layer traces, over background noise. To further improve visual clarity, a super-resolution Enhanced Deep Super-Resolution (EDSR) module is integrated to sharpen the predicted ionograms. Leveraging DPS ionosonde data from 2002 to 2015, we constructed a large-scale ionogram sequence dataset comprising 36,240 Spread-F instances and 396,931 non-Spread-F instances, which were further categorized into five distinct classes. On the 2016 test set, our model achieved an average Spread-F classification accuracy of 90.05% and a correlation coefficient of 0.8115 for the predicted F-trace. Demonstrating superior robustness under disturbance conditions, the model maintained high performance during six representative geomagnetically disturbed intervals (2023–2024), achieving a classification accuracy of up to 95.69%. Furthermore, the model's generalizability was examined by applying pre-trained weights to data from low-latitude (Brazil, Peru), mid-latitude (Irkutsk), and high-latitude (Zhigansk) stations. Quantitative Spread-F Classification Accuracy (SFCA) metrics at low latitudes and qualitative visual assessments across all regions demonstrate the morphological transferability of our approach across diverse geospatial environments.