极端降雨发生会引发严重的洪水、城市内涝等灾害事件，已成为全球最具破坏性的天气现象之一。降雨的发生伴随着大气水汽的时空变化，与对流层中的水汽变化密切相关。因此，可以根据大气水汽变化来预测降雨的变化。大气层层内水汽含量大小可用大气可降水量(Precipitable Water Vapor，PWV)定量描述，其表示地表至大气顶界单位截面积空气柱中全部水汽凝结为液态水可形成的降雨量。全球导航卫星系统(Global Navig- ation Satellite System，GNSS)信号穿过对流层时会受到层内水汽的影响，产生大气延迟效应，该延迟量可以通过计算转换为PWV。利用GNSS技术获取PWV具有全天候、高时空分辨率、高精度等优势，逐渐引起国内外学者的关注，并推动GNSS PWV在短临降雨预警方面的快速应用。目前已有学者利用GNSS PWV水汽产品构建适用于热带、亚热带及温带地区的短临降雨预测模型，但是该类模型普遍存在预报因子单一、构建原理简单、降雨预测错报率高的缺点。因此，本文针对现有基于GNSS的短临降雨预测模型正确率低和错报率高等核心问题展开研究。本文主要研究内容总结如下：

(1) 现有基于GNSS PWV和天顶对流层总延迟(Zenith Total Delay，ZTD)的降雨预测模型仅利用PWV或ZTD的变化量和变化率来预测降雨事件，导致模型降雨预测正确率低和错报率高。本文针对上述问题，提出一种联合PWV和ZTD的短临降雨预警方法，该方法选取PWV值、PWV变化量和变化率、ZTD变化量和变化率五种预报因子构建降雨预测模型，并且考虑了降雨变化的季节性影响。此外，传统降雨预测模型预报因子最优阈值常根据经验值选取，这种阈值选取方法缺乏普适性、准确性和可靠性。本文创新点之一是利用百分位法取代传统经验阈值选取方法，最优百分位数即对应各种预报因子(PWV值、PWV变化量、PWV变化率、ZTD变化量、ZTD变化率)最优阈值，此时降雨预测模型正确率达到最高，错报率最低。实验结果表明联合PWV和ZTD所建降雨预测模型精度高于现有传统降雨预测模型。

(2) 降雨发生受多种气象参数的影响，仅依靠PWV和ZTD不能精准描述降雨发生的全过程。因此，在构建更高精度降雨预测模型时，本文通过分析气象参数与降雨间关系，提出一种联合GNSS反演PWV和气象参数的短期降雨预警理论，并联合机器学习方法确定降雨预测的最优模型。利用反向传播神经网络(Back-Propagation Neural Network，BP-NN)算法构建相应短期降雨预测模型，相比传统GNSS PWV/ZTD构建短期降雨预测模型，该类模型考虑PWV、温度、气压和相对湿度等多种气象参数参与的降雨发生过程。针对BP-NN模型，根据Kolmogrov和Kung and Hwang理论分别确定隐含层节点数和学习率。实验结果表明基于BP-NN算法构建的降雨预测正确率较传统降雨预测模型提高近10%，错报率相仿。

(3) 准确、定量的降雨预测有助于气象及政府部门制定有效措施，以保护人类生命和财产免受泥石流、滑坡等灾害气象事件的威胁。现有基于GNSS PWV/ZTD及其联合气象参数构建的降雨预测模型主要研究重点是预测未来短期(30 min-6 h)内降雨是否发生，鲜有研究针对未来短期降雨量大小进行预测。此外，上述降雨预测模型构建过程中未考虑降雨自相关信息，难免会造成数据信息遗漏及模型精度较低等缺点。因此，本文在现有传统模型和BP-NN降雨预测模型基础上，提出联合支持向量机(Support Vector Machine, SVM)和GNSS PWV构建未来逐小时降雨量的预测模型。针对SVM算法，依据网格搜索法和交叉验证法理论分别确定正则化参数和核函数最优参数。实验结果表明：基于SVM构建短期降雨量预测模型精度优于现有模型，表现在实际和预测降雨量误差更小和相关性更高两个方面。

