基于BP神经网络和支持向量机的降水量空间插值对比研究——以甘肃省为例
|
摘要
为探究降水数据产品的高精度空间插值方法及其应用差异,以BP神经网络和支持向量机模型为研究对象,选取甘肃省为研究区域,构建降水量空间插值模型,分别完成基于两种模型的甘肃省降水量空间插值结果,并划分西区、中区、东区3个建模分区,对比分析两种模型空间插值的精度和应用差异。结果表明:支持向量机模型的插值精度显著高于BP神经网络模型,且支持向量机模型在降水空间分布中能体现更多细节;西区平均相对误差最大,其中,BP神经网络模型32.32%,支持向量机模型仅23.74%;东区平均相对误差最小,BP神经网络为8.28%,支持向量机模型为6.15%;另外,分区建模的插值精度有所提高,但两种模型的提高幅度存在差异,BP神经网络的平均相对误差降低了5.08%,支持向量机模型仅降低0.66%,表明支持向量机模型更加稳定,对影响降水量的经纬度和高程等因子自身变异性的适应能力更强。此研究解决了常用的反距离权重、样条函数、克里金插值等方法在降水量插值过程中准确性差,精度低的问题,为提高降水量空间插值的精度提供了新方法和思路。
In order to explore the high accuracy spatial interpolation method and its application difference of precipitation data products,current study utilized BP neural network and SVM model as research objects,selects Gansu Province as a research area,builds the precipitation spatial interpolation model,completes the model based on two models of Gansu Province,and divided the three modeling zones in the west,middle and eastern regions,and compared the accuracy and application differences between the two models in the space interpolation. The results showed that the SVM model interpolation accuracy was significantly higher than the BP neural network model,and the SVM model could reflect more detail in the spatial distribution of precipitation; the average relative error of the west region was the largest,of which BP neural network model 32. 32% Vector machine model was only 23. 74%,the average relative error was the smallest in the east,BP neural network was 8. 28%,and SVM model was 6. 15%. In addition,the interpolation accuracy of the partition modeling was improved,but there was difference between the two models. The average relative error of BP neural network was reduced by 5. 08% and the support vector machine model was only reduced by 0. 66%. This shows that the support vector machine model is more stable and adaptable to the variability of factors such as latitude,longitude and elevation that influence precipitation. This study solves the problems of poor and low accuracy of commonly used methods of inverse distance weight,spline function and Kriging interpolation in precipitation interpolation,and provides a new method and idea for improving the accuracy of spatial interpolation of precipitation.
In order to explore the high accuracy spatial interpolation method and its application difference of precipitation data products,current study utilized BP neural network and SVM model as research objects,selects Gansu Province as a research area,builds the precipitation spatial interpolation model,completes the model based on two models of Gansu Province,and divided the three modeling zones in the west,middle and eastern regions,and compared the accuracy and application differences between the two models in the space interpolation. The results showed that the SVM model interpolation accuracy was significantly higher than the BP neural network model,and the SVM model could reflect more detail in the spatial distribution of precipitation; the average relative error of the west region was the largest,of which BP neural network model 32. 32% Vector machine model was only 23. 74%,the average relative error was the smallest in the east,BP neural network was 8. 28%,and SVM model was 6. 15%. In addition,the interpolation accuracy of the partition modeling was improved,but there was difference between the two models. The average relative error of BP neural network was reduced by 5. 08% and the support vector machine model was only reduced by 0. 66%. This shows that the support vector machine model is more stable and adaptable to the variability of factors such as latitude,longitude and elevation that influence precipitation. This study solves the problems of poor and low accuracy of commonly used methods of inverse distance weight,spline function and Kriging interpolation in precipitation interpolation,and provides a new method and idea for improving the accuracy of spatial interpolation of precipitation.
引文
[1]Hutchinson.Interpolation of Rainfall Data with Thin Plate Smoothing Splines-PartⅠ:Two Dimensional Smoothing of Data With Short Range Correlation[J].Journal of Geographic Information and Decision Analysis,1998,2(2):139-151
[2]Hutchinson.Interpolation of Rainfall Data with Thin Plate Smoothing Splines-PartⅡ:Analysis of Topographic Dependence[J].Journal of Geographic Information and Decision Analysis,1998,2(2):152-167.
[3]曹立国,刘普幸,张克新,等.锡林郭勒盟草地对气候变化的响应及其空间差异分析[J].干旱区研究,2011,28(5):789-793
[4]周锁铨,薛根元,周丽峰,等.基于GIS降水空间分析的逐步插值方法[J].气象学报,2006,64(1):100-111.
[5]封志明,杨艳昭,丁晓强,等.气象要素空间插值方法优化[J].地理研究,2004,23(3):357-364.
[6]张旭东,辛吉武,王润元,等.基于DEM的甘肃省降水资源分析[J].干旱地区农业研究,2009,27(5):1-5.
[7]李爱华,柏延臣.基于贝叶斯最大熵的甘肃省多年平均降水空间化研究[J].中国沙漠,2012,32(5):1408-1416.
[8]怀保娟,李忠勤,孙美平,等.SRM融雪径流模型在乌鲁木齐河源区的应用研究[J].干旱区地理,2013,36(1):41-48.
[9]杨劲松,姚荣江,刘广明,等.黄河三角洲地区土壤盐分的空间变异性及其Co Kriging估值[J].干旱区研究,2006,23(3):439-445.
[10]陈鹏翔,毛炜峄.基于GIS的新疆气温数据栅格化方法研究[J].干旱区地理,2012,35(3):438-445.
