基于三维卷积的双偏振雷达定量降水估测研究

Research on Quantitative Precipitation Estimation Using Dual-Polarization Radar Based on 3D Convolution

关键词:  定量降水估测  双偏振雷达  四维数据集  三维卷积  深度学习

Key words:  quantitative precipitation estimation  dual-polarization radar  four-dimensional dataset  three- dimensional convolution  deep learning

摘要:

强降水常常引发洪涝灾害，因此提高雷达定量降水估测（QPE）准确性对减轻灾害损失具有重要意义。利用广州双偏振雷达数据与自动站雨量数据生成四维数据集，设计了 3DPoly-QPENet、3DTime-QPENet、3DEcho-QPENet三种三维卷积 QPE 模型进行比较试验。通过测试集的性能评估和典型暴雨个例的检验，得出以下结论：（1）四维数据集相较于传统的三维数据集，在捕捉降水分布特征和提升 QPE 拟合效果方面提供了更多的可能性；（2）三种三维卷积 QPE 模型呈现出与结构设计紧密相关的性能差异，其中 3DPoly-QPENet 在中等降水量区间（15~20 mm·h-1）的平均绝对误差（MAE）较另两种模型平均降低 13%；3DTime-QPENet 在高降水量事件（>50 mm·h-1）的 MAE较另两种模型平均降低 8.1%；3DEcho-QPENet全局误差均衡性最优，总体 MAE较另两种模型平均降低 20.4%；（3）三维卷积模型均系统性优于传统 Z-R 关系方法，平均 RMSE 降低 46.6%，MAE 下降48.6%，CC提升21.4%。

Abstract:

Heavy precipitation often triggers flooding disasters. Therefore, enhancing radar-based quantitative precipitation estimation(QPE) accuracy is critical for disaster mitigation. This study employs Guangzhou dual-polarization radar data and automatic weather station rainfall data to construct a four- dimensional dataset. Three three-dimensional convolutional QPE models—namely, 3DPoly-QPENet, 3DTime-QPENet, and 3DEcho-QPENet—were designed and evaluated through comparative experiments. Based on test-set performance assessments and validation with typical heavy rainfall cases, we draw the following condusions: (1) Compared to traditional three-dimensional datasets, the four-dimensional dataset better captures precipitation distribution characteristics and improves QPE fitting accuracy. (2) The three three-dimensional convolutional QPE models exhibit performance differences tied to their structural designs. Specifically, 3DPoly-QPENet reduces the mean absolute error (MAE) by an average of 13% in moderate precipitation (15-20 mm · h?1) compared to the other two models. 3DTime-QPENet achieves an average MAE reduction of 8.1% in high-intensity precipitation events (>50 mm · h ? 1). 3DEcho-QPENet shows the best global error balance, with an overall MAE reduction of 20.4% on average. (3) All three three-dimensional convolutional models surpass the traditional Z-R relationship method, reducing the root mean square error (RMSE) by an average of 46.6%, lowering MAE by 48.6%, and increasing the correlation coefficient (CC) by 21.4%.