Deep Convolutional Attention Network for Autoscaling of foF2 and h’F2 from Ionograms

The ionosphere, a critical component of Earth's space environment, provides essential support for engineering systems such as high-frequency communication and Radar systems. Ionospheric parameters can be derived from ionograms. However, the extraction of ionospheric parameters remains challenging due to the complex and variable nature of ionograms. To address this issue, we propose a cascaded end-to-end network architecture that operates on a coarse-to-fine principle. This architecture first establishes global context representations using large field convolution, then introduces a spatial attention mechanism to dynamically localize and adaptively focus on echo trace regions. We evaluate the performance of the proposed method using ionogram data from Daofu station (30.98° N, 101.13° E), Puer station (22.7° N, 101.05° E), and Zhangye station (39.4° N, 100.13 ° E). The correlation coefficients between automatically scaled and manual foF2 and h'F2 at Puer reach 0.9985 and 0.9393 respectively, with Root Mean Square Errors (RMSE) of 0.1448 MHz and 10.1161 km. Furthermore, the network model maintains stable measurement accuracy across ionogram datasets at different latitudes with minimal fine-tuning, providing a reliable technical approach for automated extraction and real-time processing of ionospheric parameters.