簡易檢索 / 詳目顯示

回結果列表
研究生： 林佩萱
Lin, Pei-Xuan
論文名稱： 應用於全球導航衛星系統反射測量之跨星座延遲-都卜勒圖轉換的物理資訊殘差學習方法
Physics-Informed Residual Learning for Cross-Constellation Delay-Doppler Map Conversion in GNSS Reflectometry
指導教授： 莊智清
Juang, Jyh-Ching
學位類別： 碩士
Master
系所名稱： 電機資訊學院 - 電機工程學系
Department of Electrical Engineering
論文出版年： 2026
畢業學年度： 114
語文別： 英文
論文頁數： 133
中文關鍵詞： 全球導航衛星系統反射測量 、延遲－都卜勒圖 、跨星座延遲－都卜勒圖轉換 、物理資訊殘差學習 、HydroGNSS 、真實資料驗證
外文關鍵詞： GNSS reflectometry,, delay–Doppler map, cross-constellation DDM conversion, physics-informed residual learning, HydroGNSS, real-data validation
相關次數： 點閱：90  下載：0 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 多數全球導航衛星系統反射測量（GNSS-R）的延遲－都卜勒圖（DDM）處理鏈建立於全球定位系統（GPS）L1 C/A 碼的二元相位偏移鍵控（BPSK）訊號，而伽利略（Galileo）E1 的複合二元偏移載波（CBOC）具有不同的碼相關結構，因而呈現不同的 DDM 形態。本研究將此差異視為 DDM 形態轉換問題，提出以訓練資料擬合的延遲軸卷積核提供共通修正，並以特徵線性調變（FiLM）殘差 U-Net 依觀測幾何、極化與天線增益條件化，學習樣本相依的局部修正。第一條證據線配對 WAVPY 模擬的 Galileo E1 CBOC 來源 DDM 與 TRITON 實測 GPS 目標 DDM，採用以檔案為單位的分割，並輔以空間與時間洩漏壓力測試。相對於未轉換的來源 DDM，主要高品質測試集上的平均二維峰值位移由 10.84 格降至 5.95 格，延遲軸半高全寬（FWHM）誤差由 10.36 格降至 6.21 格，都卜勒軸 FWHM 誤差由 3.91 格降至 1.78 格，有效支撐區域的二維相關係數由 0.852 升至 0.909；惟多種子重複實驗顯示，都卜勒寬度改善的可重現性低於其他三項指標。匹配模擬控制將約 5.8 格與 2.0 格的延遲與都卜勒寬度差異歸因於 CBOC 至 BPSK 的調變改變，但模擬與觀測之間仍存在殘餘差距。在以 GPS 目標為參照的配對下游評估中，風速均方根誤差（RMSE）由 2.67 m/s 降至 2.34 m/s，海面均方斜率相對平均絕對誤差由 68% 降至 23%。第二條證據線改用 HydroGNSS 實測資料，補足第一條證據線缺少真實 Galileo 輸入的限制。相同框架於任務原生網格上以較早的產品區塊重新訓練，並在時序上較晚的保留產品區塊僅使用 Galileo 來源側資訊進行推論。在近似配對樣本上，平均逐樣本像素相關係數由 0.813 升至 0.893；在由來源側相對尺度估計驅動的凍結 GPS 導向反演介面下，對 ERA5 再分析風速的 RMSE 由 8.65 m/s 降至 7.71 m/s。此 RMSE 與同一介面下使用近似配對 GPS 參考所得的 7.72 m/s 相當。此改善主要來自偏差降低與反演上限飽和減少，相關係數則由 0.228 降至 0.116。Galileo 與 GPS 觀測來自不同發射端與閃耀區，屬近似配對而非同一反射事件；此外，獨立的跨日期下游檢查未達成預定改善條件。因此，本研究的證據所支持的是實測 Galileo DDM 的形態調和，以及與 GPS 導向處理介面的部分相容性，而非操作型風速精度或絕對輻射定標。

    Most delay–Doppler map (DDM) processing in Global Navigation Satellite System reflectometry (GNSS-R) is built on Global Positioning System (GPS) L1 C/A binary phase-shift keying (BPSK) signals, whereas the Galileo E1 composite binary offset carrier (CBOC) signal has a different code-correlation structure and therefore a different DDM morphology. This thesis formulates that mismatch as a morphology-conversion problem. A delay-axis kernel fitted on the training split supplies a shared correction, and a feature-wise linear modulation (FiLM) residual U-Net, conditioned on observation geometry, polarization, and antenna gain, supplies a sample-dependent refinement. The first evidence line pairs WAVPY-simulated Galileo E1 CBOC sources with measured TRITON GPS targets under a file-level split, supported by spatial and temporal leakage stress tests. Relative to the unconverted source DDM, mean 2-D peak displacement on the primary high-quality held-out set decreases from 10.84 to 5.95 bins, delay-axis full-width-at-half-maximum (FWHM) error from 10.36 to 6.21 bins, and Doppler-axis FWHM error from 3.91 to 1.78 bins, while valid-support 2-D correlation increases from 0.852 to 0.909; repeated multi-seed training indicates that the Doppler-width gain is less reproducible than the other three metrics. A matched simulated control attributes delay- and Doppler-width differences of approximately 5.8 and 2.0 bins to the CBOC-to-BPSK change, although a residual simulation-to-observation gap remains. In paired GPS-target-referenced evaluation, wind root-mean-square