基于流形学习算法的降雨数据时空分布特征提取及重构

Spatiotemporal distribution features of rainfall data extracting and reconstructing based on manifold learning algorithm

【目的】掌握精细化的降雨时空分布特征,对于城市洪涝风险管理水平的提高具有重要的意义。我国近十几年降雨监测站网密集且数据精细程度高,但时间序列较短;历史降雨资料时间序列长,但是精细程度低。【方法】为了更有效地利用历史降雨资料,将流形学习算法引入到历史降雨资料重构中,从高分辨率降雨资料中,提取降雨的时空分别特征,基于该特征,将历史逐6 h的降雨空间数据重构为逐1 h的降雨数据,以满足城市洪涝风险分析的要求。【结果】结果表明,该方法重构数据高值区与实测值的平均误差在15%以内,低值区在20%以内,比传统插值处理的数据高值区误差降低了45%~85%,低值区降低了10%~40%。【结论】利用流形学习算法重构的历史空间降雨数据符合各地区降雨时空分布特征,可提高降雨空间数据颗粒度,实现降雨时空分布精细化特征的有效、合理的提取和总结。

[Objective]Mastering the refined spatiotemporal distribution characteristics of rainfall is of great significance for improving the level of urban flood risk management.The quality of rainfall monitoring data in China in the past decade is good,witha dense network of stations and high precision of rainfall data, but the time series is relatively short. [Methods]In order to makemore effective use of historical rainfall data, this study introduces manifold learning algorithm into the reconstruction of historicalrainfall data. From high-resolution rainfall data in the past decade, spatiotemporal features of rainfall are extracted. Based on thisfeature, 6 h interval rainfall spatial data is reconstructed into 1 h interval spatial data to meet the requirements of urban flood riskanalysis. [Results]The result show that: The average error between the high value area of the reconstructed data and the measuredvalue is within 15%, and the low value area is within 20%, the error in the high value area of the data is reduced by 45%~85%, while the error in the low value area is reduced by about 10%~40%, compared to traditional method. [Conclusion]Thereconstructed historical spatial rainfall data using manifold learning algorithms conforms to the spatiotemporal distribution characteristicsof rainfall in various regions, which can improve the granularity of rainfall spatial data and achieve effective and reasonableextraction and summary of refined features of rainfall spatiotemporal distribution.