The extreme rainfall has become one of the most destructive weather phenomenon around the world, which can induce severe floods, urban flooding, and other disaster events. There is a closely relationship between the rainfall and the water vapor in the troposphere. The spatial-temporal variation of the water vapor in the atmosphere maybe induce the occurrence of the rainfall. Therefore, the rainfall could be forecasted based on the variation of the water vapor. The content of the water vapor could be quantified by the Precipitable Water Vapor (PWV), which refers to the depth of precipitation formed by the condensation of water vapor into rain in the air column of the unit cross section from the surface ground to the top troposphere. The Global Navigation Satellite System (GNSS) signal transverses the troposphere and will be affected by the water vapor producing atmospheric delay. The delay could be further transformed into PWV. The PWV requirement have the advantages of all-weather, high spatial-temporal resolution and high precision based on the GNSS technology, which gradually attracted the attention from many scholars. And accelerate the rapid application of GNSS PWV in the short term rainfall forecast. The rainfall forecast models have been established for different regions based on the GNSS PWV products recently. However, these models always exists the disadvantages of single predictor, simple established principle and the high false rainfall forecast rate. Therefore, this study focused on the nuclear questions of high false and low right rainfall forecast rate in the previous studies, and proposed a new approach and methodology to solve these questions. The mainly research in this study are as follows:

(1) The rainfall forecast model usually be established by the variation and the first deviation of GNSS-derived PWV or Zenith Total Delay (ZTD), which caused the rainfall forecast model with high false forecast rate and low right forecast rate. In this paper, a short-term rainfall forecast model is established to forecast the rainfall, which concluded PWV and ZTD two kind of predictors. The rainfall forecast model considered PWV variations and its first deviations, seasonal ZTD variations and its first deviations and PWV monthly values. And the model also considered the characteristics of rainfall in different seasons. In addition, the empirical threshold method usually be selected to determine the forecast factor threshold of the traditional rainfall forecast model. The threshold method exists the disadvantages of university, accuracy and reliability. The selection method of the best threshold also is one of the innovations in this study. During establishing the model, the optimal threshold of various predictors (PWV values, PWV variations, PWV first deviations, ZTD variations, ZTD first deviations) be determined by the percentile theory. The principle of the optimal parameter threshold is the highest right rainfall forecast rate and the lowest false rainfall forecast. The experiment result shows the accuracy of the rainfall forecast model of the combing the PWV and ZTD is higher than the traditional rainfall forecast model.

(2) The single predictor of PWV or ZTD can not describe the accurate process of rainfall, since rainfall events are correlated with myriad atmospheric parameters. Therefore, besides the above established model, this paper established short term rainfall forecast model through combing the GNSS-derived PWV/ZTD and the meteorological parameters based on the machine learning algorithm. Comparing the traditional rainfall forecast model, the Back-Propagation Neural Network (BP-NN) was used to establish the corresponding short term rainfall forecast model, which considered PWV, surface temperature, pressure, relative humidity and other meteorological parameters. During establishing the short term rainfall forecast model based on the BP-NN model, the hidden layer nodes and learning rate be determined by the kolmogrov and the Kung and Hwang theories for BP-NN model, respectively. The experiment result shows that the true rainfall forecast rate of the short term rainfall model based on the machine learning algorithm is superior than the existed rainfall forecast model, and the false forecast rate is comparable.

(3) For preventing casualties and property losses caused by landslides and floods, accurate and quantitative rainfall forecast is necessary to the meteorology and government departments to formulate effective measures. The main focus of the existed rainfall forecast model based on the GNSS PWV/ZTD and its combined meteorological parameters is on predicting whether the rainfall will occur in the short term (30 min-6 h). There are few studies to forecast the size of the rainfall in short term of the future. In addition, the above rainfall forecast model does not consider the auto-correlation character of the rainfall, which will make the data wasted and induce the accuracy of the model low and other disadvantages. Therefore, this paper proposes combing the Support vector machine (SVM) and GNSS PWV to establish hourly rainfall forecast model. The kernel parameter and regularized parameter be confirmed by the grid search and cross-validation method for SVM algorithm, respectively. The experiment result shows that the accuracy of the rainfall forecast model based SVM is the highest among the models, and could be further applied to the actual rainfall forecast events. The experiment results also showed the accuracy of the rainfall forecast model based on the SVM algorithm is better than the existed rainfall forecasted model, which reflects on the high correlation and low root mean squared error. 

[1] McElroy, Michael B. The Atmospheric Environment. New Jersey: Princeton University Press, 2002, 34~36.

[2] Li L I, Kuang C L, Zhu J J, et al. Rainstorm nowcasting based on GPS real-time precise point positioning technology[J]. Chinese Journal of Geophysics, 2012, 55(4):1129~1136.

