【摘要】 采集新疆北疆棉田385个自然土壤样本,将筛选出的土壤样品分别过2、1、0. 5、0. 15 mm筛并测定其原始光谱反射率,利用支持向量机(Support vector machine,SVM)、偏最小二乘回归(Partial least squares regression,PLSR)和多元逐步线性回归(Stepwise multiple linear regression,SMLR)方法对土壤原始光谱及其12种光谱变换数据分别构建土壤全氮含量的估测模型,并对模型精度进行检验。结果表明,土壤原始光谱特征在各个波段与全氮含量相关性都较差,不同形式的数据变换均能够提高光谱反射率与全氮含量的相关性,同一种数据变换形式在不同粒径处理中最大相关系数所对应的波段位置差异不大。从不同粒径处理的拟合精度来看,过筛粒径越小对全氮含量的估测精度越高,3种方法的最优拟合模型都是过0. 15 mm筛的处理,其中SVM方法采用(lgR)’变换后,构建模型R2c为0. 898 7,RMSEc为0. 018 1,RPD为2. 704 9,PLSR和SMLR方法均采用R’变换,构建模型的R2c分别为0. 852 0和0. 819 6,RMSEc分别为0. 041 3和0. 043 6,RPD分别为2. 554 9和2. 437 4,3种方法在该过筛处理下均能够很好地估测土壤全氮含量。用未参与建模的样本对3种最优模型进行验证,SVM、PLSR和SMLR模型的检验R2分别为0. 822 9、0. 771 5和0. 705 4,SVM方法优于PLSR和SMLR,模型具有较好的精度和稳定性,从模型的预测误差来看,土壤全氮含量越低其预测误差也越大,在氮素含量较低的情况下无法直接通过光谱反射特征准确反演。更多还原

【Abstract】 Hyperspectral remote sensing technology is a powerful tool in the analysis of soil compositions as well as soil physical and chemical properties. Totally 385 natural soil samples were collected from cotton fields in North Xinjiang Province,the selected soil samples according to the total nitrogen content were processed by 2 mm,1 mm,0. 5 mm and 0. 15 mm sieves, and their spectral reflectance characteristics were measured. After the transformation of spectral data with twelve forms,the spectral inversion models of soil nitrogen content were established based on support vector machine( SVM),partial least squares regression( PLSR) and stepwise multiple linear regression( SMLR),and the accuracy and universality of the model were tested. The results showed that there was no significant correlation between the original spectral characteristics and soil nitrogen content,and which can be improved by different data transformations. In the same data transformation,there was no obvious difference in the band position corresponding to the maximum correlation coefficient in different particle size processing. According to the fitting accuracy of different particle size treatments,the smaller the particle size of the sieve was,the higher the precision of the total nitrogen content was,the optimal fitting models of the three methods were all processed by 0. 15 mm sieve treatment,the SVM method used( lgR) ’ transformation,the model Rc2 was 0. 898 7,the RMSEcwas 0. 018 1 and the RPD was 2. 704 9,the PLSR and the SMLR methods used R’ transformation,the Rc2 were 0. 852 0 and 0. 819 6,the RMSEc was 0. 041 3 and 0. 043 6,and the RPD was 2. 554 9 and 2. 437 4,respectively. The optimal model was checked with the samples which were not involved in building model and the R2 of SVM,PLSR and SMLR were 0. 882 9,0. 771 5 and 0. 705 4,respectively. From the prediction error of the model,the lower the soil total nitrogen content was,the greater the prediction error was,it was impossible to accurately estimate the soil total nitrogen content by spectral reflectance characteristics.更多还原