Estimation of trend and random components of conditional random field using Gaussian process regression
Estimation of trend and random components of conditional random field using Gaussian process regression
复制标题
DOI：
10.1016/j.compgeo.2021.104179
复制
发表时间：
2021-08
影响因子：
5.3
通讯作者：
I. Yoshida;Y. Tomizawa;Y. Otake
中科院分区：
文献类型：
--
作者：
I. Yoshida;Y. Tomizawa;Y. Otake
提出了一种利用多个高斯随机场叠加高斯过程回归同时估计任意位置土壤性质趋势和随机分量的方法。将该方法应用于三个合成数据集趋势分量的一维空间分布估计。对高斯、马尔可夫和二值噪声三种协方差函数的比较表明，高斯协方差最适合于趋势估计。利用极大似然估计方法对样本随机分量的波动尺度和标准差进行估计。该方法还应用于基于锥突试验实测数据的趋势分量和随机分量的三维空间分布估计。结果表明，可以估计任意位置的趋势和随机分量。基于赤池信息准则和贝叶斯信息准则，发现whittle - mat日新月异的协方差函数比马尔可夫协方差函数更适合于圆锥侵彻试验数据随机分量的估计。
A method is proposed for simultaneously estimating the trend and random component of soil properties at arbitrary locations using Gaussian process regression with the superposition of multiple Gaussian random fields. The proposed method is applied to the estimation of the one-dimensional spatial distributions of the trend component of three synthetic datasets. A comparison of three covariance functions, namely Gaussian, Markovian, and binary noise, indicates that Gaussian covariance is most suitable for trend estimation. The scale of fluctuation and the standard deviation of the random component of the examples are estimated using the maximum likelihood estimation method. The proposed method is also applied to the estimation of the three-dimensional spatial distribution of the trend and random components based on measured cone penetration test data. It is shown that the trend and random components at arbitrary locations can be estimated. The Whittle-Matérn covariance function is found to be more suitable than the Markovian covariance function for the estimation of the random component of the cone penetration test data based on the Akaike information criterion and the Bayesian information criterion.