[3] Manandhar S, Lee Y H, Meng Y S, et al. A simplified model for the retrieval of precipitable water vapor from GPS signal[J]. IEEE Transactions on Geoscience and Remote Sensing, 2017, 55, 6245~6253.

[4] Wang L, Geller M. A. Morphology of gravity-wave energy as observed from 4 years (1998–2001) of high vertical resolution US radiosonde data[J]. Journal of Geophysical Research: Atmospheres, 2003, 108, 1–10.

[5] Gong J, Geller M A. Vertical fluctuation energy in United States high vertical resolution radiosonde data as an indicator of convective gravity wave sources[J]. Journal of Geophysical Research: Atmospheres, 2010, 115, 1–16.

[6] Zhao Q, Yao Y, Yao W, et al. Near-global GPS-derived PWV and its analysis in the El Niño event of 2014–2016[J]. Journal of Atmospheric and Solar-Terrestrial Physics, 2018a, 179, 69–80.

[7] Rahimi Z, Shafri H Z M, Norman M. A GNSS-based weather forecasting approach using Nonlinear Auto Regressive Approach with Exogenous Input (NARX)[J]. Journal of Atmospheric and Solar-Terrestrial Physics, 2018: 178, 74–84.

[8] Bevis M, Businger S, Chiswell S. GPS meteorology: Mapping zenith wet delays onto precipitable water[J]. Journal of applied meteorology, 1994, 33, 379–386.

[9] Gutman S I, Benjamin S G. The role of ground-based GPS meteorological observations in numerical weather prediction[J]. GPS solutions, 2001, 4, 16–24.

[10] Wang J, Zhang L, Dai A, et al. A near-global, 2-hourly data set of atmospheric precipitable water from ground-based GPS measurements[J]. Journal of Geophysical Research Atmospheres, 2007, 112, 1–17.

[11] Xiong L H, Connor K M. An empirical method to improve the prediction limits of the GLUE methodology in rainfall–runoff modeling[J]. Journal of Hydrology, 2008, 349, 115–124.

[12] Lu K, Wang L. A novel nonlinear combination model based on support vector machine for rainfall prediction[C]. In 2011 Fourth International Joint Conference on Computational Sciences and Optimization, 2011, 1343-1346.

[13] Chang E K, Peña M, Toth Z. International research collaboration in high-impact weather prediction[J]. Bulletin of the American Meteorological Society, 2013, 94(11): ES149-ES151.

[14] 聂进. 模糊决策树算法在降雨预测中的研究与应用: [毕业论文]. 广州: 广东工业大学, 2014.

[15] Yu P S, Yang T C, Chen S Y, et al. Comparison of random forests and support vector machine for real-time radar-derived rainfall forecasting[J]. Journal of hydrology, 2017, 552, 92-104.

[16] Olson D A, Junker N W, Korty B. Evaluation of 33 years of quantitative precipitation forecasting at the NMC[J]. Weather and Forecasting, 1995, 10(3): 498-511.

[17] Prakash S, Mahesh C, Gairola R M, et al. A feasibility of six-hourly rainfall forecast over central India using model output and remote sensing data[J]. International Journal of Hydrology Science and Technology, 2012, 2(2): 138-152.

[18] Robertson D E, Shrestha D L, Wang Q J. Post processing rainfall forecasts from numerical weather prediction models for short term streamflow forecasting[J]. Hydrology & Earth System Sciences Discussions, 2013, 17(9): 3587-3603.

[19] Askne J, Nordius H. Estimation of tropospheric delay for microwaves from surface weather data[J]. Radio Science, 1987, 22, 379–386.

[20] Bevis M, Businger S, Herring T A, et al. GPS meteorology: Remote sensing of atmospheric water vapor using the global positioning system[J]. Journal of Geophysical Research, 1992, 97(D14): 15787.

[21] Oikonomou C, Tymvios F, Pikridas C, et al. Tropospheric delay performance for GNSS integrated water vapor estimation by using GPT2w model, ECMWF’s IFS operational model and in situ meteorological data[J]. Advances in Geosciences, 2018, 45, 363–375.

[22] Katsougiannopoulos S, Pikridas C, Zinas N, et al. Analysis of Precipitable Water Estimates using permanent GPS station data during the Athens heavy rainfall on February 22th 2013[M]//IAG 150 Years. Springer, Cham, 2015: 407-414.

[23] Benevides P, Catalao J, Miranda P M. On the inclusion of GPS precipitable water vapour in the nowcasting of rainfall[J]. Natural Hazards and Earth System Sciences, 2015, 15, 2605–2616.

