Identification of the Reflected Signal in the COSMIC-2 Occultation Signal Using the Improved GoogLeNet Deep Learning Model
-
摘要: GPS(Global Positioning System)掩星探测技术作为一种先进的大气探测手段,已广泛用于数值天气预报、气候和空间天气研究。掩星探测存在的问题之一是容易受到地球表面反射信号的干扰,识别和分离掩星探测信号中的反射信号有助于将掩星数据同化到数值天气预报系统中去,具有重要意义。本文提出一种基于改进的GoogLeNet深度学习模型(Im-GNet),并应用于COSMIC-2掩星探测数据来识别反射信号。本文选择了2020年1月1~9日的COSMIC-2掩星数据(conPhs文件),进行质量控制后,利用无线电全息方法得到掩星信号的无线电全息功率谱密度图像,然后训练得到Im-GNet深度学习模型,Im-GNet模型测试的准确率达到了96.4%,显著高于支持向量机(SVM)方法的结果。本文还分析了反射信号对掩星数据的影响,掩星事件的地理分布以及掩星反演数据(atmPrf文件)与NCEP再分析资料的12 h预报值(avnPrf文件)的折射率比较表明:有反射信号的掩星事件数据质量更好,所包含的大气信息更丰富。
-
关键词:
- COSMIC-2掩星 /
- 深度学习 /
- 无线电全息技术 /
- GoogLeNet模型
Abstract: As an advanced atmospheric detection method, GPS (Global Positioning System) occultation detection technology has been widely used in numerical weather forecasting, climate, and space weather research. One of the problems in occultation detection is that it is easily interfered with by the reflected signals on the surface of the earth. Identifying and separating the reflected signals in the occultation detection signal helps assimilate the occultation data into the numerical weather prediction system, which has considerable importance. This study proposes a deep learning model based on improved GoogLenet (Im-GNet) model and applies it to COSMIC-2 occultation detection data to identify reflected signals. This study selects the COSMIC-2 occultation data (conPhs file) from 1 January to 9 January 2020. After quality control, the radio holography method is used to obtain the spatial spectrum image of the occultation signal, and the Im-GNet deep learning model is trained. The accuracy rate of the Im-GNet model test reached 96.4%, which is significantly higher than the result of the support vector machine method. This study also analyzes the impact of reflected signals on occultation data. The geographic distribution of occultation events and the refractivity comparison between the occultation inversion data (atmPrf file) and the NCEP (National Centers for Environmental Prediction) 12-h forecast files (avnPrf file) shows that the quality of the occultation event data with reflection signals is better, and the atmospheric information contained is richer.-
Key words:
- COSMIC-2 occultation /
- Deep learning /
- Radio holographic technology /
- GoogLeNet model
-
图 2 一次COSMIC-2掩星事件[(1.75°N,54.02°E),掩星编号为C2E1.2020.001.03.50.G19]的(a)L1波段(1.57542 GHz)信号、(b)L2波段(1.22760 GHz)信号的时间—频率域空间谱,(c)L1波段信号、(d)L2波段信号的影响高度—弯角域空间谱
Figure 2. Spatial spectrums of time–frequency domain for the (a) L1 wave band (1.57542 GHz) and (b) L2 wave band (1.22760 GHz) signals, spatial spectrums of height–bending angle domain for the (c) L1 wave band and (d) L2 wave band signals in a COSMIC-2 occultation event [(1.75°N, 54.02°E), occultation ID is C2E1.2020.001.03.50.G19]
图 7 COSMIC-2掩星事件与NCEP再分析资料的折射率廓线的误差均值和标准差:(a)有清晰反射信号;(b)有模糊反射信号;(c)没有反射信号。黑色曲线代表误差均值,黑色水平线段代表标准差,蓝色曲线代表掩星事件的数量
Figure 7. Mean error and standard deviation of the refractive index profile between COSMIC-2 occultation events and NCEP reanalysis data: (a) Clear reflection signals; (b) unclear reflection signals; (c) no-reflection signals. The black curve represents the mean error, the black horizontal line represents the standard deviation, and the blue curve represents the number of occultation events
表 1 试验计算机的软、硬件配置
Table 1. Configuration for hardware and software
项目 配置 硬件 内存(RAM) :16 G;
处理器:Inter(R) Core(TM)i7-9850H CPU操作系统 Windows 10专业版 计算机语言 Python 3.7.0 机器学习框架/工具 GoogLeNet:Tensorflow 2.1.0
SVM:Scikit-learn 0.18.1表 2 三种模型的测试结果
Table 2. Test results for the three models
模型 模块/核函数 准确率 运算时间/s Im-GNet Inception(3a、3b、4a) 96.4% 12423 Li-SVM 线性核函数 80.9% 2053 Ga-SVM 高斯核函数 73.8% 2804 -
