基于机器学习和贝叶斯模型平均的多源降水融合方法

孟锐; 陈云瑶; 李彬权; 肖洋; 张汇明; 张涛涛; 李匡

doi:10.13928/j.cnki.wrahe.2026.03.012

水利水电技术（中英文） ›› 2026, Vol. 57 ›› Issue (3) : 167 -178. DOI: 10.13928/j.cnki.wrahe.2026.03.012

水文水资源

基于机器学习和贝叶斯模型平均的多源降水融合方法

孟锐 ¹ ,
陈云瑶 ²^,³ ,
李彬权 ²^,³ ,
肖洋 ⁴^,⁵ ,
张汇明 ²^,⁴ ,
张涛涛 ⁵ ,
李匡 ⁶

作者信息 +

Multi-source precipitation fusion method based on machine learning and Bayesian model averaging

Rui MENG ¹ ,
Yunyao CHEN ²^,³ ,
Binquan LI ²^,³ ,
Yang XIAO ⁴^,⁵ ,
Huiming ZHANG ²^,⁴ ,
Taotao ZHANG ⁵ ,
Kuang LI ⁶

Author information +

文章历史 +

PDF (14441K)

摘要

【目的】降水输入误差是洪水预报误差的主要来源，融合多源降水信息是降低降水输入误差的重要手段。高时空分辨率卫星降水产品能够更好地捕捉降水事件的时空分布信息，但因其点估计误差较大导致其应用受限。提出一种基于机器学习和贝叶斯模型平均的多源降水融合方法，以提高降水数据在时空尺度上的精度。【方法】首先采用双线性插值方法将3种卫星降水产品(GSMaP、IMERG和PERSIANN)进行空间降尺度，再利用轻量梯度增强机(LGBM)算法进行降水偏差校正，最后基于季节尺度贝叶斯模型平均(BMA)融合校正后的降水产品，得到精度更高的降水数据。【结果】选择赣江上游峡山站以上流域进行实例验证，结果表明：(1)融合后降水在6个评价指标上均明显优于原始卫星产品(均方根误差为4.73 mm、相关系数为0.92、误报率为0.28);(2)与原始卫星产品相比，融合后降水空间分布的准确性显著提高，与地面观测站的一致性更好；(3)单卫星降水校正的结果和多卫星降水融合后的结果在小雨、中雨、大雨、暴雨和大暴雨等5种量级下的均方根误差值均显著低于原始卫星产品。【结论】总体上，多卫星降水融合数据综合了各校正降水产品的优势，在不同降水量级方面均表现较好，成果可为水文模拟和预报提供准确降雨输入数据支撑。

Abstract

[Objective] Precipitation input errors are the major source of flood forecasting errors, and integrating multi-source precipitation information is an important approach to reduce such errors. Satellite precipitation products with high spatiotemporal resolution can better capture the spatiotemporal distribution of precipitation events, but their application is limited due to large point estimation errors. A multi-source precipitation fusion method based on machine learning and Bayesian model averaging(BMA) is proposed to improve the accuracy of precipitation data at spatiotemporal scales. [Methods] Firstly, a bilinear interpolation method was used to perform spatial downscaling of three satellite precipitation products(GSMaP, IMERG, and PERSIANN). Then, the light gradient boosting machine(LGBM) algorithm was utilized for precipitation bias correction. Finally, the corrected precipitation products were fused based on the seasonal-scale BMA to obtain precipitation data with higher accuracy. [Results] The river basin upstream of Xiashan station in the upper reaches of Ganjiang River was selected for case validation. The result showed that:(1) the fused precipitation was significantly better than the original satellite products in all six evaluation indicators(with a root mean square error of 4.73 mm, a correlation coefficient of 0.92, and a false alarm ratio of 0.28).(2) Compared with the original satellite products, the spatial accuracy of the fused precipitation data was significantly improved and showed better consistency with ground observation stations.(3) The root mean square error values of both the single-satellite corrected precipitation and the multi-satellite fused precipitation were significantly lower than those of the original satellite products under five precipitation magnitudes: light rain, moderate rain, heavy rain, rainstorm, and heavy rainstorm. [Conclusion] Overall, the multi-satellite fused precipitation data integrates the advantages of each corrected precipitation product and performs well across different precipitation magnitudes, providing accurate precipitation input data support for hydrological simulation and forecasting.