基于多尺度地理环境变量的随机森林空间推测建模

Random Forest Spatial Prediction Modeling Based on Multi-Scale Geographic Environment Variables

摘要: 在数据科技快速发展的时代下，空间推测(Spatial Prediction)是指利用空间理论技术对空间对象进行建模、描述和表示，并以此为基础对空间对象间的空间关系进行定性或定量分析和处理的过程。机器学习方法作为空间智能领域的常用推测方法之一，它能够融合多维辅助变量的信息，提高模型的推测精度，但是不能有效结合空间位置关系信息来进一步提高推测精度。本文基于随机森林空间推测框架，将协变量的多尺度特征图层与随机森林空间推测算法结合，提出了基于多尺度地理环境变量的随机森林空间推测建模方法。应用Meuse数据集对所提出的方法进行对比实验，并与随机森林方法、随机森林的空间推测方法相对比，检验所提方法的推测精度。实验表明，本文提出的基于多尺度地理环境变量的随机森林空间推测建模方法相较于RF、RFsp-MS (k = 0)、RFsp-MS (k = 1)、RFsp-MS (k = 2)，可获得更加精准的空间推测结果，土壤制图结果也具有更加合理的空间分布和丰富的细节信息。因此，多尺度地理环境变量的随机森林空间推测方法能够有效结合自变量信息与空间位置关系信息，有效提高土壤变量推测的精度，同时能够输出有效的不确定性表达信息。

Abstract: In the era of rapid development of data science and technology, spatial prediction refers to the pro-cess of using spatial theory technology to model, describe, and represent spatial objects, and based on this, qualitatively or quantitatively analyze and process the spatial relationships between spatial objects. Machine learning method, as one of the commonly used inference methods in the field of spatial intelligence, can fuse information from multi-dimensional auxiliary variables to improve the inference accuracy of the model. However, it cannot effectively combine spatial position relationship information to further improve the inference accuracy. Based on the framework of random forest spatial inference, this article combines the multi-scale feature layer of covariates with the random forest spatial inference algorithm, and proposes a modeling method for random forest spatial in-ference based on multi-scale geographic environment variables. Meuse dataset was used to conduct comparative experiments on the proposed method and compare it with the random forest method and the spatial inference method of the random forest to test the inference accuracy of the proposed method. The experiment shows that the random forest spatial inference modeling method based on multi-scale geographic environment variables proposed in this article can obtain more accurate spatial inference results compared to RF, RFsp-MS (k = 0), RFsp-MS (k = 1), and RFsp-MS (k = 2), and soil mapping results also have more reasonable spatial distribution and rich detailed information. Therefore, the random forest spatial inference method for multi-scale geographic environment var-iables can effectively combine independent variable information and spatial position relationship information, effectively improve the accuracy of soil variable inference, and output effective uncer-tainty expression information.