[1] Aparicio J M, Cardellach E, Rodríguez H. 2018. Information content in reflected signals during GPS radio occultation observations [J]. Atmospheric Measurement Techniques, 11(4): 1883−1900. doi: 10.5194/amt-11-1883-2018 [2] Beyerle G, Hocke K. 2001. Observation and simulation of direct and reflected GPS signals in Radio Occultation Experiments [J]. Geophys. Res. Lett., 28(9): 1895−1898. doi: 10.1029/2000GL012530 [3] Beyerle G, Hocke K, Wickert J, et al. 2002. GPS radio occultations with CHAMP: A radio holographic analysis of GPS signal propagation in the troposphere and surface reflections [J]. J. Geophys. Res.: Atmos., 107(D24): 4802. doi: 10.1029/2001JD001402 [4] Boniface K, Aparicio J M, Cardellach E. 2011. Meteorological information in GPS-RO reflected signals [J]. Atmospheric Measurement Techniques, 4(7): 1397−1407. doi: 10.5194/amt-4-1397-2011 [5] Cardellach E, Oliveras S. 2016. Assessment of a potential reflection flag product [R]. ROM SAF Report 23. [6] Cardellach E, Oliveras S, Rius A. 2008. Applications of the reflected signals found in GNSS radio occultation events [C]//Proceedings of GRAS SAF Workshop on Applications of GPS Radio Occultation Measurements. Shinfield Park, Reading: ECMWF, 133–143. [7] Cardellach E, Ao C O, De La Torre Juárez M, et al. 2004. Carrier phase delay altimetry with GPS-reflection/occultation interferometry from low Earth orbiters [J]. Geophys. Res. Lett., 31(10): L10402. doi: 10.1029/2004GL019775 [8] Gorbunov M E. 2001. Radioholographic methods for processing radio occultation data in multipath regions [J]. Danish Meteorological Institute, 1–38. [9] Gorbunov M E. 2002. Radioholographic analysis of radio occultation data in multipath zones [J]. Radio Sci., 37(1): 14-1–14-9. doi:10.1029/2000RS002577 [10] Gorbunov M E, Lauritsen K B. 2004. Analysis of wave fields by Fourier integral operators and their application for radio occultations [J]. Radio Sci., 39(4): RS4010. doi: 10.1029/2003RS002971 [11] Gorbunov M E, Gurvich A S, Kornblueh L. 2000. Comparative analysis of radioholographic methods of processing radio occultation data [J]. Radio Sci., 35(4): 1025−1034. doi: 10.1029/1999RS002247 [12] Gorbunov M E, Lauritsen K B, Leroy S S. 2016. Application of Wigner distribution function for analysis of radio occultations [J]. Radio Sci., 45(6): RS6011. doi: 10.1029/2010RS004388 [13] Gorbunov M E, Cardellach E, Lauritsen K B. 2018. Reflected ray retrieval from radio occultation data using radio holographic filtering of wave fields in ray space [J]. Atmospheric Measurement Techniques, 11(2): 1181−1191. doi: 10.5194/amt-11-1181-2018 [14] Gorbunov M E, Lauritsen K B, Rhodin A, et al. 2006. Radio holographic filtering, error estimation, and quality control of radio occultation data [J]. J. Geophys. Res. :Atmos., 111(D10): D10105. doi: 10.1029/2005JD006427 [15] Hocke K, Pavelyev A G, Yakovlev O I, et al. 1999. Radio occultation data analysis by the radioholographic method [J]. Journal of Atmospheric and Solar-Terrestrial Physics, 61(15): 1169−1177. doi: 10.1016/S1364-6826(99)00080-2 [16] Hu A D, Wu S Q, Wang X M, et al. 2018. Improvement of reflection detection success rate of GNSS RO measurements using artificial neural network [J]. IEEE Trans. Geosci. Remote Sens., 56(2): 760−769. doi: 10.1109/TGRS.2017.2754512 [17] Jensen A S, Lohmann M S, Benzon H H, et al. 2003. Full spectrum inversion of radio occultation signals [J]. Radio Sci., 38(3): 1040. doi: 10.1029/2002RS002763 [18] Jensen A S, Lohmann M S, Nielsen A S, et al. 2004. Geometrical optics phase matching of radio occultation signals [J]. Radio Sci., 39(3): RS3009. doi: 10.1029/2003RS002899 [19] Kuo Y H, Wee T K, Sokolovskiy S, et al. 2004. Inversion and error estimation of GPS radio occultation data [J]. J. Meteor. Soc. Japan, 82(1B): 507−531. doi: 10.2151/jmsj.2004.507 [20] Kursinski E R, Hajj G A, Bertiger W I, et al. 1996. Initial results of radio occultation observations of Earth’s atmosphere using the global positioning system [J]. Science, 271(5252): 1107−1110. doi: 10.1126/science.271.5252.1107 [21] Schreiner W S, Weiss J P, Anthes R A, et al. 2020. COSMIC-2 radio occultation constellation: First results [J]. Geophys. Res. Lett., 47(4): e2019GL086841. doi: 10.1029/2019GL086841 [22] Sokolovskiy S V. 2001. Modeling and inverting radio occultation signals in the moist troposphere [J]. Radio Sci., 36(3): 441−458. doi: 10.1029/1999RS002273 [23] Szegedy C, Liu W, Jia Y, et al. 2014. Going deeper with convolutions [J]. IEEE Computer Society. https://doi.org/10.48550/arXiv.1409.4842 [24] 杨真真, 匡楠, 范露, 等. 2018. 基于卷积神经网络的图像分类算法综述 [J]. 信号处理, 34(12): 1474−1489. doi: 10.16798/j.issn.1003-0530.2018.12.009Yang Z Z, Kuang N, Fan L, et al. 2018. Review of image classification algorithms based on convolutional neural networks [J]. Journal of Signal Processing (in Chinese), 34(12): 1474−1489. doi: 10.16798/j.issn.1003-0530.2018.12.009 -