在過去的研究中，傳統電離圖（ionogram）判讀演算法若遇到高背景雜訊、強烈的射頻干擾（radio frequency interference， RFI）或複雜的多層結構，在提取電離層軌跡與參數標註時容易導致其精確性及穩定性降低。因此近年在電離圖判讀演算法開發中逐漸引入深度學習結合電腦視覺技術。 在本研究中首先針對電離圖中高背景雜訊與射頻干擾的嚴重影響，在雷達資料預處理中引入恆虛警率（constant false alarm rate， CFAR）進行射頻干擾偵測，後續以三種射頻干擾抑制演算法進行抑制效果比較。結果證實自適應特徵子空間干擾抑制法為三者演算法當中表現最為優越，在極低的訊號與干擾加雜訊比（signal-to-interference-plus-noise ratio， SINR）的環境下，仍能將 SINR 大幅提升，並完整保留電離層回波功率。 在深度學習於辨識電離圖結構中，本研究改良 U-Net 模型架構經由消融實驗（ablation study）確立最佳訓練組成。於測試集之像素語義分割（semantic segmentation）任務中，模型整體表現優良；在結構辨識任務中，模型在各類別具有非常優越的準確度。經由模型標註之虛高、臨界頻率與真實標籤之標註數值比較後亦具有極高的一致性。顯示即使在電離層軌跡功率微弱的情況下，模型仍能有效辨識整體的軌跡。 與國際參考電離層模型（International Reference Ionosphere， IRI）對比後，臨界頻率兩者的表現相符，惟電離層探測儀反演之 F2 層峰高度範圍廣泛。後續進一步在地磁寧靜期下與福爾摩沙衛星七號（FORMOSAT-7/COSMIC-2， FS-7/C-2）無線電掩星（radio occultation， RO）資料進行對比，其 hmF2 相關係數較IRI模型高。本研究將電離層探測儀反演之地磁場強度與地面磁力計（magnetometer）之水平地磁場強度進行分析。在頻率域中，皆具有全日潮、半日潮以及八小時潮的明顯分量。而對電離層探測儀與地面磁力計之數據進行諧波擬合後，結果顯示兩者之間存在時間差。而此時間延遲不同步的原因可能為電離層中具有其他動力機制影響其結構。;In previous studies， traditional ionogram scaling algorithms often suffered from reduced accuracy and stability when extracting ionospheric traces and annotating parameters under conditions of high background noise， intense radio frequency interference (RFI)， or complex multi-layer structures. Consequently， recent developments in ionogram scaling algorithms have increasingly integrated deep learning with computer vision techniques. To address the severe impact of high background noise and RFI in ionograms， this study first introduces a constant false alarm rate (CFAR) detector in the radar data preprocessing stage for RFI detection， followed by a comparative analysis of three RFI mitigation algorithms. The results demonstrate that the adaptive eigen-subspace filtering for interference suppression (AEFIS) algorithm outperforms the other two methods. It significantly enhances the signal-to-interference-plus-noise ratio (SINR) even in environments with extremely low initial SINR， while completely preserving the ionospheric echo power. Regarding the deep learning-based recognition of ionogram structures， this study improves the U-Net architecture and establishes the optimal training configuration through an ablation study. In the pixel-level semantic segmentation task on the test set， the model delivers robust overall performance; in the event-level structure recognition task， the model achieves highly superior classification accuracy across all categories. Furthermore， the virtual height and critical frequency automatically annotated by the model exhibit exceptional consistency with the ground-truth values， demonstrating that the model can effectively recognize the complete traces even when the echo power is extremely weak. When compared with the International Reference Ionosphere (IRI) model， the critical frequency retrieved by the proposed system is in good agreement; however， the F2-layer peak height (hmF2) retrieved by the ionosonde exhibits a wider dynamic range. Further validation under geomagnetically quiet conditions against Formosat-7/COSMIC-2 (FS-7/C-2) radio occultation (RO) data reveals that the hmF2 retrieved by the ionosonde yields a higher correlation coefficient than that of the IRI model. Lastly， this study analyzes the geomagnetic field intensity retrieved by the ionosonde alongside the horizontal geomagnetic field intensity measured by ground-based magnetometers. In the frequency domain， both datasets exhibit prominent diurnal， semidiurnal， and terdiurnal tidal components. However， harmonic fitting of the ionosonde and magnetometer data reveals a distinct time lag between the two. This phase asynchronization suggests that other physical and dynamical mechanisms within the ionosphere actively modulate its vertical structure.
