基于BP神经网络的土壤水力学参数预测

为了获取区域土壤水分和溶质运移模拟所需的土壤水力学参数,利用黄淮海平原曲周县的试验资料建立基于BP神经网络的土壤转换函数模型。本文采用土壤粒径分布、容重、有机质含量等土壤基本理化性质,来预测土壤饱和导水率Ks、饱和含水量sθ、残余含水量θr、以及van Genuchten公式参数α、n的对数形式ln(α)和ln(n),并与多元线性逐步回归方法进行比较。t检验结果表明,BP神经网络训练和预测得到的模拟值与实测值之间吻合很好,该方法具有较高的预测精度。通过对平均相对误差的比较,得出在粒径分布的基础上增加容重、有机质含量等输入项目,可以提高部分土壤水力学参数的预测精度,而有些参数的预测精度反而降低。以误差平方和为标准的比较结果表明,BP神经网络模型的预测效果总的来看要优于多元线性回归法。     

基于BP人工神经网络的长江河口地区土壤盐分动态模拟及预测

为开展长江河口地区土壤盐分动态的中长期模拟与预测,采用人工神经网络中应用较为成熟和广泛的BP网络建立长江河口地区土壤盐分与降雨量、蒸发量、长江水电导率、内河水电导率、地下水位、地下水电导率6因子间的非线性神经网络响应模型。网络模型结构为6-11-1,隐含层单元数用"试错法"确定。选择合适的参数训练和学习网络模型后,对河口地区2003年各月平均根层土壤电导率进行预测,并与线性回归模型预测结果进行比较。结果表明:BP网络模型较线性回归模型具有更高的预测精度,平均相对预测误差为7.3%,预测值与实测值相关性良好,可以满足实际应用需求。     

饲草料地土壤水分动态变化规律及其预测的人工神经网络模型的研究

根据2005～2006年实测土壤水分资料和气象资料,研究分析了饲草料地土壤水分的动态变化规律和建立了考虑多个因素对土壤水分影响的BP人工神经网络模型,结果表明:表层土壤的含水率变幅较大,主要是受大气降水的影响,20～40cm和40～60cm土层土壤含水率的波动,除大气降水的影响外,还与植物的生长发育状况有关;土壤水分预测模型具有较好的预测效果,用神经网络建立土壤水分预测模型的方法是可行的,对于不同条件的地区具有广泛的适应性和推广应用前景。     

艾比湖流域不同土地覆盖类型土壤养分 高光谱反演模型研究

土壤养分影响着土壤的质量,也影响着植被、农作物等的生长。为快速准确地估测艾比湖流域土壤养分状况,选择艾比湖流域精河县作为研究区,以精河县内不同地表覆盖类型土壤为研究对象,基于实地采集的75个土壤样品的室内ASD Field Spec3实测光谱数据和3种光谱变换形式,利用10 nm间隔重采样进行去噪处理,再结合多元逐步回归法(SMLR)、偏最小二乘法回归法(PLSR)、人工神经网络法(ANN)分别建立土壤养分预测模型,以探索最优模型。结果表明:土壤实测光谱的一阶微分、二阶微分变换形式能显著提高光谱与土壤养分之间的相关性,尤其是一阶微分变换与土壤有机质和全氮的相关性最高分别达0.87和0.91,光谱变换技术能显著增强土壤养分与高光谱之间的敏感度,达到更好的建模效果;SMLR、PLSR和ANN这3种模型都具有良好的预测能力,其中,ANN建立的模型预测效果最好,二阶微分变换的ANN模型对有机质、全氮的预测决定系数(R2)分别为0.886和0.984,均方根误差(RMSE)分别为2.614和0.147,PLSR次之;全氮的预测效果明显优于有机质的预测效果,说明高光谱和全氮之间的敏感性更高。总体来说,光谱二阶微分变换形式的人工神经网络模型可以最精确稳定地完成土壤养分含量的快速预测,能够实现艾比湖流域的土壤养分空间分布状况和动态变化特征的动态监测。     

