基于SVR算法的林地土壤氮含量高光谱测定

提出了一种利用高光谱技术进行杉木林土壤全氮测定的新方法。以FieldSpec®3地物光谱仪采集杉木林土壤148份, 随机分成校正集(100份)和检验集(48份)。以不同方法实现了土壤光谱的预处理, 并采用偏最小二乘回归算法(PLS)建立土壤氮含量估测模型对其进行比较分析, 发现小波除噪结合多元散射校正能最有效地消除原始光谱的噪声与背景信息, 此时PLS模型校正集与预测集R²分别为0.891与0.885。为进一步优化模型, 对经小波除噪结合多元散射校正处理后的光谱采用主成分分析法(PCA)降维, 以前4个主成份为输入变量, 采用小二乘支持向量机回归算法(LS-SVR)建立了土壤氮含量估测模型, 其校正集与预测集R²分别提高至0.921与0.917, 具有比PLS算法更高的精度。结果表明:以高光谱技术进行林地土壤氮含量快速监测是可行的, 其中小波去噪结合多元散射校正系光谱预处理的优选方法, 而LS-SVR则是建模的优选方法。

Forest soil nitrogen content estimation using hyperspectra technology based on SVR algorithm

A new method was put forward to measure the total N by hyperspectra technology. 148 fir soil samples were collected using a FieldSpec®3 spectrometer. All samples were divided randomly into 2 groups, one group with 100 samples used as calibrated set, and the other with 48 samples used as validated set. The original spectra were pretreated by different methods, and then the PLS model was established with the spectra in the range of 350-2350 nm to compare the different pretreated methods. It was found that the background information and noise of the spectra could be eliminated by the method of wavelet denoising combined with multiplicative scatter correction effectively, with the calibration R-square (C-R²) Prediction R-square (P-R²) 0.891 and 0.885, respectively. In order to optimize the result, the pretreated spectra were analyzed using the principal component analysis(PCA), and the top 4 principal components were used as the input variables for the least square support vector regression( LS-SVR) model. The C-R²and P-R² of LS-SVR model increased to 0.921 and 0.917, respectively, higher than those of PLS model, which indicated LS-SVR algorithm was more accurate. The result showed that it is feasible to estimate the nitrogen content of fir soil with hyperspectra technology, and the estimation model can be improved by the pretreatment method of wavelet denoising combined with multiplicative scatter correction and the modeling algorithm of LS-SVR.