ABSTRACT

Kriging interpolation is a spatial interpolation method widely employed in the field of data analytics and prediction of environmental variables, which provides the best linear unbiased prediction of intermediate values. The core principle of Kriging interpolation is searching for data distribution regularity and predicting regionalised variable value, and it can be transferred into two descriptions of learning process: function fitting problem and coefficient optimisation problem. Although these two problems could be solved by many traditional algorithms like multiple linear regression method, the parameter estimation of variogram model becomes quite difficult when there are drifts or noises in the raw data. The purpose of this paper is to improve the Kriging interpolation algorithm with learning kernels based on Estimation of Distribution Algorithms (EDAs) and Least-Squares Support Vector Machine (LSSVM). The experiments have been carried out based on a real-world case with environmental variables. Compared with other machine learning methods, experimental results verify the effectiveness of the proposed algorithm.

摘要

克里金插值法是一种广泛应用于数据分析和环境变量预测领域的空间插值方法，可提供最佳的中间值线性无偏预测。克里金插值的核心原理是寻找数据分布规律和预测区域化变量值，可以转化为对学习过程的两种描述：函数拟合问题和系数优化问题。虽然这两个问题可以通过多元线性回归法等许多传统算法来解决，但是当原始数据中存在漂移或噪声时，变差函数模型的参数估计变得相当困难。本文的目的是改进基于分布估计算法（EDA）和最小二乘支持向量机（LSSVM）的学习核的克里金插值算法。实验是基于具有环境变量的真实案例进行的。与其他机器学习方法相比，实验结果验证了所提算法的有效性。