Please wait a minute...
浙江大学学报（农业与生命科学版）
浙江大学学报(农业与生命科学版)  2013, Vol. 39 Issue (6): 636-644    DOI: 10.3785/j.issn.1008-9209.2013.05.231
农业科学     
基于贝叶斯最大熵的多因子空间属性预测新方法
杨勇*， 张楚天， 贺立源
（华中农业大学资源与环境学院,农业部长江中下游耕地保育重点实验室，武汉 430070）
New multifactor spatial prediction method based on Bayesian maximum entropy
YANG Yong*, ZHANG Chutian, HE Liyuan
(Key Laboratory of Arable Land Conservation (Middle and Lower Reaches of Yangtse River), Ministry of Agriculture, College of Resources and Environment, Huazhong Agricultural University, Wuhan 430070, China)
 全文: PDF(5388 KB)   HTML ( ) 
摘要： 为了在空间数据预测时充分利用样点和环境数据，提出了在贝叶斯最大熵方法框架下将经典地统计方法与环境相关法结果融合、利用多源数进行空间预测的新方法；并以湖北省京山县土壤有机质含量为例，验证该方法的可行性.以由数字高程模型（digital elevation model, DEM）生成的各种相关地形因子作为环境数据，并分为密集建模集Ⅰ（330个样点）和稀疏建模集Ⅱ（100个样点），分别用普通克里金法和本文所提方法进行土壤有机质空间预测，用预留的50个样点进行精度分析.结果表明：本文所提方法的预测精度较普通克里金法的高，其Ⅰ和Ⅱ 2组建模集精度分别提高了10.95%和22.72%，特别在样点较稀疏时，在相关环境因子的辅助下，精度提高幅度更大.说明将经典地统计方法与环境相关法结果相融合的多因子空间属性预测方法使预测结果既能体现样点的空间自相关，又能体现被预测属性与其他属性间的相关性.
Abstract: The spatial distributions of soil properties (e.g., organic matter and heavy metal content) are vital to soil quality evaluation and regional environment assessment. Currently, the spatial distribution of soil properties is usually predicted with classical geostatistics or environmental correlation. These two methods are different in theory. Geostatistics is based on spatial correlation of sampling points. However, it contains some deficiencies, such as the lack of effective utilization of environmental information, the smoothing effect of predicted results, difficult to meet the assumption of single point to multipoint Gaussian distribution etc. On the other hand, the theoretical basis of environmental correlation is based on the relationship between soil and environment, but it ignores the spatial correlation among sampling points. These two methods complement each other. Thus, it is very important to study how to integrate these two methods, so that the spatial correlation among sampling points and the relationship between soil and environmental factors can both be used to improve the prediction accuracy. We propose a new spatial prediction method based on the theory of Bayesian maximum entropy (BME), which is one of the most wellknown modern spatiotemporal geostatistical techniques. The main objective is to incorporate the results of classical geostatistics and quantitative soillandscape model in the BME framework. The result of ordinary Kriging was taken as the priori probability density function (pdf), as well as the sampling data as hard data, and the results of environmental correlation as soft data. Posterior pdf is calculated with priori pdf, hard data and soft data. According to the posterior pdf, the predicted values of nonsampling points could be obtained, which not only contained the spatial correlation between sample points, but also took into account the relationship between soil properties and the environment. Meanwhile, the soil organic matter contents in Jingshan County of Hubei Province were used as experimental data. Six environmental factors closely related with soil organic matter content, including relative elevation (Hr), topographic relief (QFD), slope (β), variable rate of slope (SOS), topographic wetness index (ψ) and surface roughness (M), were used as auxiliary variables to produce soft data for prediction. To evaluate the advantages of this method and to compare with Kriging for different sampling density, the analysis and comparison were conducted on two groups of sampling points, which were a group with high spatial density (group Ⅰ, 330 points) and the other with low spatial density (group Ⅱ, 100 points) respectively. The root mean square error (RMSE) and the relative improvement of the accuracy (RI) are computed from 50 estimation and observation values. The results showed that the prediction of BME had higher accuracy (RI values were increased by 10.95% and 22.72% for group Ⅰ and group Ⅱ respectively) than that from ordinary Kriging, especially in the lowdensity group. Compared with ordinary Kriging, the method proposed in this study has solid theoretical foundation, as it integrates theories of Kriging, environmental relation, and maximum entropy. It is more flexible than Kriging, for example, the parameters used in this method, such as environmental factors and the classification number of predicted variables, can be adjusted according to the users’ needs. Moreover, this method performs better than the ordinary Kriging do for the prediction with fewer sampling points.
出版日期: 2013-11-20
CLC:  S 158.2  
服务　　
把本文推荐给朋友
加入引用管理器
E-mail Alert
RSS
作者相关文章　　
杨勇*
张楚天
贺立源

引用本文:

杨勇*， 张楚天， 贺立源. 基于贝叶斯最大熵的多因子空间属性预测新方法[J]. 浙江大学学报(农业与生命科学版), 2013, 39(6): 636-644.

YANG Yong*, ZHANG Chutian, HE Liyuan. New multifactor spatial prediction method based on Bayesian maximum entropy. Journal of Zhejiang University (Agriculture and Life Sciences), 2013, 39(6): 636-644.

链接本文:

http://www.zjujournals.com/agr/CN/10.3785/j.issn.1008-9209.2013.05.231    或    http://www.zjujournals.com/agr/CN/Y2013/V39/I6/636

null
No related articles found!