基于随机森林的土壤重金属浓度预测

Prediction of Soil Heavy Metal Concentration—Using a Random Forest Machine Learning Algorithm

摘要: 土壤中重金属的潜在危害越来越受到人们的重视。预测连续的重金属浓度面有助于确定环境管理和修复的热点问题区域。近年来，基于森林的算法在解决包括地球科学在内的各个领域的各种问题中得到了广泛的应用。随机森林回归是一种有效的预测方法。本文利用土壤样本数据和其他协变量数据，将随机森林回归算法应用于预测土壤重金属浓度栅格图。本研究中的“污水排放点”、“干洗店”、“交通数据”(命名为AADT)和EPA TRI (有毒物质释放清单)与土壤重金属如何传播的主要途径有关。它还包括佛罗里达州“主要/次要排放”和“副排放”的数据。本文采用随机森林回归方法预测了铅、锌的连续重金属浓度。将点数据分成不同的种类进行密度面计算。将这些密度数据与土壤性质数据和采样数据结合起来，创建两个数据集：预测点和训练点。两者都在研究范围内，具有相同的解释变量集。然后利用随机森林回归模型得到预测结果。最后，使用RMSE评估预测精度。结果表明，随机森林回归模型中决策树数量的增加，RMSE值降低。

Abstract: The potential hazard of heavy metals in soil has been attracted more and more attention. Prediction of a continuous heavy metal concentration surface helps identify hotspot problem areas for environmental management and remediation. In recent years, forest-based algorithm has become popular in solving many kinds of problems in different fields, including geoscience. Random Forest Regression is one of them works well in prediction. The paper represents an application of Random Forest Regression algorithm in predict concentration surface using soil sample data collected the research project by Liebens et al. (2012) and other covariates data. Such as “sewer waste points”, “dry cleaner”, “Traffic data (Named AADT) and EPA TRI (toxic release inventory) in this study relate with the main way about how soil heavy metal spread. It also includes Florida DEP “major/minor emitters” and “DEP emitters” data. In this paper, we use random forest regression to predict continuous heavy metal concentration focus on Pb and Zn. The point’s data were divided into different kinds to calculate the density surfaces. Joined these density data with soil properties data and sampling data to create two datasets: prediction points and training points. Both are in the study area and have the same set of explanatory variables. Then run a Random Forest Regression model to get the prediction results. Finally, assess prediction accuracy with RMSE. The results show that RMSE value decreases when the number of decision trees increases in the random forest regression model.