基于BP神经网络与Kriging结合的土壤有机质空间分布模拟——以福建省华安县为例

[目的]提高县域尺度耕地土壤有机质空间插值精度。[方法]基于福建省漳州市华安县215个土壤有机质野外采样数据,将样地土壤类型、土地利用方式两种定性因素转化为虚拟变量,结合土壤质地、海拔高度、坡度等定量因素,构建了BP神经网络与克里金插值(Kriging)相结合的非线性拟合法(BP_OK),并与回归克里金插值法(RK)、普通克里金插值法(OK)进行对比。[结果]利用30个验证样点计算BP_OK、RK、OK法的均方根误差分别为3.55、3.73、4.92 g·kg~(-1),相关系数分别为0.72、0.68、0.35。[结论]结合土壤有机质采样数据和外界辅助因素的BP_OK法、RK法插值精度明显优于仅考虑土壤有机质采样数据空间自相关性的OK法,其中采用非线性拟合的BP_OK法预测精度最高,证明了BP_OK法可以有效改善县域尺度下耕地土壤有机质空间分布模拟精度。     

罗江县农田土壤全氮协同克里格插值和采样数量优化研究

[目的]研究县域土壤全氮含量的空间分布和采样数量,为紫色土丘陵区采样提供参考。[方法]利用协同克里格法,以初始的1 777个土壤全氮含量数据为随机抽取的数据,分别随机抽取1 599、1 421和1 243个数据为目标变量,并以初始的1 777个土壤有机质数据为辅助变量,对四川省罗江县土壤全氮含量进行插值分析,从而利用协同克里格法对县域尺度下农田土壤全氮含量在不同样点数量下空间分布中的适用性进行评价。[结果]在相同取样数量下,全氮协同克里格法的均方根误差相对于普通克里格法降低0.019 6%～0.072 5%,预测值和实测值之间的相关系数提高0.69%～0.90%。利用协同克里格法,土壤全氮含量数据在缩减30%情况下,其估值精度高于1 777个样点下的普通克里格估值,且二者的分布图都具有较高的拟合度。[结论]协同克里格法是一种经济、精准的方法,可为县域土壤养分含量的空间分布提供基础信息。     

不同模型预测土壤有机质含量空间分布对比分析

【目的】探究土壤有机质含量空间分布预测模型的差异。【方法】在重庆市长寿区采集5162个土壤样点,结合地形、气候、植被和成土母质等9个环境变量,利用分类与回归树(CART)、随机森林(RF)、随机森林残差克里格(RFRK)和普通克里格(OK)4种预测模型对研究区土壤有机质含量空间分布进行预测制图,并利用平均绝对误差(MAE)、均方根误差(RMSE)、决定系数(R~2)和整体评价指标(GPI)评价模型精度。【结果】预测模型整体精度由低到高排序为CARTRFRFRKOK;在不同预测模型绘制的空间分布图中,土壤有机质含量空间分布的总体趋势一致,但在局部细节中存在差异。【结论】在采样密度较大区域,可以不借助辅助变量直接预测土壤有机质含量。     

基于随机森林的农耕区土壤有机质空间分布预测

以陕西省周至县农耕区为研究区,采集192个土壤样品,通过随机森林模型(random forest, RF)对土壤有机质含量进行回归预测,通过29个(15%)独立验证点对预测结果进行精度验证,并与普通克里格(ordinary kriging,OK)和协同克里格(cokriging,COK)插值结果进行对比分析。结果表明,研究区土壤有机质含量在训练集和验证集中均属于中等变异性,含量处于中等偏低水平,大致表现为中、南部黑河东岸土壤有机质含量相对较高,东北部渭河沿岸含量较低。对变量重要性进行排序,影响研究区土壤有机质的主要因素为数字高程(DEM)和降水量。与OK、COK相比, RF对土壤有机质的预测值和实测值的相关系数(0.782)更高,而平均绝对误差(0.618 g·kg^-1)和均方根误差(2.062 g·kg^-1)更低,说明RF能够更精确地反映局部土壤有机质含量的空间变化信息。     

