应用基于PLSR的土壤-环境模型预测土壤属性

土壤-环境模型对于正确理解土壤属性与环境因子间的关系,以及进行土壤属性预测与制图均具有重要的意义。研究区位于陕西省长武县内多年退耕还林还草沟壑区域,采集72个土壤表层样本,选择3/4的样本作为建模集,其余1/4的样本作为验证集;环境因子选择容易获取的地形因子和由遥感影像提取的植被因子和湿度因子,建立基于偏最小二乘回归(PLSR)的土壤-环境模型。结果表明:全氮、速效钾、全钾、有机质与环境因子间均有显著相关性;建立的PLSR模型可解释土壤属性的空间变异从23%(全氮)到27%(全钾);与逐步回归方法构建的模型相比,利用PLSR构建的土壤-环境模型可以更好地表征土壤属性与环境变量间的关系,拟合精度和预测精度也相对较高,说明PLSR建立的模型可以更好地应用于相似区域的土壤属性预测。     

青海省表层土壤属性数字制图

对于土壤景观复杂的大区域,样点往往较为稀疏,如何准确地进行土壤预测制图仍是一个需要研究的问题。本文以青海省为研究区,基于近年采集205个土系调查点数据,采用随机森林模型,分别建立了表层(0～20 cm)土壤全氮、有机碳、粉粒含量和pH四个基本土壤属性与环境协同变量(海拔、坡度、地形湿度指数、年降水量、年平均气温、归一化植被指数、地表温度和地表反射率)之间的定量关系模型,对该地区进行了土壤多要素预测制图,分析了影响土壤空间变异的控制性因素。交叉验证结果显示,全氮、有机碳、粉粒含量和pH的R~2分别是0.61、0.53、0.47和0.54,这说明随机森林模型可解释47%以上的土壤空间变异。表层土壤全氮和有机碳空间分布趋势东南高,西北低,pH呈现出相反的空间模式;粉粒含量东高西低,预测结果高值出现在柴达木盆地和南部玉树、果洛地区。环境变量的重要性分析表明,年降水量对表层土壤全氮、有机碳、pH空间分布模式具有控制性影响,夜间地表温度与表层土壤粉粒含量空间变异具有较强的协同关系。     

基于环境相关法和地统计学的土壤属性空间分布预测

土壤属性是土壤质量的重要决定因素,并强烈影响土地利用和生态过程。正确理解并充分考虑土壤空间变异,对于在景观尺度上建立生态、环境过程模型是必不可少的。在黄土高原横山县采集了254个样点,应用数字地形与遥感影像分析技术,获取相关地形因子与遥感指数,分析土壤属性（土壤容重、有机质和全磷）与环境因子相互关系,并利用环境变量进行空间预测。结果表明,土壤容重、有机质与地形因子和遥感指数之间存在较好相关性,而全磷与地形因子相关性不大;多元线性逐步回归模型对于土壤容重和有机质拟合较好,而对于全磷,预测结果较差;回归－克里格预测有效地减小了残差,消除了平滑效应,与实测值较为接近。     

流域尺度土壤厚度的模糊聚类与预测制图研究

基于土壤厚度与景观位置和特征之间的关系,运用模糊c均值聚类(FCM)方法对西苕溪流域的土壤厚度分布进行了空间预测。选取高程、坡度、平面曲率、剖面曲率、径流强度系数和地形湿度指数6个地形因子进行模糊聚类,根据相应的聚类参数将流域地形组合分为8类。利用部分调查获得的土壤剖面数据,结合样点属性和专家经验为典型区赋值,最后由加权平均得到流域土壤厚度预测图。验证结果表明,FCM方法可以对地形因子组合进行有效合理的分级,其预测精度较高,模型的稳定性较好,是一种低成本高效率的制图方法。该方法在土壤厚度预测方面具有一定的可靠性。     

DEM栅格分辨率对多元线性土壤—景观模型及其制图应用的影响

以亚热带丘陵地区为对象,以该区4 km×3 km的5 m、10 m、15 m、20 m、25 m、30 m数字高程模型(DEM)为基础,建立多元线性土壤景观模型,并应用该模型预测研究区内土壤表层有机质含量分布,进而比较不同分辨率DEM中土壤景观模型及其预测制图的精度。结果表明:在本研究区11 km2范围内,随着DEM栅格分辨率降低,坡度、曲率、比汇水面积(对数)频度均表现出了向其中值区集中的趋势;地形因子的这一变化规律对土壤景观模型的影响较小,例如模型的变量、变量系数及R2在不同分辨率DEM中的差异很小;但地形因子的这一变化规律对模型预测制图的精度具有较大影响,各项指标均说明,模型在10~25 m DEM中的制图精度较高,而在更高分辨率(5 m)或更低分辨率(30 m)DEM中较低。本研究结果对其他亚热带丘陵地区具有一定的指示意义。     

