Aofan Gong

Accurate estimation of surface precipitation with high spatial and temporal resolution is crucial for disaster weather detection and decision-making regarding water resources management. Polarimetric weather radar is an important instrument for quantitative precipitation estimation (QPE). Conventional parametric approaches, such as the radar reflectivity (Z) and rain rate (R) relations, cannot fully represent the spatial and temporal variability of clouds and precipitation due to parameterization errors and dependence on raindrop size distribution (DSD). Furthermore, these relations estimate rainfall on a grid-by-grid basis, preventing the incorporation of spatial information into precipitation estimation.In recent years, machine learning has made rapid advancements in non-linear fitting and feature extracting. Since 2020, multiple studies constructed MLP or CNN-based QPE models that used polarimetric radar observations to retrieve precipitation. These researches have consistently demonstrated that machine learning algorithms perform better than traditional parametric methods in different regions and climatic conditions(Chen & Chandrasekar, 2021; Li et al., 2023; Osborne et al., 2023; Tian et al., 2020; Zhang et al., 2021; Zhou et al., 2023).The aforementioned studies have highlighted the immense potential of deep learning for radar QPE, but they are based on S-band radar data. Because X-band radar has a shorter wavelength, the electromagnetic scattering characteristics of hydrometeors differ from those of S-band radar, especially for specific differential phase (kdp), which is closely related to rainfall. Furthermore, X-band radars have different spatial resolutions from S-band radars, which indicates that directly applying a model trained with S-band radar data to X-band radar data may introduce biases. Therefore, we develop a CNN-based QPE model using polarimetric measurements from X-band radars and compare its performance against traditional parametric methods. The input data for the CNN model is a matrix with dimensions (6, 9, 9). The matrix is composed of two matrices of size (3, 9, 9), which is the polarimetric measurements from the two lowest scan elevation angles and 9*9 surrounding range gates. This allows the input data to capture the spatial and physical characteristics of the precipitation field. The results reveal that the CNN-based model not only enhances the accuracy of radar QPE with a diminished bias but also provides a more precise depiction of the spatial distribution of precipitation in comparison to conventional methods.

Implementing the 3-Dimensional Variational (3DVar) data assimilation technique using high-density automatic weather station (AWS) observations substantially improves the precipitation simulation and forecast capabilities in the Weather Research and Forecasting (WRF) model. Given the impact of spatial distribution and quantity of observation data on assimilation effectiveness, there is a growing need to assimilate the most efficient amount of observation data to improve the precipitation forecast accuracy, especially in the context of the proliferation of data from diverse sources. This study investigates the impacts of spatial density of assimilated data on enhancing model predictions, focusing on a squall line event in Beijing on 2 August 2017 which has approximately 2400 AWSs in the simulation domain. Seven experiment groups assimilating varying proportions of AWS data (3.125, 6.25, 12.5, 25, 50, 75, and 100 percent of total AWSs) were conducted, comprising 10 experiments per group. The results were then compared with the experiment without data assimilation (CTRL) and the observations. Results show that while the WRF model roughly captured the evolution of this event, it overestimated the precipitation amount with significant deviations in precipitation locations. A general positive correlation was observed between the spatial density of assimilated data and the enhancement in model performance. However, there is a notable threshold beyond which additional data ceases to enhance forecast accuracy. The model performs best when the ratio of the number of assimilated AWSs to the model simulated area reaches 1/40 km-2. Moreover, significant variations in improvement effects across experiments within the same group indicate the substantial impact of spatial distribution of assimilated AWSs on forecast outcomes. This study provides a reference for devising more efficient and cost-effective data assimilation strategies in numerical weather prediction.

Weather radars commonly suffer from the data-missing problem that limits their data quality and applications. Traditional methods for the completion of weather radar missing data, which are based on radar physics and statistics, have shown defects in various aspects. Several deep learning (DL) models have been designed and applied to weather radar completion tasks but have been limited by low accuracy. This study proposes a dilated and self-attentional UNet (DSA-UNet) model to improve the completion of weather radar missing data. The model is trained and evaluated on a radar dataset built with random sector masking from the Yizhuang radar observations during the warm seasons from 2017 to 2019, which is further analyzed with two cases from the dataset. The performance of the DSA-UNet model is compared to two traditional statistical methods and a DL model. The evaluation methods consist of three quantitative metrics and three diagrams. The results show that the DL models can produce less biased and more accurate radar reflectivity values for data-missing areas than traditional statistical methods. Compared to the other DL model, the DSA-UNet model can not only produce a completion closer to the observation, especially for extreme values, but also improve the detection and reconstruction of local-scale radar echo patterns. Our study provides an effective solution for improving the completion of weather radar missing data, which is indispensable in radar quantitative applications.

Weather radars commonly suffer from the data-missing problem that limits their data quality and applications. Traditional methods for the completion of weather radar missing data, which are based on radar physics and statistics, have shown defects in various aspects. Several deep learning (DL) models have been designed and applied to weather radar completion tasks but have been limited by low accuracy. This study proposes a dilated and self-attentional UNet (DSA-UNet) model to improve the completion of weather radar missing data. The model is trained and evaluated on a radar dataset built with random sector masking from the Yizhuang radar observations during the warm seasons from 2017 to 2019, which is further analyzed with two cases from the dataset. The performance of the DSA-UNet model is compared to two traditional statistical methods and a DL model. The evaluation methods consist of three quantitative metrics and three diagrams. The results show that the DL models can produce less biased and more accurate radar reflectivity values for data-missing areas than traditional statistical methods. Compared to the other DL model, the DSA-UNet model can not only produce a completion closer to the observation, especially for extreme values, but also improve the detection and reconstruction of local-scale radar echo patterns. Our study provides an effective solution for improving the completion of weather radar missing data, which is indispensable in radar quantitative applications.

Weather radar plays an important role in accurate weather monitoring and modern weather forecasting, as it can provide timely and refined weather forecasts for the public and for decision makers. Deep learning has been applied in radar nowcasting tasks and has exhibited a better performance than traditional radar echo extrapolation methods. However, current deep learning-based radar nowcasting models are found to suffer from a spatial “blurry” effect that can be attributed to a deficiency in spatial variability representation. This study proposes a Spatial Variability Representation Enhancement (SVRE) loss function and an effective nowcasting model, named the Attentional Generative Adversarial Network (AGAN), to alleviate this blurry effect by enhancing the spatial variability representation of radar nowcasting. An ablation experiment and a comparison experiment were implemented to assess the effect of the generative adversarial (GA) training strategy and the SVRE loss, as well as to compare the performance of the AGAN and SVRE loss function with the current advanced radar nowcasting models. The performances of the models were validated on the whole test set and inspected in two storm cases. The results showed that both the GA strategy and SVRE loss function could alleviate the blurry effect by enhancing the spatial variability representation, which helps the AGAN to achieve better nowcasting performance than the other competitor models. Our study provides a feasible solution for high-precision radar nowcasting applications.