基于INLA-SPDE贝叶斯空间模型预测土壤有机碳含量  

Prediction of Soil Organic Carbon Contents Based on INLA-SPDE Bayesian Spatial Model

作　　者：伍维模[1,2] 吴嘉平[3] 王家强[2] 曹琦[2] WU Weimo

机构地区：[1]浙江大学环境与资源学院,浙江杭州310058 [2]塔里木大学植物科学学院,新疆阿拉尔843300 [3]浙江大学海洋学院,浙江舟山316021

出　　处：《地理空间信息》2020年第12期92-96,I0006,I0007,共7页Geospatial Information

基　　金：国家自然科学基金资助项目(40961028)。

摘　　要：采用嵌套拉普拉斯逼近积分——随机偏微分方程(INLA-SPDE)构建贝叶斯空间模型,用该模型对塔里木盆地北缘土壤有机碳含量的空间分布进行预测;采用Python语言PyMC库和R语言spBayes包构建基于马尔科夫链蒙特卡洛(MCMC)模拟的贝叶斯空间模型。比较了基于MCMC和INLA-SPDE两类贝叶斯空间模型的推断结果、预测准确性和计算时间,结果表明基于INLA-SPDE与MCMC的有机碳含量的贝叶斯空间模型具有相似的参数后验分布、后验预测分布以及预测准确度;INLA-SPDE模型比MCMC模型具有更快的运算速度。In this study,we used the integrated nested Laplace approximations with stochastic partial differential equation(INLA-SPDE)Bayesian spatial model to predict soil organic carbon contents in northern Tarim Basin,Xinjiang Province,China.Then,we compared the parameter posterior distribution,organic carbon contents posterior predictive distribution,computation times,as well as prediction accuracy between INLA-SPDE and Markov chain Monte Carlo(MCMC)approaches.R package R-INLA was used to construct the INLA-SPDE Bayesian spatial model.R package spBayes and Python module PyMC were used to construct soil organic carbon Bayesian spatial model which was implemented through MCMC method.Results show that prediction accuracies of soil organic carbon contents of Bayesian spatial models between INLA-SPDE and MCMC(using spBayes and PyMC to construct the Bayesian spatial model,respectively)are almost the same.In addition,spatial distribution patterns of posterior meansof soil organic carbon contents predicted by INLA-SPDE and MCMC are also similar.The INLA-SPDE is more efficient and takes much less computation times than MCMC.

关 键 词：贝叶斯空间模型 嵌套拉普拉斯逼近积分 随机偏微分方程 马尔科夫链蒙特卡洛 土壤有机碳 空间预测 

分 类 号：P208[天文地球—地图制图学与地理信息工程]