用高程辅助提高土壤属性的空间预测精度

 【目的】探讨土壤属性变量与高程之间在何种条件下，可利用高程变量来辅助提高土壤变量的预测精度。【方法】用两种将高程作为辅助变量的克里格插值方法（协克里格法和简单克里格加变化局部平均值法）与没有考虑高程的普通克里格插值方法进行对比分析，用均方根预测误差和预测精度的相对提高值作为标准对3种方法的预测结果进行评价。【结果】对于交换性钾和pH值，协克里格法获得最精确的预测；对于Olsen-P、土壤有机质和有效锌，简单克里格加变化局部平均值法得到最精确的预测；而有效铜、有效铁和有效锰的最精确的预测结果则由普通克里格法产生。【结论】高程数据能够用来提高土壤特征的空间预测精度，但并不是对所有的土壤属性都适合；在利用高程数据来提高土壤属性空间预测之前，应该先对高程和土壤特征变量之间的线性相关关系、结构相关关系和全局趋势等进行仔细地分析，然后再选择适宜的方法。     

基于星载高光谱数据的南京新济洲湿地土壤有机质估测研究

在高光谱数据预处理、土壤有机质高光谱敏感波段提取基础上,建立多元线性回归、最邻近法、装袋算法、多元感知器、随机森林5种遥感估测模型。用10折交叉验证方法,借助相关系数、绝对误差、均方根误差、相对误差、相对均方根误差5个指标,对遥感估测模型结果进行精度评价,选择精度最高的模型进行湿地土壤有机质遥感估测和空间分析。结果表明:土壤有机质高光谱敏感波段主要集中在925、1 144、1 477、1 780 nm 4个波段;在预测土壤有机质的5种模型中,多元线性回归模型预测精度最高,随机森林次之;土壤有机质空间分布呈现由洲滩中间向四周逐渐增加的带状分布格局;新济洲沼泽地土壤有机质含量最高,为2.22%;靠近沼泽的林地次之;植被覆盖度较低的农地和裸地的土壤有机质最低,为0.43%;这种土壤有机质空间分布格局与研究区土壤类型的带状分布存在密切联系。     

毛竹林叶面积指数和郁闭度空间分布协同克里格估算

在地面调查的基础上,利用协同克里格插值法对研究区内毛竹Phyllostachys edulis林叶面积指数(LAI,leaf area index)和冠层郁闭度(CC,canopy closure)2个冠层参数进行空间分布估算研究,并与普通克里格插值法进行了比较。研究结果表明:①球状模型可以用来反映LAI和CC的空间变异,且两者具有强烈的空间自相关特征。②协同克里格插值得到的LAI预测值与实测值之间的决定系数R2为0.635 1,而CC的决定系数R2为0.428 5;与普通克里格法相比,基于协同克里格法的LAI和CC预测精度均得到改善,其中LAI预测精度提高了1.94%,均方根误差减少2.00%,平均标准误差减少0.18%,而CC预测精度提高了4.82%,均方根误差减少1.90%,平均标准误差减少1.30%。③安吉县毛竹林LAI和CC都具有从西南到东北逐渐递减空间分布格局,在一定程度上反映了安吉县不同区域毛竹林经营水平的差异。     

西南丘陵地区土壤有机质含量的空间插值法研究

通过对重庆市梁平县仁贤镇105个土壤采样点的有机质含量与海拔高度的关系进行分析, 提出了结合海拔高度对逆距离权重法和克里格插值法改进的插值方法, 并用改进前后的方法对研究区土壤有机质含量进行插值获取土壤有机质含量的空间分布图, 进行交叉检验. 结果表明, 结合海拔高度改进的插值方法插值结果精度有比较明显的提高. 对逆距离权重法, 考虑海拔高度的平均绝对误差(MAE)从0.081 9%降到了0.000 8%, 均方根误差(RMSE)从0.103 4%降到了0.001 4%;对克里格插值法, 考虑海拔高度的平均绝对误差(MAE)从0.076 4%降到了0.003 1%, 均方根误差(RMSE)从0.098 6%降到了0.003 9%. 在4种插值法中结合海拔高度的逆距离权重法最优.     

