Indentification of Weather Radar Abnormal Data Based on Deep Learning

The world's largest weather radar observation network which consists of 236 weather radars is built up in China. The quality control of weather radar data becomes an indispensable part in operation as data grow. In real-time operation of CMA Meterological Observation Center, the abnormal data caused by radar hardware fault or calibration problem usually leads to a bad image, and the problem directly affects the quality of weather radar data, quantitative estimation of precipitation and analysis of the weather system. At present, the abnormal data are processed by artificial corrigendum in real-time operation, so that it does not affect the application of follow-up data. In recent years, artificial intelligence technology has developed rapidly, and deep learning algorithms are used to build convolutional neural network models to extract image features from abnormal data such as radar hardware failures and electromagnetic interference. According to characteristics of the abnormal data, two kinds of abnormal dataset are established. The first kind is labeled data for those samples with whole abnormal picture or precipitation. The second kind is the electromagnetic interference in fixed direction, which is easily mixed with precipitation, and it is used to label the samples for pixels. Based on the convolution pre-training network model (ResNet), the R-ResNet model is built with the first kind of abnormal data. The model evaluation shows that the R-ResNet model achieves more than 99% accuracy in both training and test datasets, and the prediction results of the model in test datasets are all consistent with the label data. For the second kind of abnormal data, based on the image semantic segmentation network model LinkNet, a group of encoders and decoders are extended, and a hop-link structure is added, and ResNet50 is used as the encoder structure. The R-LinkNet model is constructed and the accuracy is over 98% on the training and test datasets, with intersection over union of 83.4% and 83.2%, respectively. Two models can be used to monitor the abnormal data of the national radar in real-time operation, so that abnormal data can be corrected automatically and the workload of the manual on duty can be reduced greatly.