基于模拟退火算法的土壤样点布局和土壤景观模型优化

本研究利用多重线性回归方程,以地形因子为预测变量,构建关于土壤有机质的土壤景观模型,并以西南山地丘陵区的一块面积为2 km2的汇水盆地为研究区,对该区域的土壤有机质空间分布进行预测。在此基础之上,探讨最少可用多少个点来预测土壤有机质的空间分布,并使之预测精度不低于原始集合的精度;同时,找出最优土壤样点布局,确定不同地形部位的取样单元,使之预测精度最高。研究结果表明:在预测误差最小化的情况下,最少可用7个优化的样点就可以代替原始200个采样点,且优化的样点数为124时,模型预测土壤有机质空间分布的精度最高。优化后的土壤景观模型的拟合度比原始模型提高了3.28%,MAE降低了5.3%,RMSE降低了3.94%。     

紫色土丘陵地区农田土壤养分空间分布预测

为深入研究紫色土丘陵区农田土壤养分空间分布规律,在GIS技术的支持下,利用研究区450个土壤实测数据,结合地形因子和土地利用类型,运用多重线性回归构建了土壤养分预测模型,对养分的空间分布进行预测。结果表明,土壤有机质和碱解氮含量与地形因子之间的相关性较强,有效磷和速效钾含量与地形因子之间的相关性较弱。土壤水田和旱地中有机质、碱解氮和有效磷含量均值间的差异显著(P<0.01),速效钾之间不显著(P=0.34)。基于地形因子的土壤养分预测模型与基于地形因子和土地利用方式组合的土壤养分预测模型预测结果精度对比表明,在预测变量中增加土地利用类型对提高预测模型的拟合度和预测精度作用非常微小,且仅用地形因子预测土壤养分的空间分布更方便,因此选用该模型对验证集数据进行预测。以验证集数据进行预测结果与实测值进行比较,结果显示预测值与实测值之间的差异甚小,有机质、碱解氮、有效磷和速效钾的相对偏差分别为0.09、0.19、0.08和0.12,均方根误差分别为1.38、3.42、1.03和1.57,说明基于地形因子的土壤养分预测模型的精度较高,可以很好地预测土壤养分分布规律。该研究结果可为丘陵地区农田合理施肥提供理论依据。     

基于模糊集理论提取土壤—地形定量关系及制图应用

通过对研究区地形因子的模糊聚类,提取了地形因子组合与特定土壤属性的定量隶属度关系,然后对隶属度高值区土壤进行目的性采样为隶属度函数赋值,制作研究区土层厚度连续分布图。通过野外实地验证,将观测值与图中预测值比较,结果显示该方法制图精度在82%左右,具有一定的可靠性。进一步考察认为该模型在地形部位较低,地势较为平坦,土壤发育较好,土层较厚,成土环境相对稳定的地区预测效果更好,适用性更强。该方法能提高土壤制图效率,降低制图成本,提高制图精度,对土壤微域变异的表现更为详细,图面信息负载量更高。应用该方法制作大比例尺土壤详图不失为土壤调查与制图领域一种可行的新方法。     

基于人工神经网络的土壤颗粒组成制图

以浙江西苕溪流域为研究区,综合考虑地形和土壤类型等信息,采集典型土壤样本,测定土壤颗粒组成,并基于土壤颗粒组成与景观位置和特征之间的关系,利用径向基函数(RBF)神经网络建立了高程、坡度、平面曲率、剖面曲率、径流强度系数和地形湿度指数6个地形因子与土壤颗粒组成之间的非线性映射关系,预测土壤颗粒组成的空间分布。验证结果表明,RBF神经网络方法能够挖掘出地形因子信息与土壤颗粒组成之间的非线性映射关系,其预测精度较高,模型稳定性较好,是一种低成本、高效率的制图方法。     