[24] Manandhar S, Lee Y H, Meng Y S, et al. GPS-Derived PWV for Rainfall Nowcasting in Tropical Region[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018a, 56, 4835–4844.

[25] Samui P, Mandla V R, Krishna A, et al. Prediction of rainfall using support vector machine and relevance vector machine[J]. Earth Science India, 2011, 4, 188-200.

[26] Guan Z, Tian Z, Xu Y, et al. Rainfall predict and comparing research based on Arcgis and BP neural network. 2016 3rd International Conference on Materials Engineering. Manufacturing Technology and Control. https://doi.org/10.2991/icme mtc -16.2016.291 (2016).

[27] Benevides P, Catalao J, Nico G. Neural Network Approach to Forecast Hourly Intense Rainfall Using GNSS Precipitable Water Vapor and Meteorological Sensors[J]. Remote Sensing, 2019, 11(8): 966.

[28] Vapnik V, Lerner A. Pattern recognition using generalized portrait[J]. Automation and Remote Control, 1963, 24, 774-780.

[29] Lucay F A, Cisternas L A, Gálvez E D. An LS-SVM classifier based methodology for avoiding unwanted responses in processes under uncertainties[J]. Computers & Chemical Engineering, 2020, 138, 106860.

[30] Tripathi S, Srinivas V V, Nanjundiah R S. Downscaling of precipitation for climate change scenarios: a support vector machine approach[J]. Journal of hydrology, 2006, 330(3-4): 621-640.

[31] Manandhar S, Dev S, Lee Y H, et al. A data driven approach to detecting precipitation from meteorological sensor data[J]. IGARSS 2018 2018 IEEE International Geoscience and Remote Sensing Symposium, 2018b, 3872-3875.

[32] Hamidi O, Poorolajal J, Sadeghifar M, et al. A comparative study of support vector machines and artificial neural networks for predicting precipitation in Iran[J]. Theoretical and applied climatology, 2015, 119(3-4): 723-731.

[33] Gocic M, Shamshirband S, Razak Z, et al. Long-term precipitation analysis and estimation of precipitation concentration index using three support vector machine methods[J]. Advances in Meteorology, 2016, 2016.

[34] 毛节泰. GPS的气象应用[J]. 气象科技, 1993, 000(004):45-49.

[35] Yao Y, Shan L, Zhao Q. Establishing a method of short-term rainfall forecasting based on GNSS-derived PWV and its application[J]. Scientific reports, 2017, 7(1): 1-11.

[36] Zhao Q, Yao Y, Yao, W. GPS-based PWV for precipitation forecasting and its application to a typhoon event[J]. Journal of Atmospheric and Solar-Terrestrial Physics, 2018b, 167, 124-133.

[37] Zhao Q, Yao Y, Yao W, et al. Real-time precise point positioning-based zenith tropospheric delay for precipitation forecasting[J]. Scientific reports, 2018c, 8(1), 1-12.

[38] 李永华, 刘德, 金龙. 基于BP神经网络的汛期降水预测模型研究[J]. 气象科学, 2002, 22(4): 461-467.

[39] 刘莉, 叶文. 基于 BP 神经网络时间序列模型的降水量预测[J]. 水资源与水工程学报, 2010, 21(5): 156-159.

[40] 刘洋, 赵庆志, 姚顽强. 基于多隐层神经网络的GNSS PWV和气象数据的降雨预测研究[J]. 测绘通报(S1), 2019, 36-40.

[41] 杨淑群, 芮景析, 冯汉中. 支持向量机方法在降水分类预测中的应用[J]. 西南农业大学学报(自然科学版), 2006, (02): 81-86.

[42] 罗芳琼, 吴建生, 金龙. 基于最小二乘支持向量机集成的降水预报模型[J]. 热带气象学报, 2011, 27(004): 577-584.

[43] 农吉夫. 基于主成分分析和支持向量机的区域降水预测应用研究[J]. 广西民族大学学报(自然科学版), 2009, 15(2): 89-93.

[44] Saastamoinen J. Atmospheric correction for the troposphere and stratosphere in radio ranging satellites[J]. The use of artificial satellites for geodesy, 1972, 15, 247–251.

[45] Desai S, Kuang D, Bertiger W. GIPSY/OASIS (GIPSY) overview and under the hood. Near Earth Tracking Syst. Appl. Groups, Jet Propuls. Lab., California Inst. Technol., Pasade- na,CA,USA,Tech.Rep.Onlineat.ftp://ehzftp.wr.usgs.gov/svarc/GIPSY_pdfs/GIPSY_Overview.Pdf (2014).

