Synthetic‐Data‐Driven Supervised Learning for Pixel‐Wise Recognition of Ionogram Layer Traces

Abstract

The significance of automatic ionogram data interpretation is growing with the increasing number of ionosondes operating at minute‐level cadence. In recent years, deep‐learning models have substantially advanced ionogram pattern recognition. However, they typically require extensive labeled ionospheric data sets for training. A key challenge is developing models for pixel‐wise identification of ionogram layer traces with sufficient generalization to maintain consistent performance across independent test sets from different years and phases of the solar cycle. Traditionally, entire trace labeling is not routine, as human resources are limited to determining key ionospheric parameters. We propose a method to generate ground truth without manual labeling and train a U‐Net with attention for robust pixel‐wise segmentation of ionogram layers at high latitudes. The model recognizes E, F1, and F2 traces and handles complex F‐spread patterns and overlapping higher‐order reflections. This capability was demonstrated using an independent test data set of ionograms. Following layer segmentation, we evaluated the mean absolute errors of key parameters: minimum frequency ( f _min ), maximum frequency (fmax), and minimum virtual height ( h _min ). The mean absolute errors for the E layer are f _min  = 0.79 pixels (0.047 MHz), fmax = 0.85 pixels (0.051 MHz), and h _min  = 0.54 pixels (3.10 km); for the F1 layer, f _min  = 0.66 pixels (0.040 MHz), f _max  = 0.91 pixels (0.055 MHz), and h _min  = 0.58 pixels (3.33 km); and for the F2 layer, f _min  = 1.19 pixels (0.071 MHz), f _max  = 0.75 pixels (0.045 MHz), and h _min  = 1.24 pixels (7.12 km).