运用模糊隶属度进行土壤属性制图的研究——以黑龙江鹤山农场研究区为例

通过传统土壤类型图所得的土壤属性图已不能满足精准农业和生态环境模型所需土壤属性的精度,而目前应用较多的统计方法和地统计方法均存在一定的局限性。鉴于此,本文探索了一种采用模糊聚类获取模糊隶属度进行土壤属性制图的方法。首先,采用模糊c均值聚类(Fuzzyc-means clustering,FCM)方法对环境因子进行聚类,通过野外采样(称为建模点)建立土壤-环境关系知识;然后,计算区域内各像元点对土壤类型的模糊隶属度;最后,对模糊隶属度采用加权平均的方法获取土壤属性值。将该方法应用于黑龙江鹤山农场老莱河流域的研究小区,以土体厚度和表层有机质为例进行土壤属性制图。为了评价该方法的有效性,将其与采用环境因子所建立的多元线性回归模型进行比较,通过野外验证点集评价两种模型所得的土壤属性,评价指标为观测值和预测值的相关系数、平均绝对误差(MAE)、均方根误差(RMSE)和准确度(AC)。结果表明,尽管通过建模点建立的多元线性回归方程R2较大,但该方程并不适用于研究区内的其他样本点,这表明多元线性回归方法在该区具有一定的局限性。与之相比,模糊隶属度加权平均的方法则可以通过较少的建模点得到更好的预测效果。     

预测性土壤有机质制图中模糊聚类参数的优选

对数字高程模型(DEM)中的地形特征进行模糊C均值聚类,分别采用3种方法来选择最优模糊度和分类数组合及分类结果;在最优分类结果上,用多元线性回归方法建立土壤A层有机质含量与地形景观之间的定量关系,并应用该关系进行土壤制图应用.结果表明:3种方法选择的最优模糊度比较接近,主要为1.5,还包括1.4和1.6,但3种方法选择的最优分类数却有很大差别;尽管依据回归模型γ2选择的分类结果较多地解释了土壤A层有机质含量的变异,但基于这种分类结果的制图偏差较大,与实测值相比较的结果也说明基于这种分类结果的制图精度较低;用内部判据选择的分类结果在制图过程中产生的偏差较小,制图精度也较高.     

Match method and its application for the development of a large-scale soil map

A method for interpolating field soil data to obtain the maps of soil taxa is suggested. It is based on representation of categorical data in the form of Voronoi map with barriers limiting the areas with particular combinations of indicative soil-landscape relationships. The predictive capacity of the proposed model depends on the level of the taxon and exceeds 80% for soil types and subtypes in the study area. At different levels of detail of the indicative soil-landscape relationships, the accuracies of prediction are different. The proposed method significantly reduces the time of soil mapping and opens new possibilities for investigating the soil-landscape relationships.     

How good is a reconnaissance soil map for agronomic purposes?

Abstract. Information about the soil fertility status in irrigated ricelands at regional scales (1:50 000–1:250 000) is commonly not contained in classical soil maps. To assess the agronomic suitability of two different reconnaissance soil maps, we conducted a detailed soil survey in the Nueva Ecija province, Philippines. Soil samples were collected from 384 farmers' fields, and soil properties were measured for topsoil and subsoil samples. For most soil properties, a soil map made in 1940 (1:125 000) had within-map unit variances that were smaller than the total variance, whereas a new soil map of 1992 (1:50 000) did not significantly reduce the within-class variance. In both soil maps, classification into mapping units accounted for 0–40% of the variance of 14 agronomically important soil properties and large within-map unit variabilities were found. Underlying strategies of classical soil survey supported the partition of variance for relatively stable soil properties, such as soil texture, CEC, and organic matter. If reconnaissance soil maps are used in quantitative land evaluation studies, existing maps require upgrading by adding quantitative information about relevant soil properties and their within-map unit variability The sampling demand for upgrading a reconnaissance soil map was large, but pedotransfer functions can be used as cost-saving tools. Measures of soil nutrient status were highly variable within all mapping units and differences among farmers were much greater than the differences between soil types. Therefore, nutrient management in the study region should be based on individual field or farm recommendations rather than on soil-map based recommendations.     