基于地形分类的黄土丘陵区土壤有机质插值方法

黄土丘陵山区地形复杂,耕地细碎,采样点密度不足,因此研究适宜该区域的空间插值方法具有重要意义.以山西省偏关县为研究区,基于数字高程模型(digital elevation model,简称DEM)数据生成地形因子,依据地形因子将研究区划分为不同地形单元,利用Kriging插值法对各地形单元进行分层插值,通过叠加不同地形单元的插值结果,得到全区域的土壤有机质含量空间分布,并利用均方根误差和决定系数评价空间插值方法的预测精度.结果表明,采用划分地形单元分层克里金(Kriging)插值法得到的预测值与实测值的决定系数为0.3753,明显高于全局插值法;均方根误差精度高于全局插值法,预测值无偏性好.因此,基于地形单元的空间插值方法可以更精确有效地获取土壤有机质空间分布特征,为复杂山区低密度采样下的耕地质量调查与评价提供技术参考.     

县域农田土壤铜含量的协同克里格插值及采样数量优化

 【目的】研究县域农田土壤铜含量的空间分布和采样数量,为农田土壤环境质量调查提供帮助。【方法】采用协同克里格方法,以初始的623个土壤铜含量数据及在此基础上随机抽取的560、498和432个数据为目标变量,并以初始的623个土壤有机质含量数据为辅助变量,对四川省双流县农田土壤铜含量进行插值分析,并对不同样点数量下协同克里格法在县域尺度农田土壤铜含量空间分布研究中的适用性进行评价。【结果】相同取样数量下,协同克里格法的均方根误差相对于普通克里格法可降低0.9%～7.77%,预测值和实测值之间的相关系数可提高1.76%至9.76%。利用协同克里格法,在土壤铜含量数据量缩减10%的情况下,其估值精度仍高于初始的623个土壤铜含量数据的普通克里格估值,且二者的分布图具有高度相似性。【结论】协同克里格作为一种更为精确和经济的方法,可为县域尺度农田土壤重金属含量的空间分布研究提供更多的信息和帮助。     

Improved Prediction and Reduction of Sampling Density for Soil Salinity by Different Geostatistical Methods

The spatial estimation for soil properties was improved and sampling intensities also decreased in terms of incorporated auxiliary data. In this study, kriging and two interpolation methods were proven well to estimate auxiliary variables： cokriging and regression-kriging, and using the salinity data from the first two stages as auxiliary variables, the methods both improved the interpolation of soil salinity in coastal saline land. The prediction accuracy of the three methods was observed under different sampling density of the target variable by comparison with another group of 80 validation sample points, from which the root-mean-square error （RMSE） and correlation coefficient （r） between the predicted and measured values were calculated. The results showed, with the help of auxiliary data, whatever the sample size of the target variable may be, cokriging and regression-kriging performed better than ordinary kriging. Moreover, regression-kriging produced on average more accurate predictions than cokriging. Compared with the kriging results, cokriging improved the estimations by reducing RMSE from 23.3 to 29% and increasing r from 16.6 to 25.5%, regression-kriging improved the estimations by reducing RMSE from 25 to 41.5% and increasing r from 16.8 to 27.2%. Therefore, regression-kriging shows promise for improved prediction for soil salinity and reduction of soil sampling intensity considerably while maintaining high prediction accuracy. Moreover, in regression-kriging, the regression model can have any form, such as generalized linear models, non-linear models or tree-based models, which provide a possibility to include more ancillary variables.     

