• 关键词：稻田土壤  径向基神经网络模型  砷污染  多源数据  主成分分析

• 基金项目：国家重点研发项目（No.2017YFD0301603）

• 江叶枫
• 1. 江西农业大学, 江西省鄱阳湖流域农业资源与生态重点实验室, 南昌 330045;2. 江西农业大学, 国土资源与环境学院, 南昌 330045

• 郭熙
• 1. 江西农业大学, 江西省鄱阳湖流域农业资源与生态重点实验室, 南昌 330045;2. 江西农业大学, 国土资源与环境学院, 南昌 330045

• 摘要：合适的方法和多源的辅助数据对于准确预测土壤重金属的空间分布具有重要意义.该研究提出一种径向基函数神经网络结合普通克里格法的模型（RBFNN_OK），由主成分分析（PCA）提取的地形因子、遥感数据和邻近信息等多源辅助数据作为自变量，预测江西省都昌县稻田土壤砷空间分布.为验证RBFNN_OK的可行性：首先在全县范围内采集144个稻田表层（0~20 cm）土壤样品，运用ArcGIS地统计模块随机抽取115个（80%）采样点作为测试集，29个（20%）采样点作为验证集.其次多源辅助数据包括地形因子、遥感数据和邻近信息等14个定量因子作为预测变量，将预测变量进行主成分分析，得到前10个主成分的累积贡献率达到97.62%.再次一个特定的RBFNN_OK被用来预测土壤砷空间分布.最后将RBFNN_OK模型的预测结果与径向基神经网络模型（RBFNN）、回归克里格模型（RK）和多元逐步线性回归模型（MSLR）进行比较.结果表明，RBFNN_OK的测量值标准偏差与均方根误差的比值（RPD）较其它3种方法分别提高了14.92%、35.71%和44.67%.此外，RBFNN_OK还提供了更加真实且有关土壤砷空间分布的细节信息.RBFNN_OK取得最优效果可能归因于引入多源辅助数据，考虑多源辅助数据和土壤砷之间的多重共线性和非线性关系.该方法可为稻田土壤砷调查与环境保护提供更为精准的信息.

• Abstract：A suitable method and appropriate auxiliary data are important for accurately predicting heavy metal distribution in soils. Here we propose a radial basis function neural network combined with ordinary kriging (RBFNN_OK), multi-source auxiliary data such as topographic factors, remote sensing data, and neighboring information extracted by principal component analysis (PCA) as independent variables for predicting the spatial distribution of arsenic in paddy soils in Duchang County, Jiangxi Province, China. First, surface (0~20 cm) soil samples were collected from 144 sampling points in paddy fields across the study area; 115 (80%) of the sampling points were selected at random as the calibration set and 29 (20%) were selected as the validation set using ArcGIS Geostatistical Analyst. Next, multi-source auxiliary data includes 14 quantitative factors such as the topographic factors, remote sensing data, and neighboring information were selected as auxiliary variables, these variables were used for PCA and the cumulative contribution of the first 10 principal components to the total variance reached 97.62%. Then, a particular RBFNN_OK model was adapted to predict the spatial distribution of soil arsenic using the first 10 principal components. Finally, the predictions of RBFNN_OK were compared with those of radial basis function neural network (RBFNN), regression kriging (RK), and multiple stepwose linear regression (MSLR) for assessment of prediction accuracy. Results showed that the ratio of standard deviation of measured values to root mean square error of predictions was 2.85 for RBFNN_OK and 1.97~2.48 for the other three models, RBFNN_OK was increased by 14.92%, 35.71% and 44.67%, respectively, compared with the others. In addition, RBFNN_OK provides more detailed information about the spatial distribution of soil arsenic. The improved performance of RBFNN_OK can be attributed to the introduction of multi-source auxiliary data, to consider the multicollinearity and nonlinear relationship between multi-source auxiliary data and soil arsenic. This method may provide more accurate information for the investigation of arsenic in the paddy soil and environmental protection.

• Key words：paddy soils  radial basis function neural network  arsenic contamination  auxiliary data  principal component analysis

• 
  摘要点击次数： 941 全文下载次数： 802