Nowcasting Prediction Based On Radar Echo Images

In the area of nowcasting precipitation,the requirements of radar echo image prediction resolution and forecasting time precision are much higher than the traditional weather forecasting.However,the main drawbacks of traditional nowcasting precipitation methods include the over-complex construction of physical atmosphere models and the cumulative multi-step prediction errors.This thesis explored the nowcasting precipitation methods based on radar echo image sequences,and proposed a precipitation-related model based on convolutional auto-encoder and long short-term memory recurrent neural network for radar echo image sequences feature prediction.The main work and contributions of this thesis are summarized as the following.(1)Image Space Feature Extraction and Image Reconstruction based on Convolutional Auto-Encoder for Radar Echo Data Sequences.By training the constructed Convolutional Auto-Encoder,image space features can be learned for radar echo sequence data.Then the image space features can be extracted from the input history radar echo image sequence based on the encoder module.Furthermore,with the predicted future time-sequence feature from LSTM network,the radar echo image sequence can be reconstructed based on the decoder module of the trained Convolutional Auto-Encoder.Experimental results verify the effectiveness of the proposed method.(2)Time-Sequence Feature Detection and Prediction via Long Short-Term Memory based Recurrent Neural Network.By implementing a Long Short-Term Memory based recurrent neural network with encoding-reconstruction structure,time-sequence features can be learned from radar echo data.Then,the time-sequence features can be extracted from the input history echo image sequence based on the encoder module from LSTM network.Furthermore,the future time-sequence features can be predicted based on the prediction module from LSTM network.Experimental results verify the effectiveness of the proposed method.