Comparison of Estimation Models for Hyperspectral-based Rape Leaf SPAD

  Objective  In order to compare the estimation model effect of SPAD of rape leaves based on hyperspectral parameters.

  Method  Models of partial least squares regression (PLSR), back propagation neural network (BPNN), support vector regression (SVR), and deep neural network (DNN) based on the spectral parameters selected from the correlation analysis between the spectral reflectance parameters and SPAD data were constructed and compared for the estimation of chlorophyll SPAD of rape leaves.

  Result  The SPADs and the original spectra in the blue wave of 425-495 nm and red wave of 665-680 nm of the leaves had a weak positive correlation. However, significantly inverse correlations between the SPADs and the green-yellow band of 510-650 nm and between that and the red edge band of 690-735nm were observed. The negative correlation coefficient between S_Db and S_Dy was as high as −0.98, while the positive correlation coefficient between CARI and MCARI, CI and NDVI₇₀₅ 0.99. The 3 sets of S_Db and S_Dy, CARI and MCARI, and CI and NDVI₇₀₅ had significant linear correlations with the leaf SPAD. They could be somewhat interchangeable rendering them potential for accuracy improvement. The DNN model had an R² of 0.93 and an RPD of 3.92 indicating a high predictability of the two models. They were followed by SVR, while PLSR and BPNN models being similar.

  Conclusion  There were different degrees of correlation between the SPADs and the spectral parameters of rape leaves. The non-linear prediction model based on machine learning showed higher stability and predictability than the others, and the deep learning algorithm more effective in estimating SPAD of rape leaves.