Neural Network Methods For Radar Data For Precipitation Type Classification

In this thesis, a Back-Propagation Model of Artificial Neural Network is used forportioning the reflectivity into straitiform and convective rain classifications using the dataobserved by the China new generation S band A series radar (CINRAD/SA) in Hefei during2001 to 2003. And the trained neural network is applied in a precipitation progress.The radar mentioned above is located at 117°15′28′′E,31°52′1′′N. The altitude of theantenna is 165.0m.The highest shelter elevation is 1.5°. The best observing elevation is 0.5°.The aerial gain of the antenna is 45.69dBZ. The peak power is 690kw and the wavelength is10cm. The data used in the thesis is the CAPPI reflectivity data at different levels from 5 km to10 km under rectangular coordinates. It is obtained from the raw volume scanning radar dataafter azimuth complementary processing, dual-linear interpolation in horizontal direction andlinear interpolation in vertical direction. The horizontal grid interval is 1.5km, and vertical gridinterval is 1.0km. Among the data mentioned above, the portion used for convectiveprecipitation classification is the data from 2001 to 2003 of 973 Program, and the rest portionused for straitiform precipitation classification is the data observed by the same radar in 2003.100 matrixes with the size 30km×30km of each type are selected from the reflectivity datamentioned above as the database. 20 ones out of the 100 matrixes for each type are trainingdatabase, and the rest are the database for validation. A Back-Propagation Model with 2 neural cells in in-put layer, 3 neural cells in out-putlayer and 1 hidden layer with 6 neural cells is used in this thesis. The main conclusions and results are as follows:(1) The Artificial Neural Network can classify the radar reflectivity intoconvective/straitiform precipitation types successfully. Given the appropriate frameand parameters, the successfully-identify rate is higher than 93.75% with trainingdatabase, validation database and the data from the validating progress.(2) Since the former statistic model is not needed for the ANN to do the identification,and the character dimension for the ANN is 2. Therefore, the ANN is better thanthe traditional statistic methods for practical application.(3) Validate the conclusion which is mentioned in other former researches that theperformance of ANN is influenced greatly by its frame. If the frame is not proper,the network will perform badly. In this thesis, when given the 2 dimentions of thein-put layer, 3 dimentions of the out-put layer and single hidden layerBack-Propagation model, 6 neural cells for the hidden layer, the network will reachthe highest successfully-identify rate.(4) The results of the experiment show that the number of the neural cells of the hiddenlayer influences the performance of ANN greatly. If the number of the neuralcells is too small, the network will perform badly,. But it doesn't mean that thelager the number of the neural cells the better the network performs. For a certainnetwork, a most proper number exists. In this experiment, when the number is 6,the network reaches the highest successfully-identify rate.(5) Validate the conclusion which is mentioned in other former researches that theperformance of ANN is influenced greatly by the learning rate A. Given a properlearning rate A, the network will convergent quickly and reach a highsuccessfully-identify rate. But if the learning rate is not appropriate, the networkwill perform badly and even can not convergent. In this thesis, the results of theexperiment indicate that when given the learning rate A of 0.7, the network willreach the highest successfully-identify rate.(6) Validate that the quality and number of the samples of the training database issignificant and can influence the generalizing capability of the network directly.The more modalities of the portioned model the training database includes, thebetter generalizing capability the network will have.(7) Validate that when training the network, the input order of the samples of thetraining database will influences the performance of the network remarkably.