基于多维特征的区域性暴雨过程相似性检索方法研究

Similarity Retrieval Method for Regional Rainstorm Processes Based on Multidimensional Features

关键词:  多维特征  区域性暴雨  欧式距离  学习感知图像块相似度  相似性检索

Key words:  multidimensional parameters  regional rainstorm  Euclidean distance  learned perceptual image patch similarity  similarity retrieval

摘要:

针对一次区域性暴雨过程，检索与之相似的历史案例，对区域性暴雨过程的评估、防灾减灾和决策服务等具有重要意义。以京津冀地区为例，利用 1961—2023年京津冀 175个气象观测站逐日降水数据，定义和识别了京津冀区域性暴雨过程，提出了一种基于多维特征的区域性暴雨过程相似判别方法。该方法基于最大过程降水量、最大日降水量、暴雨范围、持续时间、平均范围和平均强度等特征，采用欧式距离度量属性维度相似性；基于过程降水量、最大日降水量和平均降水量分布图，采用 LPIPS (learned perceptual image patchsimilarity，学习感知图像块相似度)度量空间维度相似性；将属性和空间相似度加权求和，得出最终综合相似度。通过本方法确定“23·7”暴雨过程最相似的前 3次暴雨过程为“63·8”、“96·8”和“16·7”，与专家经验分析结果一致。进一步采用检索覆盖率和不同强度等级过程随机检索进行验证，结果表明与现有常见方法相比，本方法检索结果覆盖率最高，且检索结果具有合理性。

Abstract:

For the assessment of regional rainstorm processes, disaster prevention and mitigation efforts, and decision-making services, it is crucial to retrieve historical cases with similar characteristics and compare their meteorological parameters. With a focus on the Beijing-Tianjin-Hebei region, we used daily precipitation data during 1961—2023 from 175 meteorological observation stations in this area, and identified 517 regional rainstorm processes. To measure the similarity of these processes, we proposed a similarity discrimination method for regional rainstorm processes based on multidimensional parameters. The Euclidean distance was used to measure attribute similarity based on parameters such as maximum process precipitation, maximum daily precipitation, influence range, duration, average intensity, and average range. The learned perceptual image patch similarity was used to measure spatial similarity based on the distribution of process precipitation, maximum daily precipitation, and average daily precipitation. The composite similarity score was calculated using a weighted summation of attribute and spatial similarity. The similarity analysis of the rainstorm process in July 2023 reveals that the three most similar rainstorm processes were in August 1963, August 1996, and July 2016, consistent with experts’empirical analyses. Compared with other methods, this method offers the broadest retrieval coverage and reasonable retrieval results, as validated by retrieval coverage rates and the random retrieval of rainstorms across different intensity levels.