[46] Hecht-Nielsen R. Neurocomputer applications[M]//Neural computers. Springer, Berlin, Heidelberg, 1989, 445-453.

[47] Hu A, Zhang K. Using Bidirectional Long Short-Term Memory Method for the Height of F2 Peak Forecasting from Ionosonde Measurements in the Australian Region[J]. Remote Sensing, 2018, 10, 1658.

[48] Rani B K, Srinivasa K. Govardhanb A. Rainfall Prediction with TLBO Optimized ANN[J]. Journal of Scientific and Industrial Research, 2014, 73, 643–647.

[49] Hashim F R, Daud N N, Ahmad K A, et al. Prediction of rainfall based on weather parameter using artificial neural network[J]. Journal of Fundamental and Applied Sciences, 2017, 9, 493–502.

[50] Li F. Liu C. Application study of BP neural network on stock market prediction[J]. 2009 Ninth International Conference on Hybrid Intelligent Systems IEEE, 2009, 3, 174–178.

[51] Dharia A. Adeli H. Neural network model for rapid forecasting of freeway link travel time[J]. Engineering Applications of Artificial Intelligence, 2003, 16, 607–613.

[52] Xu T, Zheng W, Sun P. Transient power quality recognition based on BP neural network theory[J]. Energy Procedia, 2012, 16, 1386–1392.

[53] Gao M, Wu Z. Personalized Context-Aware Collaborative Filtering Based on Neural Network and Slope One[J]. International Conference on Cooperative Design, Visualization, and Engineering, 2009, 109–116.

[54] Yu F, Xu X. A short-term load forecasting model of natural gas based on optimized genetic algorithm and improved BP neural network[J]. Applied Energy, 2014, 134, 102–113.

[55] Kung S Y, Hwang J N. An Algebraic Projection Analysis for Optimal Hidden Units Size and Learning Rates in Back-Propagation Learning[J]. Proceedings IEEE International Conference on Neural Networks, 1988, 1, 363–370.

[56] Manandhar S, Dev S, Lee Y H, et al. Systematic study of weather variables for rainfall detection[C]. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, 2018c, 3027-3030. IEEE.

[57] Rahman M M, Davis D. Cluster based under-sampling for unbalanced cardiovascular data[J]. Proceedings of the World Congress on Engineering, 2013, 3, 3–5.

[58] Yen S J, Lee Y S. Cluster-based under-sampling approaches for imbalanced data distributions[J]. Expert Systems with Applications, 2009, 36, 5718–5727.

[59] Laza R, Pavón R, Reboiro-Jato M, et al. Evaluating the effect of unbalanced data in biomedical document classification[J]. Journal of integrative bioinformatics, 2011, 8, 105–117.

[60] Yuan Executive, R.O.C. The Republic of China Yearbook 2014. ISBN 9789860423020. Retrieved 11 June 2016. 2014.

[61] Yeh T K, Hong J, Wang C, et al. Applying the water vapor radiometer to verify the precipitable water vapor measured by GPS[J]. Terrestrial Atmospheric and Oceanic Sciences, 2014, 25(2): 189-201.

[62] Wauters M, Vanhoucke M. Support vector machine regression for project control forecasting[J]. Automation in Construction, 2014, 47, 92-106.

[63] Were K, Bui D T, Dick Ø B, et al. A comparative assessment of support vector regression, artificial neural networks, and random forests for predicting and mapping soil organic carbonstocks across an Afromontane landscape[J]. Ecological Indicators, 2015, 52, 394-403.

[64] Deka P C, Raghavendra N S, Support vector machine applications in the field of hydrology: a review[J]. Applied soft computing, 2014, 19, 372-386.

[65] Liu Z. A method of SVM with normalization in intrusion detection[J]. Procedia Environmental Sciences, 2011, 11, 256-262.

[66] Min J H, Lee Y C. Bankruptcy prediction using support vector machine with optimal choice of kernel function parameters[J]. Expert Systems with Applications, 2005, 28(4): 603-614.

[67] Kisi O, Cimen M. Precipitation forecasting by using wavelet-support vector machine conjunction model[J]. Engineering Applications of Artificial Intelligence, 2012, 25(4): 783-792.

[68] Daut I, Yusoff M I, Ibrahim S, et al. Relationship between the Solar Radiation and Surface Temperature in Perlis[J]. Advanced Materials Research, 2012, 512, 143-147.