青海三江源地区土壤水分常数转换函数的建立与比较

利用土壤理化性质数据建立转换函数是间接获得土壤水力参数的重要手段之一。基于测定的土壤理化性质和土壤水分常数数据,本文采用回归分析、BP神经网络和基于BP神经网络的Rosetta模型3种方式分别建立了青海三江源地区土壤饱和含水量、毛管持水量和田间持水量的转换函数,并对其预测精度进行了比较。结果表明:(1)回归分析方法总体预测效果比较理想,特别是田间持水量的平均误差(ME)和均方根误差(RMSE)都在3.397%以下,决定系数(R2)高达0.868;(2)BP神经网络方法的预测效果非常理想,各土壤水分常数平均误差和均方根误差都在4.685%以下,并且决定系数均在0.857以上;(3)Rosetta模型的预测效果相对较差,特别是饱和含水量和毛管持水量,平均误差(ME)和均方根误差(RMSE)相对较大,决定系数(R2)相对较小。3种方式中,BP神经网络方法所建立的毛管持水量和饱和含水量转换函数均为最佳,回归方法所建立的田间持水量的转换函数要好于BP神经网络方法和Rosetta模型,Rosetta模型对土壤水分常数的预测效果不如其他两种方式。研究可为青海三江源地区土壤水力特性参数研究以及区域尺度上土壤水分估算提供科学依据。     

Soil Particulate Organic Matter (POM) Prediction in a Mountainous Watershed using Artificial Neural Networks

Soil organic matter is a very important component of soil that supports the sustainability and quality in all ecosystems, especially in arid and semi-arid regions. A comparison study was carried out to verify when the artificial neural network (ANN) and multiple linear regression (MLR) models are appropriate for the prediction of soil organic matter (SOM) and particulate organic matter (POM). Discussions of advantages and disadvantages are given for both methods. Three different sets of easily available properties (soil properties alone, topographic and vegetation index, a combination of soil and topographic data) were used as inputs and the one affecting the model the most was determined. The smallest prediction errors were obtained by the ANN method; however, the prediction accuracies of the constructed MLR models using different data sets were closed to the ANN models in many cases.     

Stand-scale spatial patterns of soil microbial biomass in natural cold-temperate beech forests along an elevation gradient

This study focuses on spatial heterogeneity in the soil microbial biomass (SMB) of typical climax beech (Fagus crenata) at the stand scale in forest ecosystems of the cold-temperate mountain zones of Japan. Three beech-dominated sites were selected along an altitudinal gradient and grid sampling was used to collect soil samples at each site. The highest average SMB density was observed at the site 1500 m a.s.l. (44.9 gC m⁻²), the lowest was recorded at the site 700 m a.s.l. (18.9 gC m⁻²); the average SMB density at the 550 m site (36.5 gC m⁻²) was close to the overall median of all three sites. Geostatistics, which is specifically designed to take spatial autocorrelation into account, was then used to analyze the data collected. All sites generally exhibited stand-scale spatial autocorrelation at a lag distance of 10-18 m in addition to the small-scale spatial dependence noted at <3.5 m at the 550 m site. Correlation analysis with an emphasis on spatial dependency showed SMB to be significantly correlated with bulk density at the 550 and 1500 m sites, dissolved organic carbon (DOC) at the 700 and 1500 m sites, and nitrogen (N) at the 550 and 700 m sites. However, no soil parameter showed a significant correlation with SMB at every site, and some variables were also differently correlated (negative or positive) with SMB at different sites. This suggests that the factors controlling the spatial distribution of SMB are very complex and responsive to local in situ conditions. SMB regression models were generated from both the ordinary least-squares (OLS) and generalized least-squares (GLS) models. GLS performance was only superior to OLS when cross-variograms were accurately fitted. Geostatistics is preferable, however, since these techniques take the spatial non-stationarity of samples into account. In addition, the sampling numbers for given minimum detectable differences (MDD_s) are provided for each site for future SMB monitoring.     

