基于组合模型的庐山森林土壤有效铁光谱间接反演研究

铁是植物生长的重要微量营养元素之一,土壤有效铁含量对林地环境起着重要的影响,利用土壤光谱预测技术获取土壤有效铁含量信息具有重要意义。而要通过土壤光谱直接预测土壤有效铁含量是难以实现的,因此提出利用土壤有机质含量与有效铁含量之间的相关性,探讨间接估算土壤有效铁含量的可行性。以庐山森林土壤样本为研究对象,研究基于偏最小二乘回归(PLSR)和径向基函数(RBF)神经网络的组合模型预测土壤有机质含量的适用性,并且通过构建有机质含量与有效铁含量的二项式线性模型,对土壤有效铁含量进行间接反演,探讨不同权重下的最优组合模型。结果表明,组合模型的预测效果优于偏最小二乘回归和RBF神经网络单个模型,并且熵值组合为最优组合模型,其中,土壤有机质的反演模型验证的决定系数(R~2)为0.81,均方根误差(RMSE_p)为11.54 g kg~(-1),测定值标准差与标准预测误差的比值(RPD)为2.18;有效铁的间接反演模型R~2为0.70,RMSE_p为21.60 mg kg~(-1),RPD为1.77。通过土壤有机质构建土壤有效铁含量的光谱反演间接模型,在光谱反演模型中,组合模型能较大限度地利用各种预测样本信息,能有效减少单个预测模型中随机因素的影响,增强预测稳定性,提高模型的预测能力。因此,组合模型可对土壤有机质含量的光谱预测及土壤有效铁的间接预测发挥更好的作用。

Composite-Model-Based Indirect Reversion of Soil Available Iron Spectrum of Forest Soil in Lushan

Objective]As iron is one of the nutrient elements essential to plant growth,the content of soil available plays an important role in evolution of forest environment. The technology of hyper-spectral remote-sensing(RS)provides a new means for determination of soil physical and chemical components in laboratory.Method]In this study,the relationship between soil organic matter and available iron was used to predict the content of available iron in soil indirectly. Besides the traditional single factor prediction model has its own limitation. In order to solve the problem of errors with the single-factor model,this study brought forth a composite model to improve accuracy of the prediction of soil organic matter contents in forest soils at a regional scale with the Vis-NIR spectrum technique. A total of 190 soil samples were collected from the 0~20 cm soil layers of the forests typical of Lushan region in Jiangxi Province. An ASD FieldSpec3 spectrograph diameter equipped with a high intensity contact probe was used to measure original spectral reflectance of the samples in line with standard procedure of the laboratory conditions,and mean while,the soil samples were analyzed for physical and chemical properties. Out of the 190 soil samples,143 were picked out as samples for modeling and the remaining 47 verification ones. Result]The results showed that a significant positive correlation was found between the contents of soil organic matter and soil available iron,and then the binomial model can be built. Based on the results of spectral inversion of soil organic matter content, the contents of soil available iron were retrievable indirectly. Among the spectral inversion models,based on the full band(400~2450 nm)of soil spectra in this study,PLSR(Partial Least Square Regression)of the optimal linear fitting model and RBF(Radial Basis Function)neural network of the nonlinear fitting model were selected to form a combination to figure out arithmetic mean weight coefficients and to project an optimal combination model based on squared,reciprocal and entropy weight coefficients. Accuracies of the predictions of soil available iron content were evaluated by root mean squared error(RMSEp),ratio of partial deviation (RPD)and determination coefficients(R2). Results show that the combination model is superior to the two separate models in prediction accuracy. Among the combination models,the entropy weight coefficient combination model is the best,with determination coefficient(R2)in verification model,root mean squared error(RMSEp)and ratio of standard deviation of determination to standard deviation(RPD)of the soil organic matter prediction being 0.81,11.54 g kg-1 and 2.18,the soil available iron indirect prediction being 0.70,21.60 mg kg-1 and 1.77,respectively. The combination model is able to make use to a maximum margin of various information of the samples for prediction,reduce effectively the impacts of random factors in using single prediction models,enhance prediction stability and raise prediction capability of the models.Conclusion]All the findings of the study demonstrate that it is feasible to in directly predict soil available iron contents in forest soils by making use of hyper-spectral RS data. In the end,it can be concluded that the combination model can play a pretty good role in predicting soil organic matter content and indirect predicting soil available iron content.