電離層位於地球表面約 50 至 1000 公里高度，因太陽輻射電離大氣的中性
粒子而形成，對高頻（HF）無線電波的傳播特性具有重要影響，並在遠距離通
訊、GPS 導航、軍事與太空天氣預報等領域有重要應用。電離圖（Ionograms）
由電離層觀測系統網資料生成，記錄了無線電波反射訊號的虛擬高度與頻率關
係。其中，普通波（O-mode）軌跡對提取 F2 層的電離層參數—臨界頻率
（foF2）與虛擬高度（ℎ’F2）至關重要。中央大學的原始自動分析方法使用模糊
邏輯理論處理，例如 Tsai 等人提出的模糊辨識與連通演算法。然而，這方法在
背景噪音高或多跳軌跡複雜的情況下仍會有誤差。近年，隨著深度學習的發
展，卷積神經網絡已被應用於電離層資料分析。例如：Huang 等人使用 Spatial
Attentional U-net（SA U-net）成功從花蓮站電離圖中識別出 F2 層的普通波
和雜散 Es 層訊號；本研究採用 SA U-net 深度學習網路模型，結合空間注意力
機制，自動檢測垂直與斜射電離圖中的 O-mode 軌跡，並與中央大學 Fuzzy 算法
進行性能對比。為了實現這一目標，使用中央大學蔡龍治教授的花蓮站 VIPIR
的電離圖數據，建立本地環境進行 SA U-net 模型訓練，並在 NVIDIA Tesla
P100 GPU 與 Docker 容器環境下試著優化模型參數。實驗結果顯示，SA U-net
在垂直電離圖中能夠較好識別 F2 層普通波軌跡，但在處理雜亂多跳或 X-mode
訊號時效果不如預期；使用垂直電離圖訓練之模型預測斜射電離圖中，僅在少
部分情況下成功辨識斜射電離圖軌跡，多數斜射特徵未被完整捕捉。可見，深
度模型對原始電離圖具有一定的辨識潛力，但對複雜場景仍需更大規模數據與
本地環境改進。

The ionosphere, located approximately 50 to 1000 km above Earth's surface, is
formed by the ionization of neutral atmospheric particles due to solar radiation. It
significantly influences the propagation of high-frequency radio waves and plays a
critical role in applications such as long-distance communication, GPS navigation,
military operations, and space weather forecasting.
Ionograms, generated from Hualien VIPIR, record the relationship between virtual
height and frequency of reflected radio signals. The ordinary wave (O-mode) trace is
essential for ionospheric parameters of the F2 layer, namely the critical frequency
(foF2) and virtual height (ℎ’F2).The original fuzzy logic–based approach adopted by
National Central University (NCU), including the fuzzy identification and connectivity
algorithms proposed by Tsai et al., may still produce recognition errors when
analyzing ionograms with high background noise or complex multi-hop trace
conditions. In recent years, advancements in deep learning have led to the
application of convolutional neural networks in ionospheric data analysis. For
instance, Huang et al. employed a Spatial Attentional U-net (SA U-net) to successfully
identify F2-layer O-mode and sporadic Es-layer signals in ionograms from the Hualien
station.
This study adopts the SA U-net deep learning model with a spatial attention
mechanism to automatically detect O-mode traces in both vertical and oblique
ionograms, comparing its performance with NCU fuzzy algorithms. To achieve this
objective, using ionogram data from the Hualien VIPIR station, provided by Professor
Lung-Chih Tsai of National Central University (NCU), a local environment was
established to train the SA U-net model, with optimization performed on an NVIDIA
Tesla P100 GPU within a Docker container.
Experimental results reveal that SA U-net effectively identifies F2-layer O-mode
traces in vertical ionograms but exhibits suboptimal performance in handling
complex multi-hop or X-mode signals. When the model, trained on vertical
ionograms, was applied to predict oblique ionograms, it successfully identified
oblique traces in only a few cases. Most oblique features not fully captured. The
findings reveal that deep learning models possess notable potential for interpreting
raw ionograms, yet their effectiveness in complex conditions is constrained by
limited data and insufficient accommodation of local environmental factors.

[1]
Baumjohann, W., & Treumann, R. A. (2012). Basic space plasma physics.
[2]
Chang, Y. C., Lin, C. H., Dmitriev, A. V., Hsieh, M. C., Hsu, H. W., Lin, Y. C., ... & Tsogtbaatar, E. (2022). State-of-the-art capability of convolutional neural networks to distinguish the signal in the ionosphere. Sensors, 22(7), 2758.
[3]
Guo, C., Szemenyei, M., Yi, Y., Wang, W., Chen, B., & Fan, C. (2021, January). Sa-unet: Spatial attention u-net for retinal vessel segmentation. In 2020 25th international conference on pattern recognition (ICPR) (pp. 1236-1242). IEEE.
[4]
Huang, G. H., Dmitriev, A. V., Lin, C. H., Chang, Y. C., Hsieh, M. C., Tsogtbaatar, E., ... & Li, Y. H. (2022). The development of Spatial Attention U-Net for the recovery of ionospheric measurements and extraction of ionospheric parameters. Radio Science, 57(8), 1-12.
[5]
Ian Julian (n.d.). The Ionosphere [Online image]. https://www.radioheritage.com/Shortwave-With-A-Difference/radiopic/ionosphere.jpg
[6]
Jenn, D. (n.d.). Ionospheric Wave Propagation (Version 1.6.5) [PPT image]. Naval Postgraduate School. https://faculty.nps.edu/jenn/EC3630/Ionosphere(v1.6.5).pdf
[7]
Maucher, J. (2020, October 30). Convolution filter demo [GIF image]. Hochschule der Medien Stuttgart. http://maucher.home.hdm-stuttgart.de/Pics/ConvolutionFilterDemo3.gif
[8]
Tsai, L. C., & Berkey, F. T. (2000). Ionogram analysis using fuzzy segmentation and connectedness techniques. Radio Science, 35(5), 1173-1186.
[9]
Zolesi, B., & Cander, L. R. (2013). The general structure of the ionosphere. In Ionospheric Prediction and Forecasting (pp. 38). Berlin, Heidelberg: Springer Berlin Heidelberg.
[10]
GeeksforGeeks. (2019, July 21). CNN | Introduction to pooling layer. https://www.geeksforgeeks.org/cnn-introduction-to-pooling-layer/
[11]
NVIDIA Corporation. (2016, October). NVIDIA Tesla P100 datasheet.
[12]
鄭惇仁（1999）。應用模糊集合理論於自動化電離圖判讀之研究。﹝碩士論文。國立中央大學﹞臺灣博碩士論文知識加值系統。
59
[13]
鄭子琳（2024）。應用Spatial Attention U-Net神經網路於VIPIR垂直電離圖的自動判圖處理。﹝碩士論文。國立中央大學﹞臺灣博碩士論文知識加值系統。