基于FY-3B被动微波数据的青藏高原降尺度机器学习雪深反演  

Downscaling machine learning snow depth inversion on the Qinghai-Xizang Plateau based on FY-3B passive microwave data

作　　者：周舟 朱灵龙 张永宏 阚希[2,3,4] 刘旭 曹海啸[2] 王剑庚 ZHOU Zhou;ZHU Linglong;ZHANG Yonghong;KAN Xi;LIU Xu;CAO Haixiao;WANG Jiangeng(School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Internet of Things Engineering,Wuxi University,Wuxi 214105,Jiangsu,China;Collaborative Innovation Center of Atmospheric Environment and Equipment Technology,Nanjing University of Information Science and Technology,Nanjing 210044,China;Internet of Vehicles Laboratory,Wuxi University,Wuxi 214105,Jiangsu,China;School of Atmospheric Physics,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学自动化学院,江苏南京210044 [2]无锡学院物联网工程学院,江苏无锡214105 [3]南京信息工程大学江苏省大气环境与装备技术协同创新中心,江苏南京210044 [4]无锡学院车联网实验室,江苏无锡214105 [5]南京信息工程大学大气物理学院,江苏南京210044

出　　处：《冰川冻土》2024年第2期539-554,共16页Journal of Glaciology and Geocryology

基　　金：国家自然科学基金项目(42305158,42105143);江苏省高等学校基础科学(自然科学)研究面上项目(23KJB170025,21KJB170006);国家重点研发计划项目(2021YFE0116900);无锡市“太湖之光”科技攻关计划基础研究项目(K20231021)资助。

摘　　要：作为中国三大积雪区之一,青藏高原的积雪变化在气候系统、水文地质以及生态环境方面发挥着关键作用。已有的被动微波积雪深度反演方法存在数据分辨率低、不确定性高等问题,不适用于青藏高原复杂的山区地形。因此,本文基于FY-3B被动微波数据开发了青藏高原降尺度雪深反演模型,利用机器学习算法,将筛选后的亮温差作为参数输入,同时引入了高程、经纬度、植被覆盖度、积雪覆盖度和积雪天数等特征,最终进行了500 m分辨率的青藏高原雪深制图。结果显示,极端梯度提升XGBoost算法的决定系数(R^(2))和均方根误差(root mean square error,RMSE)分别为0.762和5.732 cm,明显优于支持向量回归和随机森林算法。从积雪天数、积雪覆盖度和植被覆盖度三个方面探讨了模型精度的变化,结果表明,在积雪天数为30~60 d时,模型表现良好,平均相对误差(mean relative error,MRE)最低为36.79%,RMSE为2.78 cm;随着积雪覆盖度的增加,模型的RMSE逐渐增大,在积雪覆盖度为0.25~0.50时,MRE和RMSE分别达到39.97%和3.12 cm;植被覆盖度对模型精度的影响较为复杂,可能与具体的土地覆盖类型相关,在0.25~0.50范围内模型表现出较高的精度,MRE和RMSE分别为32.77%和2.94 cm。As one of the three major snow regions in China,the snow variation on the Qinghai-Xizang Plateau plays a crucial role in the climate system,hydrogeology,and ecological environment.Existing passive microwave snow depth(SD)inversion methods face challenges such as low resolution and high uncertainty,making them unsuitable for the complex mountainous terrain of the Qinghai-Xizang Plateau.Therefore,this study develops a downscaling SD inversion model for the Qinghai-Xizang Plateau based on FY-3B passive microwave data.Utilizing machine learning algorithms,the selected brightness temperature differences are used as input parameters.Additionally,features such as elevation,latitude,longitude,fractional vegetation cover(FVC),fractional snow cover(FSC),and snow cover days(SCD)are introduced.The final SD mapping is conducted at a resolution of 500 m.Results demonstrate that the Extreme Gradient Boosting(XGBoost)algorithm exhibits superior performance with a coefficient of determination(R^(2))of 0.762 and a Root Mean Square Error(RMSE)of 5.732 cm,outperforming support vector regression and random forest algorithms.We investigated the variation in model accuracy from three perspectives:SCD,FSC,and FVC.The findings indicate that the model performs well when the SCD is between 30 and 60 days,with the lowest Mean Relative Error(MRE)at 36.79% and RMSE at 2.78 cm.As FSC increases,the model’s RMSE gradually increases,reaching 39.97%for MRE and 3.12 cm for RMSE in the FSC range of 0.25 to 0.50.The impact of FSC on model accuracy is complex and may be related to specific land cover types.Within the range of 0.25 to 0.5 for FVC,the model demonstrates higher accuracy,with MRE and RMSE at 32.77% and 2.94 cm,respectively.

关 键 词：青藏高原 降尺度 雪深反演 FY3B-MWRI 机器学习 

分 类 号：P468.025[天文地球—大气科学及气象学] P426.635