Research On Spatiotemporal Downscaling Algorithm Of Conventional Meteorological Elements Based On Deep Learnin

In meteorological services,gridded meteorological elements play an essential role,among which temperature is one of the most concerned meteorological elements.Due to the limitation of the number and distribution of ground stations,the observation data are sparsely distributed,which leads to the lack of meteorological data refinement,and improving the spatio-temporal resolution of data has great practical significance for meteorological operations.The need for spatio-temporal level refinement of grid data can be further subdivided into spatial downscaling(also called super-resolution)as well as temporal downscaling reconstruction,i.e.,improving the resolution in both spatial and temporal dimensions.Currently,the interpolation methods commonly used in operations can produce high-resolution fields,but the overall tendency is smooth and there are large errors with the observed fields.This study will focus on completing the algorithm implementation and result validation of the temporal downscaling reconstruction of gridded temperature data through an in-depth study of deep learning related methods,and verifying the generalizability by other conventional meteorological elements.The specific research contents include:(1)Designing the Spatial Downscaling GAN(SDGAN)model for spatial downscaling reconstruction of low spatial resolution data based on the optimization idea of adversarial neural networks.The model includes a generator model for recovering high-resolution data and a discriminator model for adversarial training.The temporal autocorrelation of the data is analyzed,the data are organized by time series,and the convolutional time series model Conv GRU is selected as the basic structure and improved by adding a module composed of deformable convolutions and a bidirectional structure to fuse the bidirectional time series features while adaptively adjusting the perceptual field of the convolutional kernel to extract the features of irregular spatial distribution.The overall structure uses kriging geospatial interpolation for low-resolution data and fuses the output layer with the model feature map to learn geospatially relevant information.At the optimization level,a cosine annealing restart strategy is used to dynamically adjust the decay and reset of the learning rate to avoid oscillation of the parameters around the local optimum and to converge to the global optimum by jumping out of the local optimum.(2)Using the SDGAN and optical flow time interpolation methods based on the content of research topic 1 to design an end-to-end spatiotemporal downscaling model for coarse time intervals and spatially low-resolution data,improving the temporal and spatial resolution in a single step.The optical flow time interpolation structure is improved based on the U-Net model by adding a residual structure and a self-attentiveness mechanism,predicting the bi-directional optical flow field at adjacent input moments for synthesizing intermediate moment data frames.The resulting optical flow fields are fused with the spatial downscaling model for channel-level features,and experiments show that the spatiotemporal fused network model has better accuracy than the combination of mutually independent temporal and spatial methods.(3)Designed experiments for research topics(1)and(2),respectively,and the results show that the models designed in this study outperform the operational method in terms of metrics such as Root Mean Square Error(RMSE)and practical performance.The ablation experiment verifies the effectiveness of the improved method.Additionally,the analysis experiments on the restart parameter of the cosine annealing strategy show that the learning rate reset in the training process of this task will cause instability in the training process and cannot find the global optimum.Therefore,the restart parameter is not used as the optimum in this task.