Application Of Machine Learning Methods In Quantitative Precipitation Estimation And Nowcasting Using Radar Data

In this thesis, based on the data observed by the China new generation S band radar inHefei and the data collected by the raingauges around Hefei, three kinds of artificial neuralnetworks (ANN),such as radial basis function network (RBF), generalized regression neuralnetwork (GRNN) and wavelet BP network (BPNN), are introduced to apply in quantitativeprecipitation estimation, and their results are compared with the result of Z-R relationshipmethod. Based on TREC, with the combination of ANN and support vector machine (SVM)one hour forecast on radar reflectivity has been done and the forecast results with crosscorrelation method (TREC)have been compared.The data used in the thesis is the CAPPI reflectivity data at constant altitude underrectangular coordinates.It is obtained from the raw volume scanning radar data after azimuthcomplementary processing,dual-linear interpolation in horizontal direction and linearinterpolation in vertical direction. Five indexes are applied,such as correlation coefficient(CC),root mean square error (RMSE), average relative error (Wabs),average deviation (Bias)and average root mean square difference (σ),to comprehensively compare the quantitativeprecipitation estimation results of artificial neural network and Z-R relationship. Six indexesare applied, such as hit rate (HR),false alarm rate (FAR),the rate of missing report(NAP),critical success index (CSI),correlation coefficient (CC) and root mean square error(RMSE),to inspect the forecast effects of artificial neural network, support vector machineand TREC.The main conclusions and results are as follows:(1) Based on the capabilities of artificial neural network in solving non-linear problems, theestablished models of radar reflectivity factor and precipitation intensity with artificialneural network had good performance and promotion ability. The models could identifycertain evolution of the complex reflectivity-precipitation intensity sequences in changingtime and space, and estimate the precipitation in similar rainfall process with highprecision.(2) Compared with Z-R relationship, estimation of per hour rainfall with artificial neuralnetwork and the measured values were in good agreement, where even though the rainfallintensity was larger (>45mm/h), the error of estimation and the measured value was stillsmall relatively.While Z-R relationship apparently did overestimation for the strongprecipitation intensity which was more than45mm/h,and did insufficient estimation foractual intensity between5to45mm/h.(3) The performance of artificial neural network was influenced by its structure andparameter. When the structure selected was not ideal or parameter set is not reasonable,the network performance was poor. (4) RBF, GRNN, BPNN and SVM could nowcast the radar reflectivity echoes in time andspace.(5) The threshold setting had some impact on the evaluation of forecast results when HR,FAR, NAP and CSI are used. When the forecast time was shorter (about30minutes),TREC could do better than SVM, using high and low threshold,evaluated withthe hit rate (HR),the false alarm rate (FAR), the no alarm probability (NAP) and thecritical success index (CSI),together correlation coefficient (CC) and root mean squareerror (RMSE),vice versa.(6) Compared with TREC, SVM could do better than TREC on forecasting strong convectiveweather changes and development for one hour on the whole.