基于辅助变量和回归径向基函数神经网络(R-RBFNN)的土壤有机质空间分布模拟

为快速准确地获取土壤有机质的空间分布状况,以江西省万年县齐埠镇为例,运用四方位搜索法、地统计学和遥感影像分析技术提取环境因子和邻近信息,构建基于环境因子和邻近信息的回归克里金法(RK)和回归径向基函数神经网络法(R-RBFNN),对齐埠镇耕地表层(0~20 cm)土壤有机质空间分布进行模拟,并与普通克里金法(OK)相比较。结果显示:齐埠镇耕地表层土壤有机质含量在17.30~53.58 g·kg^-1,平均值为35.03 g·kg^-1,变异系数为23.61%,呈中等变异性。半变异函数分析显示,土壤有机质的块金效应值为0.59,表现为中等空间相关性,自相关范围较大。利用62个采样点进行建模、16个采样点进行独立验证,误差分析表明,应用环境因子和邻近信息作为辅助变量的RK和R-RBFNN预测结果的均方根误差、平均绝对误差、平均相对误均差较OK降低,测试集中的相对提高度分别为66.67%和71.79%,显示出较高精度。但R-RBFNN无须计算半方差函数,使用简单,因此更具优势。     

Cokriging of Soil Cation Exchange Capacity Using the First Principal Component Derived from Soil Physico-Chemical Properties

As soil cation exchange capacity (CEC) is a vital indicator of soil quality and pollutant sequestration capacity,a study was conducted to evaluate cokriging of CEC with the principal components derived from soil physico-chemical properties.In Qingdao,China,107 soil samples were collected.Soil CEC was estimated by using 86 soil samples for prediction and 21 soil samples for test.The first two principal components (PC1 and PC2) together explained 60.2％ of the total variance of soil physico-chemical properties.The PC1 was highly correlated with CEC (r=0.76,P＜0.01),whereas there was no significant correlation between CEC and PC2 (r=0.03).The PC1 was then used as an auxiliary variable for the prediction of soil CEC.Mean error (ME) and root mean square error (RMSE) of kriging for the test dataset were -1.76 and 3.67 cmolc kg-1,and ME and RMSE of cokriging for the test dataset were -1.47 and 2.95 cmolc kg-1,respectively.The cross-validation R2 for the prediction dataset was 0.24 for kriging and 0.39 for cokriging.The results show that cokriging with PC1 is more reliable than kriging for spatial interpolation.In addition,principal components have the highest potential for cokriging predictions when the principal components have good correlations with the primary variables.     

Spatial Interpolation and Sample Size Optimization for Soil Copper （Cu） Investigation in Cropland Soil at County Scale Using Cokriging

Knowledge on spatial distribution and sampling size optimization of soil copper （Cu） could lay solid foundations for environmetal quality survey of agricultural soils at county scale. In this investigation, cokriging method was used to conduct the interpolation of Cu concentraiton in cropland soil in Shuangliu County, Sichuan Province, China. Based on the original 623 physicochmically measured soil samples, 560, 498, and 432 sub-samples were randomly selected as target variable and soil organic matter （SOM） of the whole original samples as auxiliary variable. Interpolation results using Cokriging under different sampling numbers were evaluated for their applicability in estimating the spatial distribution of soil Cu at county sacle. The results showed that the root mean square error （RMSE） produced by Cokriging decreased from 0.9 to 7.77%, correlation coefficient between the predicted values and the measured increased from 1.76 to 9.76% in comparison with the ordinary Kriging under the corresponding sample sizes. The prediction accuracy using Cokriging was still higher than original 623 data using ordinary Kriging even as sample size reduced 10%, and their interpolation maps were highly in agreement. Therefore, Cokriging was proven to be a more accurate and economic method which could provide more information and benefit for the studies on spatial distribution of soil pollutants at county scale.     

