基于主成分分析和随机森林的橡胶小苗叶片磷含量高光谱预测

Prediction of Rubber Seedling Leaf Phosphorus Contents Based on Hyperspectral Technology and Random Forest Approach

为指导橡胶苗施肥管理,通过设置不同磷素水平的橡胶小苗砂培试验,利用高光谱仪测得不同磷处理水平下的橡胶苗叶片光谱反射率,进而运用主成分分析法对原始光谱反射率数据进行压缩及提取主要信息,最后以提取的主要信息为输入变量,结合多重线性回归、偏最小二乘回归和随机森林模型对橡胶苗叶片磷含量进行预测。结果表明：不同磷处理水平下橡胶苗叶片磷含量差异显著;在可见光波谱范围401~669 nm内,光谱反射率随施磷量的增加而增加;不论在训练集还是在验证集,随机森林模型的预测精度最高,训练集和验证集中预测值和实测值之间的相关系数r分别为0.985 0和0.988 4,均方根误差RMSE分别为0.016 5和0.018 2,平均相对误差MRE分别为5.74%和5.99%,模型性能指数RPD分别为3.83和4.01,证明高光谱技术可以快速、准确地诊断橡胶苗叶片磷含量。

The aim of this study was to test the utility of hyperspectral technology in combination with random forest （RF） approach and principal component analysis （PCA） to estimate ~ rubber seedlings leaf phosphorus contents. A sand culture experiment was carried out to grow rubber seedlings. These rubber seedlingswere cultivated with Hoagland＇s nutrient solutions set at five levels of phosphorus concentration. Leaves of rubber seedlings were sampled at 85, 100, 115, 133, 148 and 163 days after the culture, respectively. A total of 150 samples were collected. At each sampling date, collected leaves were sent to laboratory as soon as possible, and then their leaf hyperspectral reflectance was measured by ASD FieldSpec 3 spectrometer. Phosphorus contents of the corresponding leaves were also analyzed using the conventional chemical analysis method. A second order low-pass digital Butterworth filter with normalized cutoff frequency 0.5 was used to the original spectra to filter out the noise information. Next, PCA was applied to the denoising leaf hyperspectral reflectance to extract the principal components （PCs）. Then, these selected PCs were used as input variables, and multiple linear regression （MLR）, partial least-squares regression （PLSR）, as well as RF model were employed to estimate rubber seedlings leaf phosphorus contents, respectively. Results indicated that rubber seedling leaf phosphorus contents at various treatments were significantly different at p〈0.01 level. Within the range 401 4669 nm, rubber seedling leaf reflectance increased with increasing leaf phosphorus contents. RF model got the best prediction （esults. Correlation coefficients （r） between predicted leaf phosphorus contents and measured leaf phosphorus contents were 0.9850 and 0.9884 for train dataset and test dataset, respectively; values of root mean squared errors （RMSE） were 0.0165 and 0.0182 for train dataset and test dataset, respectively; values of mean relative error （MRE） were 5.74% and 5.99% for train dataset and test dataset, respectively, and values of ratio of performance to deviation （RPD） were 3.83 and 4.01 for train dataset and test dataset, respectively. The results verified that hyperspectral remote sensing could be used to rapidly, and accurately predict the leaf phosphorus contents for rubber seedlings.