Research On Weather Nowcasting Technology Based On Radar Echo Image Sequences

Nowcasting has always been a hot spot in the field of meteorological research,which plays an important role in social operation and human production and life.This paper presents three research methods about nowcasting:a traditional image-processing algorithm based on TV-L~1 variable spectral flow and two image-sequence prediction methods based on deep learning.The main research contents include:(1)Based on the TV-L~1 variable spectral flow model,we design a method of extrapolating radar echo image-sequence.We fuse the conservation hypothesis of Hessian matrix in TV-L~1 data items and combine global constraints with local constraints.We also optimize the optical flow calculation using image pyramid technique and joint bilateral filtering method.Experimental results show that the proposed scheme is effective in predicting radar echo image-sequence compared with the original TV-L~1 variable splitter flow algorithm.(2)Since the traditional image-processing algorithm cannot effectively utilize the time sequence characteristics of radar echo image,we propose a prediction model of radar echo image based on time and space dual-channel encoder-decoder in this paper.The model can effectively predict the radar echo image by using the temporal and spatial characteristics of radar echo image-sequence.Through experimental comparison and analysis,it verifies that the network model has higher prediction accuracy than the two-channel self-coding network.Under the threshold of 50d BZ,the CSI index is 1.83%higher than that of dual-channel encoder-decoder network,the FAR index is 1.92%lower than that of dual-channel encoder-decoder network,and the POD index is 1.25%higher,and all indexes are optimal on the whole.(3)Because the two-channel encoder-decoder network model proposed in this paper predicted the generated image sequence is fuzzy,and the longer the prediction time,the more blurred the image.Based on this defect,we propose a prediction algorithm about radar echo image based on WGAN architecture(WGAN-MSCGRU)of deep learning.The network combines generative adversarial network,codec structure,multi-scale convolution kernel and multi-weight loss function.Our scheme effectively improves the fuzzy prediction of image and improve the prediction ability of echo image details.The experimental results show that the proposed network has higher prediction accuracy.Under the threshold of 50d BZ,the CSI index is 1.83%higher than that of dual-channel encoder-decoder network,the FAR index is 1.92%lower than that of dual-channel encoder-decoder network,and the POD index is 1.25%higher.Experiments verify that the three algorithms proposed in this paper can successfully predict the future rainfall distribution with certain accuracy.The three algorithms have their own advantages,the traditional image processing method does not need a large number of datasets for training,and is more convenient to use.The prediction network of radar echo image-sequence based on double channel encoder-decoder solves the problem that the traditional algorithm cannot effectively utilize the time sequence features of echo image,and the network model is easy to train.The multi-scale codec network based on WGAN can not only predict the rainfall in a longer time but also have higher prediction accuracy,and the generated image has more details,but the network training time is too long and the difficulty is relatively greater.