error (RMSE) decreases from 2.67 to 2.34 m/s and mean-square-slope relative mean absolute error from 68% to 23%. The second evidence line uses measured HydroGNSS data, supplying the real Galileo E1 observations absent from the first line. The same framework is retrained on earlier product blocks in the mission-native grid and applied to a chronologically later held-out block using Galileo-side information only. Across near-matched samples, mean per-sample pixel correlation increases from 0.813 to 0.893. Under a frozen GPS-oriented retrieval interface driven by a source-only relative-scale estimate, RMSE against ERA5 reanalysis wind decreases from 8.65 to 7.71 m/s. Its RMSE is comparable to the 7.72 m/s obtained with the near-matched GPS reference under the same interface. The reduction reflects lower bias and less saturation at the upper retrieval bound, whereas correlation falls from 0.228 to 0.116. Because the Galileo and GPS observations are near matches from different transmitters and glistening zones rather than the same reflection event, and because a separate cross-date downstream check did not meet its predefined improvement criteria, the evidence supports morphology harmonization of measured Galileo DDMs and partial compatibility with GPS-oriented processing interfaces, not operational wind accuracy or absolute radiometric calibration.

    摘要 II Abstract IV 致謝 VI Content VIII List of Tables XIII List of Figures XV List of Acronyms XVII Chapter 1. Introduction 1 1.1 Motivation and Objectives 1 1.2 Literature Review 3 1.3 Contributions 7 1.4 Thesis Organization 8 Chapter 2. Background 9 2.1 GNSS-R Measurement Principles 9 2.1.1 Bistatic Geometry and DDM Formation 9 2.1.2 TRITON CorDDM Data Source 11 2.1.3 Simulated Galileo E1 CBOC DDMs 11 2.1.4 HydroGNSS Real Galileo and GPS DDMs 12 2.1.5 Ocean Surface Scattering and the Z–V Model 14 2.2 GPS and Galileo Signal Modulation 14 2.2.1 GPS L1 C/A BPSK Modulation 15 2.2.2 Galileo E1 CBOC Modulation 15 2.2.3 Modulation-Related DDM Morphological Differences 18 2.3 GNSS-R Retrieval Observables and GMFs 19 2.3.1 DDM Morphology and Observable Extraction 20 2.3.2 Implications for GPS-Oriented Observables and GMFs 20 2.4 Summary 20 Chapter 3. Methodology 22 3.1 Research Design Overview 22 3.2 TRITON Sample Definition and Screening 23 3.3 Leakage-Aware Data Partitioning 24 3.4 Matched Galileo–GPS DDM Pairing 24 3.4.1 Common Delay–Doppler Coordinate Frame 25 3.4.2 Galileo E1 CBOC Source Simulation 27 3.4.3 Paired-Sample Quality Criteria 28 3.4.4 Paired Data Representation 31 3.5 Fitted Baselines 31 3.5.1 Global Affine Baseline 32 3.5.2 Delay-Domain Physics Prior 32 3.6 Physics-Informed Residual Network 33 3.6.1 Morphology-Preserving Normalization 33 3.6.2 Physics-Informed Residual Formulation 34 3.6.3 Geometry-Conditioned Residual Learning 35 3.6.4 Residual U-Net Architecture 37 3.6.5 Training Objective and Optimization 38 3.6.6 Sensitivity Variants 40 3.7 Inference and Evaluation 42 3.7.1 Paired-Scale Reconstruction 43 3.7.2 Morphology and Observable Evaluation 43 3.8 GPS-Oriented Retrieval Compatibility 51 3.9 HydroGNSS Real-Data Validation Methodology 54 3.9.1 Real-Data Evaluation Setup 55 3.9.2 Near-Matched Pairing 56 3.9.3 Source-Only Conversion and Morphology Evaluation 56 3.9.4 Source-Only Scale