Accurate nowcasting provides key information for disaster weather warnings. Nowcasting is mostly based on radar echo extrapolation, where the echo evolution results from complex interactions among cloud systems and various thermal–dynamic features of the weather background. However, existing research has not integrated high spatiotemporal resolution weather background information. In this study, multiple data fusion units (Radarcells) integrating the radar mosaic and rapid-refresh data are constructed as input for the radar echo extrapolation network architecture designed according to Attention-Resnet Unet with Radarcells (ARUR). In addition, a self-defined loss function combining weighted mean square error and structural similarity index is further proposed to improve the extrapolation effect. The rapid-refresh data includes relative humidity, zonal wind, meridional wind, and vertical velocity within the range from 115.48 to 117.48°E, 38.81 to 40.81°N during June–August from 2018 to 2021. Four ARUR-based models with different Radarcells as input and an ARU-based model (Attention-Resnet Unet, without fusing physical data) are trained for 120 min of extrapolation, respectively. The models are evaluated with the indicators, including critical success index (CSI), probability of detection (POD), and false alarm rate (FAR), by the test dataset. The results show that the performance of echo prediction and timeliness by ARUR-based models is better than the ARU-based model, especially for strong echo prediction. Under the reflectivity thresholds of 25 and 35 dBZ, the average values of CSI, FAR, and POD calculated by ARUR-based models are improved by 8.42%, 7.76%, 8.52% and 10.36%, 7.36 %, 9.10% than those by ARU-based, respectively. The study suggests that integrating weather background information can significantly enhance the effect of extrapolation by means of improving the issues of echo blurriness, formation, and dissipation compared with the previous deep learning-based models.

Using 82892 volume scanning data at 0.5° elevation angle from the S-band dual polarization radar deployed in Guangzhou (CINRAD/SAD) and 538560 1-minute rainfall data from 1109 stations within the radar's 100 km detection range from 2018 to 2020, three deep learning networks（Z-Rnet, K_DP-Rnet and Pol-Rnet）are designed for radar quantitative precipitation estimation (QPE) based on single and three radar moments, respectively. Furthermore, based on the three networks and with K_DP = 0.5°/km as the threshold to divide the training dataset as heavy, light, and all rain data, a total of 9 QPE models are built. On the basis of using the common mean square error as the loss function, a self-defined loss function is proposed by adjusting the weight for different precipitation intensity. Several indexes including ratio deviation, relative deviation, mean square error (MSE), mean absolute error (MAE) and mean relative error (MRE) are then used to evaluate the performance of the models. Finally, three precipitation processes that are respectively dominated by cumulus-stratiform mixed, convective and stratiform clouds are used to test the effect of QPE. The results suggest that the models fitted by deep learning have better QPE results, and the QPE accuracy for the data that are divided into heavy and light rain is better than that for the data that includes all types of rainfall. The MSE, MAE and MRE with the self-defined loss function are improved by 8.62%, 12.52%, 16.34% than that with the traditional mean square error loss function. Among them, the QPE with Pol-Rnet, i.e., Z_H, Z_DR and K_DP are used as input factors, is the best, and the above indexes are respectively increased by 6.82%, 8.43%, 7.22% than that with Z-Rnet, and by 12.33%, 17.61%, 17.26% than that with K_DP-Rnet.

Radar echo extrapolation is one of the main reference criteria for nowcasting, weather modification operation and evaluation of its effectiveness. Therefore, rapid and accurate echo extrapolation technology is always a research hotspot in the field of radar meteorology. In recent years, deep learning-based spatiotemporal sequence prediction models have been widely applied in radar echo extrapolation. However, most of the inputs to these extrapolation networks are only grayscale images converted from the radar echo intensity products shown with 16-level color code pseudo colors, and thus some echo details are lost. The error inevitably increases with the extrapolation time. The initiation, disappearance, movement, and evolution of echoes are closely related to weather background. Based on this consideration, some physical products in the initial zero field of the rapid-refresh multi-scale analysis and prediction system-nowcasting (RMAPS-NOW) developed by the Institute of Urban Meteorology, CMA, are integrated with the raw data of the North China radar mosaic to construct multiple radar cells. Based on convolutional long short-term memory (ConvLSTM) network, a multi-channel radar echo extrapolation (MR-ConvLSTM) network is designed by using the radar cells as inputs. In addition, considering the smoothness of the convolutional algorithm, a self-defined loss function is designed to increase the spatiotemporal weight of the echo intensity for spatiotemporal attenuation correction. A total of 13000 samples of radar mosaic and RMAPS-NOW data during June—September from 2018 to 2021 over the area of (40.65°—41.65°N, 114°—115.4°E) are selected. 80% of the samples, i.e., 10400 samples, are used as the training data; and 20% of the samples, i.e. 2600, are used as the test data. The introduced physical products include u, v wind (1350 m), relative humidity (150 m), and horizontal divergence (1350 m), etc. at multiple altitude levels. Based on ConvLSTM and MR-ConvLSTM with the self-defined loss function, 5 extrapolation models are then trained respectively. Using critical success index (CSI), hit rate (POD), and false alarm rate (FAR) as evaluation indexes, the models are evaluated by the test data. At the values of reflectivity threshold 20, 30, and 35 dBz, the average values of CSI calculated by the MR-ConvLSTM-based and the self-defined function models with integrated physical products are 4.67%, 13.8%, 5.98% higher and the average values of POD are 3.1%, 7.68%, 8.38% higher, while average values of FAR are 6.37%, 8.54%, 10.17% lower than those by the ConvLSTM-based model without integrated physical products, respectively. The model with three physical products introduced (RH, u-wind, v-wind) performs the best for all the indexes, and average CSI and POD are respectively 16.01% and 13.38% higher while FAR is 14.88% lower than those without physical products. From the visualization cases of model application, it can also be seen that the introduction of physical quantities effectively improves the accuracy of radar echo extrapolation. These results indicate that the MR-ConvLSTM models and self-defined loss function have robust generalization ability.

Raindrop size distribution (DSD) is crucial in the study of precipitation microphysics, but the parameters that characterize the DSD is not easy to be derived accurately. Wind profile radar (WPR) provides rich data with high spatiotemporal resolution and low attenuation, and the previous studies using WPR power spectral distribution to extract DSD parameters demonstrate certain limitations. In this study, a WPR-based Gamma DSD parameter estimation network (WPR-DSDnet) combined with a neural network (NN) and a long short-term memory (LSTM) network is designed, and then a WPR-DSDnet-based model is trained to retrieve the Gamma DSD parameters including normalized intercept parameter (N_w) and mass-weighted average diameter (D_m) by means of the high-resolution WPR reflectivity factor and velocity spectral width as the inputs, and the spatiotemporal matching disdrometer data collected from 2017 to 2021 in Beijing as labels. Two precipitation cases are used to validate the performance of the model, and the results demonstrate that the model can retrieve the Gamma DSD parameters effectively, in which the ratio between the estimated and measured values of both lgN_w and D_m exceeds 99%. The estimation accuracy of lgN_w is slightly better than that of D_m due to smaller relative deviation of lgN_w although the absolute deviation between the estimated and true values of lgN_w is larger than that of D_m, which are probably caused by the larger distribution range of lgN_w. The vertical distributions of lgN_w closely aligns with variations in precipitation intensity, which hints it is a useful indicator for precipitation intensity change. On the other hand, the distribution of D_m is closely associated with the degree of convection, which is valuable for precipitation recognition and classification. The two parameters extracted by the deep learning-based model can facilitate further in-depth analysis on precipitation characteristics and mechanisms with WPR data.