基于Cubist多元混合回归的稻米富集Cd模型构建研究

为了给农产品质量安全预测提供技术支撑,基于土壤环境因子对稻米富集镉(Cd)的影响在不同条件下存在差异性这一原理,以湖南省典型区域为例,在土壤多参数检测获取基本大数据前提下,采用Cubist多元混合线性回归技术,构建了研究区土壤-稻米Cd传输模型。结果表明,在所筛选的23个土壤理化参数中,Ca、Fe、S、Cd、Cl和pH 6个参数对稻米富集Cd影响最为突出,并被纳入Cubist模型中。土壤Ca(以CaO计)与土壤p H在Ca含量小于1%时线性显著相关,S与土壤有机质(SOM)呈指数显著相关。土壤pH对稻米富集Cd的影响在不同值阈范围内差异明显。研究表明,混合线性回归技术克服了传统多元回归仅能构建单一性质模型的缺陷,可有效筛选及表征环境因子对稻米富集Cd的影响,是构建土壤-稻米Cd传输模型有效方法。     

Using pXRF and vis-NIR spectra for predicting properties of soils developed in loess

Visible near-infrared (vis-NIR) and portable X-ray fluorescence (pXRF) spectrometers have been increasingly utilized for predicting soil properties worldwide. However, only a few studies have focused on splitting the predictive models by horizons to evaluate prediction performance and systematically compare prediction performance for A, B, and combined A+B horizons. Therefore, we investigated the performance of pXRF and vis-NIR spectra, as individual or combined, for predicting the clay, silt, sand, total carbon (TC), and pH of soils developed in loess, and compared their prediction performance for A, B, and A+B horizons. Soil samples (176 in A horizon and 172 in B horizon) were taken from Mollisols and Alfisols in 136 pedons in Wisconsin, USA and analyzed for clay, silt, sand, pH, and TC. The pXRF and vis-NIR spectrometers were used to measure the pXRF and vis-NIR soil spectra. Data were separated into calibration (n=244, 70%) and validation (n=104, 30%) datasets. The Savitzky-Golay filter was applied to preprocess the pXRF and vis-NIR spectra, and the first 10 principal components (PCs) were selected through principal component analysis (PCA). Five types of predictor, i.e., PCs from vis-NIR spectra, pXRF of beams at 0-40 and 0-10 keV (XRF40 and XRF10, respectively) spectra, combined XRF40 and XRF10 (XRF40+XRF10) spectra, and combined XRF40, XRF10, and vis-NIR (XRF40+XRF10+vis-NIR) spectra, were compared for predicting soil properties using a machine learning algorithm (Cubist model). A multiple linear regression (MLR) model was applied to predict clay, silt, sand, pH, and TC using pXRF elements. The results suggested that pXRF spectra had better prediction performance for clay, silt, and sand, whereas vis-NIR spectra produced better TC and pH predictions. The best prediction performance for sand (R²=0.97), silt (R²=0.95), and clay (R²=0.84) was achieved using vis-NIR+XRF40+XRF10 spectra in B horizon, whereas the best prediction performance for TC (R²=0.93) and pH (R²=0.79) was achieved using vis-NIR+XRF40+XRF10 spectra in A+B horizon. For all soil properties, the best MLR model had a lower prediction accuracy than the Cubist model. It was concluded that pXRF and vis-NIR spectra can be successfully applied for predicting clay, silt, sand, pH, and TC with high accuracy for soils developed in loess, and that spectral models should be developed for different horizons to achieve high prediction accuracy.     

集成土壤-环境关系与机器学习的干旱区土壤属性数字制图

【目的】土壤属性的空间分布是影响农业生产力、土地管理和生态安全的重要因素。通过土壤环境耦合关系,在机器学习算法框架下,定量预测出干旱区土壤酸碱度（pH）、土壤盐分含量（Soil Salt Content,SSC)与土壤有机质（Soil Organic Matter, SOM）3种土壤属性的空间分布,为干旱区农业生产和生态安全提供科学依据。【方法】在渭干河—库车河绿洲干旱区于2017年7月设计采集典型表层（0—20 cm）土壤样品82个,依据土壤-环境之间的关系,集成DEM数据和Landsat 8数据提取出32种环境协变量,利用栅格重采样将提取出的32种变量重采样为90 m空间分辨率并转换为Grid格式参与建模。借助梯度提升决策树（Gradient Boosting Decision Tree,GBDT）模型依次对3类土壤属性的32种环境协变量进行重要性排序,并通过均方根误差（Root Mean Square Error,RMSE）界定出协变量重要性阈值点,从而筛选出参与3类土壤属性制图的环境协变量。进而运用随机森林（Random Forest, RF）、Bagging和Cubist 3种非线性模型建模,并引入多元线性回归模型（Multiple Linear Regression,MLR）进行对比分析,选出最优模型并绘制出90 m分辨率新疆渭干河-库车河绿洲干旱区pH、SSC与SOM 3种土壤属性图。【结果】梯度提升决策树能有效筛选出重要协变量,高程（Elevation）、剖面曲率（Profile Curvature）、差值植被指数（Difference Vegetation Index）、扩展增强型植被指数（Extended Normalized Difference Vegetation Index）、调整土壤亮度植被指数（Modified Soil Adjusted Vegetation Index）、盐分指数S1（Salinity Index S1）以及盐分指数S6 （Salinity Index S6） 7类环境变量均参与3类土壤属性建模,其中SSC遴选出参与建模协变量15种,pH和SOM则均为17种,且遥感指标在预测土壤属性图中起到强大的作用。机器学习3种算法的结果均优于MLR。通过3种非线性模型对比发现,随机森林在3种土壤属性中均表现最佳。在随机森林预测的3种土壤属性中,土壤pH验证集效果R ²=0.6779,RMSE =0.2182,ρ_c=0.6084;在SSC预测中,验证集R ²=0.7945,RMSE =3.1803,ρ_c=0.8377;在SOM预测中,验证集R ²=0.7472,RMSE =3.5456,ρ_c=0.7009。 【结论】GBDT所筛选出的重要性因子借助机器学习算法可以用于干旱区土壤属性制图,且随机森林模型均对3类土壤属性表现出最佳预测能力。依据所绘制的土壤属性图并结合土壤分类图厘清了3种制图属性的空间分布。