摘要
准确预测土壤重金属含量的空间分布对监测耕地污染和确保生态农业可持续发展至关重要.然而,能否在基于环境变量预测土壤重金属含量分布的基础上,通过考虑表征元素空间变异规律的协变量,从而提升预测结果的准确性和可靠性,尚需开展进一步的研究.鉴于此,本文以杭州市富阳区耕地土壤重金属As元素含量预测为例,借助土壤重金属观测样本构建反映重金属空间异质性特征的协变量(Spatial Regionalization variables,SRs),并结合常规环境变量和反映空间自相关性特征的协变量,建立偏最小二乘(Partial Least Squares Regression,PLSR)、随机森林(Random Forest,RF)和深度森林(Deep Forest 21,DF21)3种回归模型.通过交叉验证对预测模型进行评估和优选,最后获得研究区耕地土壤As元素含量的空间分布,并对其分布规律和影响因子进行分析.结果表明:①加入基于3种插值方法(普通克里金、趋势面插值、反距离加权插值)反映重金属空间异质性的SRs后,PLSR模型、RF模型和DF21模型精度均有所提升(PLSR:0.208→0.488;RF:0.367→0.522;DF21:0.381→0.538),其中,DF21模型预测精度最高;②通过使用不同数量样本点构建SRs并对比模型预测结果,可以发现,随着用于构建SR变量的样本数量的减少,3种模型的精度均呈下降趋势;然而,预测效果仍优于未考虑空间结构变量的模型;③变量重要性分析结果显示,As含量空间分异主要受空间区域化变量SRs、到河流距离和地形地貌等因素的共同影响,其中,基于反距离加权插值构建的SR变量为最关键影响因素.综上,考虑空间异质性的空间预测方法,有望为区域土壤污染的调查、评价和管控提供有效的方法支撑.
Accurately predicting the spatial distribution of soil heavy metal content is essential for monitoring pollution in cultivated land and ensuring the sustainable development of eco-agriculture.However,further research is needed to improve the accuracy and reliability of the predictions by incorporating variables that characterize the spatial variability of the elements,in addition to the environmental variables used to predict the distribution of soil heavy metal content.Therefore,we take the prediction of soil heavy metal As content in cultivated land of Fuyang District,Hangzhou City as an example.Utilizing soil heavy metal observation samples,we construct variables(Spatial Regionalization variables,SRs)that reflect the spatial heterogeneity characteristics of heavy metals.These are then combined with conventional environmental variables and covariates reflecting spatial autocorrelation characteristics to establish three regression models:Partial Least Squares Regression(PLSR),Random Forest(RF),and Deep Forest 21(DF21).The predictive models are evaluated and optimized through cross-validation,ultimately obtaining the spatial distribution of soil As content in the study area and analyzing its distribution patterns and influencing factors.The results indicate that:①Incorporating SRs constructed using three interpolation methods(ordinary kriging,trend surface interpolation,inverse distance weighting interpolation)to reflect the spatial heterogeneity of heavy metal content improves the accuracy of the PLSR,RF,and DF21 models(PLSR:0.208→0.488;RF:0.367→0.522;DF21:0.381→0.538),with the DF21 model demonstrating the highest prediction accuracy;②By using different numbers of sample points to construct the SRs and comparing the model prediction results,it was found that as the number of samples used to construct SR variables decreases,the accuracy of the three models decreases as well.Nevertheless,the prediction performance remains superior to models that do not consider spatial structural variables;③Variable importance analysis results indicate that the spatial differentiation of As content is predominantly influenced by the SR variables,distance to rivers,and topography,with the SR variable constructed based on inverse distance weighting interpolation being the most critical influencing factor.In conclusion,the spatial prediction methods considering spatial heterogeneity are expected to provide effective methodological support for the investigation,evaluation and management of regional soil pollution.
作者
赵若言
王健
吴雄辉
刘馨钰
ZHAO Ruoyan;WANG Jian;WU Xionghui;LIU Xinyu(College of Geography and Planning,Chengdu University of Technology,Chengdu 610059)
出处
《环境科学学报》
CAS
CSCD
北大核心
2024年第11期288-300,共13页
Acta Scientiae Circumstantiae
基金
国家自然科学基金(No.42002295)。
关键词
土壤重金属
机器学习
空间分布预测
空间异质性
深度森林
soil heavy meal
machine learning
spatial distribution prediction
spatial heterogeneity
Deep Forest 21
作者简介
赵若言(1999-),女,E-mail:zhaoruoyan@stu.cdut.edu.cn;王健,E-mail:jwang@cdut.edu.cn。