Combined use of hyperspectral VNIR reflectance spectroscopy and kriging to predict soil variables spatially

Hyperspectral visible near infrared reflectance spectroscopy (VNIRRS) and geostatistical methods are considered for precision soil mapping. This study evaluated whether VNIR or geostatistics, or their combined use, could provide efficient approaches for assessing the soil spatially and associated reductions in sample size using soil samples from a 32 ha area (800 × 400 m) in northern Turkey. Soil variables considered were CaCO₃, organic matter, clay, sand and silt contents, pH, electrical conductivity, cation exchange capacity (CEC) and exchangeable cations (Ca, Mg, Na and K). Cross-validation was used to compare the two approaches using all grid data (n = 512), systematic selections of 13, 25 and 50% of the data and random selections of 13 and 25% for calibration; the remaining data were used for validation. Partial least squares regression (PLSR) analysis was used for calibrating soil properties from first derivative VNIR reflectance spectra (VNIRRS), whereas ordinary-, co- and regression-kriging were used for spatial prediction. The VNIRRS-PLSR method provided better prediction results than ordinary kriging for soil organic matter, clay and sand contents, (R ² values of 0.56–0.73, 0.79–0.85, 0.65–0.79, respectively) and smaller root mean squared errors of prediction (values of 2.7–4.1, 37.4–43, 46.9–61, respectively). The EC, pH, Na, K and silt content were predicted poorly by both approaches because either the variables showed little variation or the data were not spatially correlated. Overall, the prediction accuracy of VNIRRS-PLSR was not affected by sample size as much as it was for ordinary kriging. Cokriging (COK) and regression kriging (RK) were applied to a combination of values predicted by VNIR reflectance spectroscopy and measured in the laboratory to improve the accuracy of prediction of the soil properties. The results showed that both COK and RK with VNIRRS estimates improved the predictions of soil variables compared to VNIRRS and OK. The combined use of VNIRRS and multivariate geostatistics results in better spatial prediction of soil properties and enables a reduction in sampling and laboratory analyses.     

基于RBF神经网络耕地土壤全氮插值方法的研究

[目的]通过耕地土壤全氮的空间变异特性研究,可以更好地调整耕地管理措施、合理施用氮肥、减少资源浪费。[方法]利用RBF(Radial Basis Function,径向基函数)神经网络插值法对区域土壤全氮进行空间插值,同时与普通克里格法进行比较。[结果]RBF神经网络插值法在拟合能力和插值能力方面要明显优于普通克里格法。[结论]RBF神经网络法具有很好的应用前景。     

Estimating soil organic matter using interpolation methods with a electromagnetic induction sensor and topographic parameters: a case study in a humid region

Soil organic matter (SOM) is a key indicator of soil quality although, usually, detailed data for a given area is difficult to obtain at low cost. This study was conducted to evaluate the usefulness of soil apparent electrical conductivity (EC_a), measured with an electromagnetic induction sensor, to improve the spatial estimation of SOM for site-specific soil management purposes. Apparent electrical conductivity was measured in a 10-ha prairie in NW Spain in November 2011. The EC_a measurements were used to design a sampling scheme of 80 locations, at which soil samples were collected from 0 to 20 cm depth and from 20 cm to the boundary of the A horizon (ranging from 25 to 48 cm). The SOM values determined at the two depths considered were weighted to obtain the results for the entire A Horizon. SOM distribution maps were obtained by inverse distance weighting and geostatistical techniques: ordinary kriging (OK), cokriging (COK), regression kriging either with linear models (LM-RK) or with random forest (RF-RK). SOM ranged from 46.3 to 78.0 g kg^?1, whereas EC_a varied from 6.7 to 14.7 mS m^?1. These two variables were significantly correlated (r = ?0.6, p < 0.05); hence, EC_a was used as an ancillary variable for interpolating SOM. A strong spatial dependence was found for both SOM and EC_a. The maps obtained exhibited a similar spatial pattern for SOM; COK maps did not show a significant improvement from OK predictions. However, RF-RK maps provided more accurate spatial estimates of SOM (error of predictions was between four and five times less than the other interpolators). This information is helpful for site-specific management purposes at this field.