摘要
针对GPM降水产品低空间分辨率的缺陷,基于多源环境数据提取了表征区域地形、水汽、地表覆被、海陆位置等20个辅助变量,采用遗传算法(GA)提取最相关变量做为降尺度因子,运用深层神经网络(DNN)算法建立2019年江苏省GPM逐月产品的降尺度模型,并利用地面站点数据进行验证。结果表明,GA算法能很好地排除冗余信息、约简降尺度模型;基于地面资料的独立验证表明降尺度后数据的决定系数R^(2)介于0.43~0.93之间,相对误差M_(RE)在7.47%~23.77%之间,平均绝对误差M_(AE)、均方根误差R_(MSE)分别为2.18~26.84、3.23~29.54 mm;与GPM原始产品相比,降尺度后的平均R^(2)增加了0.05,M_(AE)、R_(MSE)分别减小了1.44、2.04 mm,M_(RE)降低了2.97%。提出的GA-DNN降尺度模型可为粗级降水产品的精细化提供技术参考。
Aiming at the low spatial resolution of GPM precipitation products,20 auxiliary variables representing regional terrain,water vapor,surface cover and land sea location were extracted based on multi-source environmental data.Then,genetic algorithm(GA)was used to extract the most relevant variables as downscaling factors.Finally,deep neural network(DNN)was used to establish the downscaling model of GPM monthly products in Jiangsu Province in 2019,and the model was verified by ground station data.The results show that GA algorithm can eliminate redundant information and reduce downscaling model.Independent verification based on ground data shows that the R^(2) of downscaling data is between 0.43-0.93,M_(RE) is between 7.47%-23.77%,M_(AE) and R_(MSE) are 2.18-26.84 mm and 3.23-29.54 mm,respectively.Compared with the original GPM product,the average R^(2) increased by 0.05,M_(AE) and R_(MSE) decreased by 1.44 mm and 2.04 mm,and M_(RE) decreased by 2.97%.The proposed GA-DNN downscaling model can provide technical reference for the refinement of coarse precipitation products.
作者
徐秀丽
张艺铭
张俊瑞
张彩云
吴泽雄
XU Xiu-li;ZHANG Yi-ming;ZHANG Jun-rui;ZHANG Cai-yun;WU Ze-xiong(Taizhou Branch of Jiangsu Provincial Bureau of Hydrology and Water Resources Survey,Taizhou 225300,China)
出处
《水电能源科学》
北大核心
2021年第10期18-21,62,共5页
Water Resources and Power
