A Deep Learning-based Radar Echo Extrapolation Study with Fusing Multi-Source Data
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
Accurate nowcasting provides key information for disaster weather warning and artificial weather operations. Nowcasting is mostly based on radar echo extrapolation, where the evolution of echoes results from complex interactions among cloud systems and various thermal-dynamic features of meteorological elements. In this study, a multi-channel radar echo extrapolation architecture (UGR) based on UNet and GAN with Radarcells is proposed, and a self-defined loss function combining weighted mean square error and binary cross-entropy is designed by introducing a penalty term in the GAN network to improve model training. Four sets of radar units (Radarcells) are encoded using radar combined reflectivity mosaic data and four types of physical elements acquired from the Beijing rapid updating circular numerical forecasting system (CMA-BJ). Then, 20 UGR-based models are trained with the Radarcells as inputs for sequential forecasting every six minutes to achieve a rolling 120-minute echo extrapolation. To verify the improvement effect after introducing weather background on echo extrapolation, UNet-based models and ConvLSTM-based models that use only radar echoes as input are trained for comparison, respectively. Critical Success Index (CSI), Probability of Detection (POD), False Alarm Rate (FAR), and Bias score (Bias) are used to evaluate the model on the test set. The results show that the echo intensity range and spatiotemporal evolution predicted by UGR-based models incorporating weather background information perform better than the UNet-based and ConvLSTM-based models driven solely by radar data, especially in predicting strong echoes more accurately. Under the reflectivity thresholds of 25 and 35 dBz, the average values of CSI, FAR, POD, and Bias calculated by UGR-based models are improveed by 10.5%, 8.6%, 10.3%, 4.8%, and 13.4%, 4.6%, 11.0%, and 7.4%, respectively, compared to those by UNet-based models. The study suggests that extrapolation models incorporating weather background information can effectively improve issues of echo blurriness and insufficient echo formation-dissipation compared with previous deep learning-based extrapolation models.
Keywords:
Project Supported:
Clc Number:
You are thevisitors
Supported by:Beijing E-Tiller Technology Development Co., Ltd.
Supported by:Beijing E-Tiller Technology Development Co., Ltd.