宁镇丘陵区村域小流域不同土壤景观下表土质量变化及评价

中国南方丘陵区存在着自然过程和人类活动交互影响的、土地利用—土地覆盖叠加变化的多种土壤景观,其土壤质量对乡村农业产业发展具有重要影响。在南京市远郊的溧水区晶桥镇芝山村域小流域,选取流域内生态保护的林地、农业利用的园地、旱地和稻田4种土壤景观,于秋季分别采集表土样品,测定土壤基础理化性质、土壤团聚体粒径组成及土壤微生物磷脂肪酸和胞外酶活性分布,分析土壤肥力、土壤团聚化、土壤生物活性等关键性质在不同土壤景观中的变异情况,并采用土壤质量评价方法探明土壤景观与表土质量的变化关系。结果表明,与原生林地相比,农业土壤景观中有机质减少50.93%~69.63%,土壤团聚体平均重量直径降低41.34%~68.71%;相应地,土壤微生物总磷脂脂肪酸含量也降低19.20%~42.04%,土壤归一化酶活性降低22.48%~63.27%。因此,与林地景观土壤相比,农业土壤景观的土壤生态系统服务功能已经显著削弱。不过,在农业土壤景观中,稻田的土壤有机质储量和微生物活性相对较高。回归分析表明,土壤有机质含量是影响土壤性质变化的最强因子。基于总数据集和最小数据集的土壤质量评价和基于土壤健康理念的土壤功能质量评价均表明,表土总体土壤质量的变化趋势为林地>稻田>旱地>园地。同时,基于土壤健康理念的评价体系能综合地反映不同土壤景观间土壤的质量变化及其生态系统功能意义。     

基于RBF神经网络的土壤有机质空间变异研究方法

通过研究土壤性质的空间变异和空间插值方法,快速准确获取土壤性质的空间分布是精确农业和环境保护的基础。该文以四川眉山一块约40 km2的区域为试验区,采集表层土壤（0～20 cm）样点80个,利用径向基函数(RBF)神经网络建立空间坐标和邻近样点与土壤有机质间的非线性映射关系（RBF2）,模拟土壤有机质的空间分布。与普通克里法（OK）和仅以坐标为网络输入的神经网络方法（RBF1）相比,RBF2的插值精度有显著的提高;相同样点密度下其相对预测误差分别较OK和RBF1减小了9.87%、1.97%（样本A）和13.09%、2.36%（样本B）;即使样点数减半的情况下RBF2的相对预测误差也分别较OK和RBF1减小了10.23%和2.33%,并且插值图差异相对较小,可以更好地反映土壤有机质空间分布的异质性。因此,利用以坐标和邻近样点为输入的神经网络方法可以相对准确、快速地获取区域土壤性质空间分布的异质性信息。     

The application of expert knowledge in Bayesian networks to predict soil bulk density at the landscape scale

This paper investigates the use of expert knowledge as a resource for digital soil mapping. To do this, three models of topsoil soil bulk density (D_b) were produced: (i) a random forest model formulated and cross‐validated with the limited data available (which served as the benchmark), (ii) a naïve Bayesian network (BN) where the conditional probabilities that define the relations between D_b and explanatory landscape variables were derived from expert knowledge rather than data and (iii) a ‘hierarchical’ BN where model structure was also defined by expert knowledge. These models were used to generate spatial predictions for mapping topsoil D_b at a landscape scale. The results show that expert knowledge‐based models can identify the same spatial trends in soil properties at a landscape scale as state‐of‐the‐art mapping algorithms. This means that they are a viable option for soil mapping applications in areas that have limited empirical data.     

轮作模式在农耕区土壤有机质推测制图中的应用

人类活动近年成为数字土壤制图亟需考虑的要素。本文以农业活动中轮作模式为例,将轮作信息应用于数字土壤制图,探讨其对土壤空间变异刻画的有效性。以安徽宣城两个县市的耕地平区为研究区,通过野外调查获得近年三种主要轮作模式,基于监督分类对多期遥感影像解译得到轮作类型空间分布图,使用方差分析探讨轮作对土壤表层有机质空间变异是否有显著性影响,采用随机森林重要性指标对自然环境因子、轮作模式、土地利用方式和归一化植被指数进行重要性排序,并构建不同的环境因子组合,利用基于相似度的土壤推测模型和随机森林模型进行制图和交叉验证。结果表明,轮作模式对土壤表层有机质有显著性影响,其重要性排序为第二,引入轮作使得基于相似度的土壤推测模型和随机森林模型制图精度分别提高4.8%~65.9%和1.9%~2.7%。     

