一种机器学习海面风场快速融合的方法  

Sea surface wind field smart fusion base on machine learning method

作　　者：张巍 杜超凡[2] 郭安博宇 宋晓姜 沈世莹[2] Zhang Wei;Du Chaofan;Guo Anboyu;Song Xiaojiang;Shen Shiying(National Marine Environmental Forecasting Center,Beijing 100081,China;School of Computer Science and Technology,Ocean University of China,Qingdao 266100,China)

机构地区：[1]国家海洋环境预报中心,北京100081 [2]中国海洋大学计算机科学与技术学院,山东青岛266100

出　　处：《海洋学报》2022年第11期144-158,共15页

基　　金：国家重点研发计划(2018YFC1407001)。

摘　　要：基于多源资料进行海面风场的同化融合或插值融合,目前受到计算能力的较大制约。本文提出在多源卫星数据和ERA-5再分析数据重叠区域,训练基于XGBoost的机器学习ERA-5数据修正融合模型。然后基于该模型快速修正ERA-5数据(机器学习推理)。由于机器学习推理的快速性,ERA-5全区域修正融合的时间仅需2 s左右,可以较小计算代价构建整个海面融合风场。本文以10 m风速、10 m风向、U10分量和V10分量等典型风场变量展开,考虑了海陆分布差异使用陆地掩膜消除陆地区域,分别构建D_S_A_XGBoost、D_S_O_XGBoost、U_V_A_XGBoost、U_V_O_XGBoost 4个ERA-5修正模型,并最终生成海面融合风场。通过修正前后的ERA-5再分析数据与卫星数据进行比较,上述4个模型均减小了ERA-5再分析数据与卫星数据的差距。特别是在风速方面,不论是均方根误差(RMSE)还是绝对误差(MAE)都得到有效降低。在风向方面上,RMSE以及MAE也呈现降低趋势。在利用热带大气海洋观测计划(Tropical Atmosphere Ocean Array,TAO)浮标数据对4种XGBoost模型进行评价发现,U_V_O_XGBoost模型对于ERA-5数据的修正结果最好,其相关性达到0.893,提高了约0.011,结果表明本文在保证风场精度的情况下较大地提高了融合速度。The assimilation fusion or interpolation fusion of the sea surface wind field based on multi-source data is currently restricted by computing power.This paper proposes to train the XGBoost-based machine learning ERA-5data correction fusion model in the overlapping area of the multi-source satellite data and the ERA-5 reanalysis data,and then use the model to quickly correct (machine learning inference) ERA-5 data,of which the ERA-5whole area correction fusion it only takes about 2 seconds.Due to the rapidity of machine learning inference,the entire sea surface fusion wind field can be constructed at a lower computational cost.This paper expands on typica wind field variables such as 10 m wind speed,10 m wind direction,U10 component and V10 component,taking into account the difference in sea and land distribution,using land masks to eliminate land areas,and constructing D_S_A_XGBoost,D_S_O_XGBoost,U_V_A_XGBoost,U_V_O_XGBoost corrections model,and finally generate sea surface fusion wind field.By comparing the ERA-5 reanalysis data before and after the correction with the satellite data,the above four models all reduce the gap between the ERA-5 reanalysis data and the satellite data.Especially in terms of wind speed,both root mean square error (RMSE) and mean absolute error (MAE) are effectively reduced.In terms of wind direction,RMSEand MAEalso show a decreasing trend.Using Tropical Atmosphere Ocean Array (TAO) buoy data to evaluate the four XGBoost models,it is found that the U_V_O_XGBoos model has the best correction results for ERA-5 data,and its correlation reaches 0.893,an increase of about 0.011and the results show that the fusion speed is greatly improved under the condition of ensuring the accuracy of wind field.

关 键 词：XGBoost HY-2B CFOSAT MetOp-B ERA-5 海面风场 

分 类 号：P717[天文地球—海洋科学] P732