Comparing the spatial predictions of soil organic matter determined by two laboratory methods

Abstract. The most common way of assessing soil organic matter content is by loss on ignition, which is both simple and inexpensive. This method tends to overestimate organic matter content because additional weight losses occur during ignition. An alternative, more expensive and time-consuming method for determining soil organic matter content is by an acid dichromate oxidation. This paper compares the results of applying these methods to soil on different parent materials in two arable fields. Summary statistics and correlation coefficients showed that there were consistent relationships between the two sets of values: the stronger was for the sandy soil and the weaker was for the clay loam. This relationship can be used to improve the accuracy with which soil organic matter content is known while using fewer of the expensive measurements and more of the inexpensive ones. Two approaches to prediction were compared: the geostatistical method of cokriging, and simple linear regression. These were used to predict organic matter determined by an acid dichromate oxidation from the loss on ignition. The estimates from cokriging were more accurate but the method requires the spatial correlation to be modelled reliably. The regression results showed it to be a valuable and practical approach. Using the information from nine carefully selected sampling sites a regression line could be fitted that was representative of the full data.     

Calibration model of microbial biomass carbon and nitrogen concentrations in soils using ultraviolet absorbance and soil organic matter

There is a need for a rapid, simple and reliable method of determining soil microbial biomass (SMB) for all soils because traditional methods are laborious. Earlier studies have reported that SMB‐C and ‐N concentrations in grassland and arable soils can be estimated by measurement of UV absorbance in soil extracts. However, these previous studies focused on soils with small soil organic matter (SOM) contents, and there was no consideration of SOM content as a covariate to improve the estimation. In this study, using tropical and temperate forest soils with a wide range of total C (5–204 mg C g^?1 soil) and N (1–12 mg N g^?1 soil) contents and pH values (4.1–5.9), it was found that increase in UV absorbance of soil extracts at 280 nm (UV₂₈₀) after fumigation could account for 92–96% of the variance in estimates of the SMB‐C and ‐N concentrations measured by chloroform fumigation and extraction (P < 0.001). The data were combined with those of earlier workers to calibrate UV‐based regression models for all the soils, by taking into account their varying SOM content. The validation analysis of the calibration models indicated that the SMB‐C and ‐N concentrations in the 0–5 cm forest soils simulated by using the increase in UV₂₈₀ and SOM could account for 86–93% of the variance in concentrations determined by chloroform fumigation and extraction (P < 0.001). The slope values of linear regression equations between measured and simulated values were 0.94 ± 0.03 and 0.94 ± 0.04, respectively, for the SMB‐C and ‐N. However, simulation using the regression equations obtained by using only the data for forest profile soils gave less good agreement with measured values. Hence, the calibration models obtained by using the increase in UV₂₈₀ and SOM can give a rapid, simple and reliable method of determining SMB for all soils.     

Neural network models to predict soil water retention

Soil water retention curves are needed to describe the availability of soil water to plants and to model movement of water through unsaturated soils. Measuring these characteristics is time-consuming, labour-intensive and therefore expensive. This study was conducted to develop and evaluate functions based on neural networks to predict soil water retention characteristics. Dutch and Scottish data sets were available; they contained data on 178 and 165 soil horizons, respectively. A series of three neural networks (A, B and C) was developed. Neural network A had the following input variables: topsoil, bulk density, organic matter, clay, silt and sand content. In addition neural network B had matric potential as input, and network C included soil structural data expressed as the upper and lower boundary of the ped-size class. Neural network A had three output variables: the volumetric water content at matric potentials of 0, –100 and –15 000 hPa. Both models B and C had volumetric water content, at the matric potential given as input, as output variable. The networks were tested against independent data that were extracted from the original sets of soil profiles. Accuracy of the predictions was quantified by the root of the mean squared difference (RMSE) between the measured and the predicted water contents, and the coefficient of determination (R²). For network A the RMSE varied for the three estimated water contents from 0.0264 to 0.0476 cm³ cm^–3, and R² varied from 0.80 to 0.93 for the individual model outputs. Networks B and C had an RMSE of 0.0435 and 0.0426 cm³ cm^–3, respectively. For both networks, R² was 0.89. The neural networks performed somewhat better than previous regression functions, but the improvements were not significant.