Explaining and mapping total phosphorus content in French topsoils

Total phosphorus (TP) build‐up in agricultural soils represents both a threat to aquatic ecosystems and a valuable resource for future crop production, given the context of increasing food demand combined with the rapid depletion of the world's phosphate reserves. Therefore, it is crucially important (i) to identify the main factors controlling topsoil TP and (ii) to develop methods for mapping its spatial distribution. Multiple linear regression models were used with two distinct approaches to calculate TP and covariates linked to the P cycle. Firstly, covariates were selected from the Réseau de Mesures de la Qualité des Sols database, the French soil monitoring network, which consists of soil samples collected from 2158 sites on a 16‐km regular grid. Secondly, covariates were selected to map TP from spatially exhaustive datasets in France. The first approach explains 80% of variability in topsoil TP. The variables selected are linked to the autochthonous origin of P (parent material), to allochthonous origin (organic carbon and nitrogen contents) and to the retention capacity of soil (Al, Fe, Ca and clay contents). The predicted map obtained from the second approach provides a mean TP of 0.76 g/kg. This study demonstrates that creating national scale maps of TP, based on detailed soil sampling and many variables, is feasible and can be used to model the P cycle and P transfer processes. Such maps can be used in P erosion and transfer models over river basins, and therefore to predict P exports to surface waters.     

Extrapolating regional soil landscapes from an existing soil map: Sampling intensity,validation procedures,and integration of spatial context

This paper aims to investigate the potential of using soil-landscape pattern extracted from a soil map to predict soil distribution at unvisited location. Recent machine learning advances used in previous studies showed that the knowledge embedded within soil units delineated by experts can be retrieved and explicitly formulated from environmental data layers However, the extent to which the models can yield valid prediction has been little studied. Our approach is based on a classification tree analysis which has underwent a recent statistics advance, namely, stochastic gradient boosting. We used an existing soil-landscape map to test our methodology. Explanatory variables included classical terrain factors (elevation, slope, curvature plan and profile, wetness index, etc.), various channels and combinations of channels from LANDSAT ETM imagery, land cover and lithology maps. Overall classification accuracy indexes were calculated under two validation schemes, either taken within the training area or from a separated validation area. We focused our study on the accuracy assessment and testing of two modelling parameters: sampling intensity and spatial context integration. First, we observed strong differences in accuracy between the training area and the extrapolated area. Second, sampling intensity, in proportion to the class extent, did not largely influence the classification accuracy. Spatial context integration by the use of a mean filtering algorithm on explanatory variables increased the Kappa index on the extrapolated area by more than ten points. The best accuracy measurements were obtained for a combination of the raw explanatory dataset with the filtered dataset representing regional trend. However, the predictive capacity of models remained quite low when extrapolated to an independent validation area. Nevertheless, this study offers encouragement for the success of extrapolating soil patterns from existing soil maps to fill the gaps in present soil map coverage and to increase efficiency of ongoing soil survey.     

基于多变量与RF算法的耕地土壤有机碳空间预测研究——以福建亚热带复杂地貌区为例

耕地土壤有机碳(Soil Organic Carbon,SOC)含量不仅是土壤质量的重要表征,还是农业温室气体的重要源库,而基于环境变量建立的随机森林算法(Random Forest,RF)是当前提高土壤有机碳空间预测精度的方法,但不同组合环境变量对RF模型预测精度的影响仍需深入研究。本文以福建闽东南复杂地貌区为例,以两种环境变量组合(遥感变量+气候因子和遥感变量+气候因子+土壤属性)为输入数据集,利用RF算法对耕地表层SOC含量进行模拟预测和精度对比,并与普通克里格(OrdinaryKriging,OK)插值模型进行比较。结果表明,基于全部环境变量构建的RF模型表现最佳,其模型拟合度和预测精度相较于未加入土壤属性的模型有显著提高(r提高7.95%,为0.95,RMSE下降45.13%),且对SOC空间分异信息的捕获更精确,OK模型总体预测精度最弱。利用最优模型反演得到的研究区耕地SOC含量为14.70±2.95 g·kg~(–1),东部沿海低于西部内陆。变量贡献率分析显示,除了与土壤碳紧密相关的水解性氮(N),遥感变量中数字高程模型(Digital Elevation Modecs,DEM)也是影响闽东南地区SOC预测精度的重要变量,因此,遥感变量、气候因子和土壤属性共同驱动的随机森林模型可作为闽东南复杂地貌区耕地有机碳含量空间预测的有效方法。