and Downstream Evaluation 57 3.9.5 Claim Boundary 58 3.10 Summary 58 Chapter 4. Experiments and Results 59 4.1 Experimental Data 59 4.1.1 Source-Target DDM Data 59 4.1.2 Leakage-Aware Data Splits 59 4.1.3 Sample Distributions 60 4.2 Evaluation Setup 62 4.2.1 Models and Baselines 62 4.2.2 Evaluation Metrics 64 4.3 Morphology Conversion Results 64 4.3.1 Primary DDM Morphology Metrics 64 4.3.2 Representative DDM Conversion 67 4.3.3 Delay and Doppler Waveform Analysis 68 4.3.4 Modulation-Specificity Controls 69 4.4 Model Analysis and Ablation 72 4.4.1 Physics-Base and Residual Contributions 73 4.4.2 Spatial Morphology Error Analysis 76 4.4.3 Morphology Diagnostics 77 4.4.4 Paired Primary-Metric Comparison 79 4.4.5 FiLM Conditioning Contribution 79 4.4.6 Training-Data, Loss, and Physics-Base Variants 81 4.4.7 Doppler Over-Broadening and Physics-Base Design 83 4.5 Downstream Retrieval Compatibility 84 4.5.1 Wind Retrieval 84 4.5.2 MSS Retrieval 87 4.6 Robustness Checks 89 4.6.1 Training-Seed Robustness 89 4.6.2 Cross-Quality, Cross-Split, and Holdout Robustness 90 4.6.3 Stratified and Perturbation Analyses 90 4.6.4 Mechanism and Extrapolation Controls 92 4.7 HydroGNSS Real-Data Evaluation 94 4.7.1 Real-Data Setup and Source-Only Workflow 95 4.7.2 Morphology Results and Specificity Diagnostics 96 4.7.3 Source-Only Scale and Downstream Compatibility 99 4.7.4 Interpretation and Limitations 101 4.8 Discussion 102 4.8.1 Interpretation of Main Findings 102 4.8.2 Failure Cases 104 4.8.3 Limitations 106 4.9 Summary 107 Chapter 5. Conclusion and Future Work 108 5.1 Conclusion 108 5.2 Future Work 110 References 112

    [1] V. U. Zavorotny and A. G. Voronovich, “Scattering of GPS signals from the ocean with wind remote sensing application,” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 2, pp. 951–964, Mar. 2000, doi: 10.1109/36.841977.
    [2] S. Gleason et al., “Detection and processing of bistatically reflected GPS signals from low Earth orbit for the purpose of ocean remote sensing,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 6, pp. 1229–1241, Jun. 2005, doi: 10.1109/TGRS.2005.845643.
    [3] M. P. Clarizia, C. S. Ruf, P. Jales, and C. Gommenginger, “Spaceborne GNSS-R minimum variance wind speed estimator,” IEEE Trans. Geosci. Remote Sens., vol. 52, no. 11, pp. 6829–6843, Nov. 2014, doi: 10.1109/TGRS.2014.2303831.
    [4] M. P. Clarizia and C. S. Ruf, “Wind speed retrieval algorithm for the Cyclone Global Navigation Satellite System (CYGNSS) mission,” IEEE Trans. Geosci. Remote Sens., vol. 54, no. 8, pp. 4419–4432, Aug. 2016, doi: 10.1109/TGRS.2016.2541343.
    [5] C. S. Ruf and R. Balasubramaniam, “Development of the CYGNSS geophysical model function for wind speed,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 12, no. 1, pp. 66–77, Jan. 2019, doi: 10.1109/JSTARS.2018.2833075.
    [6] C. S. Ruf et al., “New ocean winds satellite mission to probe hurricanes and tropical convection,” Bull. Amer. Meteorol. Soc., vol. 97, no. 3, pp. 385–395, Mar. 2016, doi: 10.1175/BAMS-D-14-00218.1.
    [7] J.-C. Juang, Y.-F. Tsai, and C.-T. Lin, “TRITON GNSS-R mission: Preliminary results and perspectives,” in Proc. 2024 IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Athens, Greece, Jul. 2024, pp. 6725–6728, doi: 10.1109/IGARSS53475.2024.10642499.
