Integration of Multiscale Spectral Features as the Intermediate Variables for Improved Prediction of Grain Protein Concentration in Winter Wheat

Grain protein concentration (GPC) could be predicted at high accuracy using the reflectance of dried grain powder in the shortwave infrared (SWIR) region. However, the acquisition of grain-level hyperspectral data is time-consuming and such a procedure must be conducted with laboratory conditions. This study proposed an efficient approach to integrating multiscale spectral features as the intermediate variables for improved prediction of GPC in winter wheat by transporting the sensitive features from grain to canopy levels. The performance of the proposed method was compared in the form of ’Dependent variable ~ Independent variable ~ Intermediate variable’ with the traditional canopy-level ’GPC ~ Spectral feature (SF) ~ Leaf nitrogen concentration (LNC)’ models. Our results demonstrated that the models of ’GPC ~ Wavelet feature (WF) ~ LNC’ exhibited higher validation accuracies than the traditional ’GPC ~ Vegetation index (VI) ~ LNC’ models. Specifically, the canopy-level sensitive WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">canopy-720,5</sub> (WF at 720 nm and the wavelet decomposition scale of 2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">5</sup>) yielded pronounced improvement for GPC prediction, with RMSE values of 1.02% and 0.83% for the anthesis and the post-anthesis stages. Compared with models using the SF- or LNC-based intermediate variables, the model using the intermediate variable of WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">grain-1510,4</sub> (grain-level sensitive WF) performed remarkably better, such as the single-feature model of ’GPC ~ WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">canopy-1510,4</sub> ~ WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">grain-1510,4</sub>, at the post-anthesis stage (RMSE = 0.78%) and multi-feature model of ’GPC ~ (WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">canopy-1510,4</sub> & WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">canopy-1610,5</sub>) ~ (WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">grain-1510,4</sub> & WF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">grain-1610,5</sub>) at the anthesis stage (RMSE = 0.93%). The grain-level spectral features are not expressed in the final model and only canopy-level data are required for model applications. This study represents a first attempt to couple grain and canopy-level SFs to improve the spectroscopic prediction of GPC and provides new insight into the quantitative estimation of grain quality parameters with hyperspectral remote sensing.