基于随机森林法的棉花叶片叶绿素含量估算

为了高效和无损地估算棉花叶片的叶绿素含量,本研究测定了棉花光谱反射率及叶绿素含量(soilandplant analyzerdevelopment,SPAD)值,对光谱数据进行包络线去除处理、立方根转换和倒数转换,以SPAD值与反射光谱之间的相关性为基础,通过随机森林法筛选出对棉花叶片SPAD值影响较大的特征波段,构建估算棉花叶片SPAD值的BP神经网络(back propagation artificial neural networks, BP ANN)、偏最小二乘回归(partial least squares regression,PLSR)两个模型。结果表明,在605～690nm范围内的反射率与SPAD值相关性达0.01显著水平,均呈负相关,相关系数最高值为-0.619。与原始光谱相比,经过变换后的棉花反射率与SPAD值相关性结果相差较大,其中去除包络线光谱在550～750 nm波段范围有效提高了相关性,相关性效果优于倒数转换数据和立方根转换数据。随机森林法能够有效评出对SPAD值影响较大的特征波段,进而提高模型估算精度。在两种模型中,基于去除包络线光谱建立的PLSR和BP神经网络模型的决定系数R~2分别为0.92、0.83,说明这两种模型的估算能力较好;两种模型RMSE分别为0.88、1.26, RE分别为1.30%、1.89%,表明PLSR模型的估算精度比BP神经网络模型高。从模型的验证效果来看,PLSR模型在估算棉花SPAD值方面有一定的优势和参考价值。

Estimation of leaf chlorophyll content in cotton based on the random forest approach

The main objective of this study is the estimation of the leaf chlorophyll content efficiently and harmlessly. SPAD values and spectral data were collected from field observation. Original spectra processed to continuum-removal transformation, cube-root transformation and reciprocal transformation. Based on the correlation between SPAD values and canopy spectral reflectance, we selected characteristic bands by random forest approach to establish two kinds of estimating models, including back propagation artificial neural network (BP ANN) model and partial least squares regression (PLSR) model. The reflectivity in the range of 605-690 nm was negatively correlated with the SPAD value at P < 0.01, with the correlation coefficient of -0.619. After transformations, the spectral reflectance exhibited different correlations with SPAD value, continuum-removal spectra improved the correlation in the range of 550-750 nm, and had a better correlation with SPAD value than cube-root and reciprocal transformations. Random forest approach effectively evaluated the characteristic bands with large influence on SPAD value, which can help improve the estimation accuracy of the model. R ² of the PLSR and BP neural network model based on continuum-removal spectra was 0.92 and 0.83 respectively, show the two models with good stability in estimation of cotton SPAD values. The RMSE of the two models was 0.88, 1.26, and RE was 1.30% and 1.89% respectively, which indicates that estimation accuracy of PLSR model is higher that of BP neural network model. From the validation of the model, PLSR model has certain advantages and reference value in estimating chlorophyll content of cotton.