Prediction of Ionograms With/Without Spread‐F at Hainan by a Combined Spatio‐Temporal Neural Network

An intelligent high‐definition and short‐term prediction of ionograms with/without Spread‐F for the observation at Hainan (19.5°N, 109.1°E, magnetic 11°N) is presented in this paper, which comprises a spatio‐temporal ConvGRU network and a super‐resolution EDSR network. Our prediction is based on spatio‐temporal features in the ionogram graph only. There are 469,227 ionograms classified into 5 categories, that is, frequency/range/mix/strong range/no Spread F, over a solar cycle (14 years) labeled manually by the research group, and we process these ionograms into two data sets for training the two networks mentioned above. A series of comprehensive experiments have been designed and conducted to determine the optimal super‐parameters. Our method inputs 8 consecutive authentic ionograms (lasting 2 hr) and generates the next 2 figures (next 30 min). Remarkably, all predicted figures achieve a high accuracy rate of over 94% in predicting the occurrence of Spread‐F. Plain Language Summary Since the clear trace indicates the ionospheric electron density background and the Spread‐F indicates disturbances/irregularities, the prediction of both ionograms with/without Spread‐F is important to the research and application. There are currently no well‐established neural networks specifically designed to predict the extensive information encompassed within ionograms, particularly the intricate characteristics of Spread‐F. In this paper, we approach the short‐term prediction (next 30 min) of ionograms with/without Spread‐F. To achieve this, we employ the ConvGRU network to generate ionograms blurred but still captured the primary spatio‐temporal features of various types of Spread‐F (FSF/RSF/MSF/SSF) in ionogram sequences, as well as features without Spread‐F. Subsequently, we refine these rough images by EDSR network to obtain clear and detailed predictions. Our treatment to blurred prediction figures and our focus on the Spread‐F key area are innovative. Through our work, the auto‐prediction for ionosonde observation benefits ionosphere research and monitoring. Key Points Rather than numerical prediction methods, our prediction is based on spatio‐temporal features only Predicting high‐definition ionograms by combining ConvGRU and EDSR on the basis of establishing two ionogram data sets Achieving a high accuracy of 94.28% for Spread‐F prediction