    [8] Taiwan Space Agency (TASA), “TRITON Algorithm Theoretical Basis Document v2.0,” 2025. Accessed: Jul. 11, 2026. [Online]. Available: https://tacc.cwa.gov.tw/documents/TRITON_ATBD_v20.pdf
    [9] D. Pascual, H. Park, A. Camps, A. Alonso-Arroyo, and R. Onrubia, “Simulation and analysis of GNSS-R composite waveforms using GPS and Galileo signals,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 7, no. 5, pp. 1461–1468, May 2014, doi: 10.1109/JSTARS.2014.2311116.
    [10] J.-C. Juang, C.-T. Lin, and Y.-F. Tsai, “Comparison and synergy of BPSK and BOC modulations in GNSS reflectometry,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 13, pp. 1959–1971, 2020, doi: 10.1109/JSTARS.2020.2989824.
    [11] G. W. Hein et al., “MBOC: The new optimized spreading modulation recommended for Galileo L1 OS and GPS L1C,” in Proc. IEEE/ION Position, Location Navig. Symp. (PLANS), San Diego, CA, USA, Apr. 2006, pp. 883–892, doi: 10.1109/PLANS.2006.1650688.
    [12] H.-Y. Wang and J.-C. Juang, “Retrieval of ocean surface wind speed using reflected BPSK/BOC signals,” Remote Sens., vol. 12, no. 17, Art. no. 2698, Sep. 2020, doi: 10.3390/rs12172698.
    [13] European Union, Galileo Open Service Signal-in-Space Interface Control Document (OS SIS ICD), issue 2.2, Nov. 2025, doi: 10.2878/7420173.
    [14] M. L. Hammond et al., “First assessment of geophysical sensitivities from spaceborne Galileo and BeiDou GNSS-reflectometry data collected by the UK TechDemoSat-1 mission,” Remote Sens., vol. 12, no. 18, Art. no. 2927, Sep. 2020, doi: 10.3390/rs12182927.
    [15] Surrey Satellite Technology Ltd., “HydroGNSS PDGS Product Manual,” SSTL Ref. 00729123, Release 5.0, Mar. 12, 2026.
    [16] G. Foti and C. Gommenginger, “HydroGNSS: A-Posteriori L1 Calibration Processors L1B-OP-PX(-L)—Algorithm Theoretical Basis Document,” National Oceanography Centre, Version 2.0, Dec. 1, 2023.
    [17] F. Fabra, E. Cardellach, W. Li, and A. Rius, “WAVPY: A GNSS-R open source software library for data analysis and simulation,” in Proc. 2017 IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Fort Worth, TX, USA, Jul. 2017, pp. 4125–4128, doi: 10.1109/IGARSS.2017.8127908.
    [18] M. Asgarimehr, I. Zhelavskaya, G. Foti, S. Reich, and J. Wickert, “A GNSS-R geophysical model function: Machine learning for wind speed retrievals,” IEEE Geosci. Remote Sens. Lett., vol. 17, no. 8, pp. 1333–1337, Aug. 2020, doi: 10.1109/LGRS.2019.2948566.
    [19] H. Du, W. Li, E. Cardellach, S. Ribó, A. Rius, and Y. Nan, “Deep residual fully connected network for GNSS-R wind speed retrieval and its interpretation,” Remote Sens. Environ., vol. 313, Art. no. 114375, Nov. 2024, doi: 10.1016/j.rse.2024.114375.
    [20] Z. Zhang, J. Xu, G. Jing, D. Yang, and Y. Zhang, “Physics-informed transformer networks for interpretable GNSS-R wind speed retrieval,” Remote Sens., vol. 17, no. 23, Art. no. 3805, Nov. 2025, doi: 10.3390/rs17233805.
    [21] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015, Munich, Germany, Oct. 2015, pp. 234–241, doi: 10.1007/978-3-319-24574-4_28.
    [22] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Las Vegas, NV, USA, Jun. 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90.
    [23] E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, “FiLM: Visual reasoning with a general conditioning layer,” in Proc. 32nd AAAI Conf. Artif. Intell. (AAAI-18), New Orleans, LA, USA, Feb. 2018, pp. 3942–3951, doi: 10.1609/aaai.v32i1.11671.
    [24] H. Hersbach et al., “The ERA5 global reanalysis,” Q. J. R. Meteorol. Soc., vol. 146, no. 730, pp. 1999–2049, Jul. 2020, doi: 10.1002/qj.3803.
    [25] I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. 7th Int. Conf. Learn. Represent. (ICLR), New Orleans, LA, USA, May 2019. Accessed: Jul. 13, 2026. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

    下載圖示
    校外：立即公開
    QR CODE