[11]魏智,金会军,蓝永超,等.基于Kriging插值的黑河分水后中游地下水资源变化[J].干旱区地理,2009,32(2):196-203.
[12]刘新安,于贵瑞,范辽生,等.中国陆地生态信息空间化技术研究Ⅲ-温度、降水等气候要素[J].自然资源学报,2004,19(6):818-825.
[13]马学款,普步次仁,唐书乙,等.人工神经网络在西藏中短期温度预报中的应用[J].高原气象,2007,26(3):491-495.
[14]邵月红,张万昌,刘永和,等.BP神经网络在多普勒雷达降水量的估测中的应用[J].高原气象,2009,28(4):846-853.
[15]李法然,周之栩,陈卫锋,等.湖州市大雾天气的成因分析及预报研究[J].应用气象学报,2005,16(6):794-803.
[16]胡广义,张秋文,张勇传,等.基于BP人工神经网络的分布式降雨量插值估算[J].华中科技大学学报,2009,37(4):107-110.
[17]冯汉中,陈永仪.支持向量机回归方法在实时业务预报中的应用[J].气象,2005,31(1):41-45.
[18]李智才,马文瑞,李素敏,等.支持向量机在短期气候预测中的应用[J].气象,2006,32(5):57-61.
[19]冯汉中,陈永仪,成永勤,等.双流机场低能见度天气预报方法研究[J].应用气象学报,2006,17(1):94-99.
[20]张旭东,辛吉武,王润元,等.基于DEM的甘肃省降水资源分析[J].干旱地区农业研究,2009,27(5):1-5.
[21]安兴琴,马安青,王惠林,等.基于GIS的兰州市大气污染空间分析[J].干旱区地理,2006,29(4),576-581.
[22]李纯斌.草原综合顺序分类系统第二级亚类的定量化研究—以甘肃省为例[D].兰州:甘肃农业大学,2012.
[23]郭婧,柳小妮,任正超.基于GIS模块的气象数据空间插值方法新改进—以甘肃省为例[J].草原与草坪,2011,31(4):41-45.
[2]Hutchinson.Interpolation of Rainfall Data with Thin Plate Smoothing Splines-PartⅡ:Analysis of Topographic Dependence[J].Journal of Geographic Information and Decision Analysis,1998,2(2):152-167.
[3]曹立国,刘普幸,张克新,等.锡林郭勒盟草地对气候变化的响应及其空间差异分析[J].干旱区研究,2011,28(5):789-793
[4]周锁铨,薛根元,周丽峰,等.基于GIS降水空间分析的逐步插值方法[J].气象学报,2006,64(1):100-111.
[5]封志明,杨艳昭,丁晓强,等.气象要素空间插值方法优化[J].地理研究,2004,23(3):357-364.
[6]张旭东,辛吉武,王润元,等.基于DEM的甘肃省降水资源分析[J].干旱地区农业研究,2009,27(5):1-5.
[7]李爱华,柏延臣.基于贝叶斯最大熵的甘肃省多年平均降水空间化研究[J].中国沙漠,2012,32(5):1408-1416.
[8]怀保娟,李忠勤,孙美平,等.SRM融雪径流模型在乌鲁木齐河源区的应用研究[J].干旱区地理,2013,36(1):41-48.
[9]杨劲松,姚荣江,刘广明,等.黄河三角洲地区土壤盐分的空间变异性及其Co Kriging估值[J].干旱区研究,2006,23(3):439-445.
[10]陈鹏翔,毛炜峄.基于GIS的新疆气温数据栅格化方法研究[J].干旱区地理,2012,35(3):438-445.
[11]魏智,金会军,蓝永超,等.基于Kriging插值的黑河分水后中游地下水资源变化[J].干旱区地理,2009,32(2):196-203.
[12]刘新安,于贵瑞,范辽生,等.中国陆地生态信息空间化技术研究Ⅲ-温度、降水等气候要素[J].自然资源学报,2004,19(6):818-825.
[13]马学款,普步次仁,唐书乙,等.人工神经网络在西藏中短期温度预报中的应用[J].高原气象,2007,26(3):491-495.
[14]邵月红,张万昌,刘永和,等.BP神经网络在多普勒雷达降水量的估测中的应用[J].高原气象,2009,28(4):846-853.
[15]李法然,周之栩,陈卫锋,等.湖州市大雾天气的成因分析及预报研究[J].应用气象学报,2005,16(6):794-803.
[16]胡广义,张秋文,张勇传,等.基于BP人工神经网络的分布式降雨量插值估算[J].华中科技大学学报,2009,37(4):107-110.
[17]冯汉中,陈永仪.支持向量机回归方法在实时业务预报中的应用[J].气象,2005,31(1):41-45.
[18]李智才,马文瑞,李素敏,等.支持向量机在短期气候预测中的应用[J].气象,2006,32(5):57-61.
[19]冯汉中,陈永仪,成永勤,等.双流机场低能见度天气预报方法研究[J].应用气象学报,2006,17(1):94-99.
[20]张旭东,辛吉武,王润元,等.基于DEM的甘肃省降水资源分析[J].干旱地区农业研究,2009,27(5):1-5.
[21]安兴琴,马安青,王惠林,等.基于GIS的兰州市大气污染空间分析[J].干旱区地理,2006,29(4),576-581.
[22]李纯斌.草原综合顺序分类系统第二级亚类的定量化研究—以甘肃省为例[D].兰州:甘肃农业大学,2012.
[23]郭婧,柳小妮,任正超.基于GIS模块的气象数据空间插值方法新改进—以甘肃省为例[J].草原与草坪,2011,31(4):41-45.