摘要
基于人工神经网络方法,利用海面水温、海面风速以及海面气压反演南海近海面气温,采用的基础数据集是国际综合海洋-大气数据集(International Comprehensive Ocean-Atmosphere Data Set,2.4 Release,ICOADS2.4)1981—2008年的观测资料,其中1981—2000年的观测资料用来建立模型,2001—2008年的观测资料用来进行模型检验。采用的人工神经网络方法是引入动量因子并采用批处理梯度下降法的BP(Back propagation)算法。试验结果表明,基于人工神经网络建立的近海面气温反演方法明显优于多元线性回归方法,尤其是在春季和冬季,海面水温、海面风速以及海面气压与近海面气温之间存在较强的非线性关系,人工神经网络的优势更加明显。总体而言,人工神经网络在各月的反演效果较均衡,均方根误差介于1.5—1.8℃之间,平均绝对误差为1.1—1.3℃。
Based on artificial neural network(ANN),the authors retrieved near-surface air temperature(AT) from sea surface temperature(SST),wind speed(WS) and sea level pressure(SLP) of the International Comprehensive Ocean-Atmosphere Dataset(ICOADS).Modeling sample spans from 1981 to 2000,while validating sample spans from 2001 to 2008.The adopted ANN introduces momentum factor to back propagation(BP) algorithm to escape from local extremes.In addition,batch processing gradient descent method was used to remove the effect of sequential training.Retrieving results in the South China Sea(SCS) demonstrates that ANN is better than multi-factor linear regression,especially for coastal areas during spring and winter,where strong non-linear relation exists between SST,WS,SLP and AT.In conclusion,ANN behaves similarly for each month,with root mean square error(RMSE) between 1.5℃ and 1.8℃ and mean absolute error(MAE) between 1.1℃ and 1.3℃.
出处
《热带海洋学报》
CAS
CSCD
北大核心
2012年第2期7-14,共8页
Journal of Tropical Oceanography
基金
国家自然科学基金项目(41030854
40906015
40906016)
国家科技支撑计划项目(2011BAC03B02-01-04)
关键词
人工神经网络
BP算法
多元线性回归
Artificial neural network
BP algorithm
Multi-factor linear regression