Laboratory shortwave infrared reflectance spectroscopy for estimating grain protein content in rice and wheat

Grain protein content (GPC), as an important agronomic indicator of nutritional value, is commonly used for optimized cultivation, classified harvesting, and quality grading. Previous studies on the spectroscopic estimation of GPC were often conducted at the canopy level with vegetation indices (VIs) derived from the visible and near-infrared (VNIR) bands, which are apart from the absorption features of protein predominately located in the shortwave infrared (SWIR) region. Since those absorption features are mostly masked by water for living plants, the causal relationships between sensitive spectral bands and protein absorption features are poorly understood. Recently, continuous wavelet spectral analysis has been commonly applied to the spectroscopic estimation of vegetation parameters in the way of extracting the optimal wavelet feature (denoted as WF (lambda,S), lambda is the shifting factor in wavelength and S is the scaling factor in the power number of two) or the red-edge position. This study proposed the wavelet-based SWIR edge position (SWEP) for estimating the GPC in cereal crops through extending the red edge position extraction technique to laboratory SWIR reflectance spectroscopy. The performance of SWEPs and WFs was evaluated in coefficient of determination (R (2)), root mean squared error (RMSE) and relative RMSE (RRMSE) with experimental data over field trials. Our results demonstrated that the protein absorption features in the SWIR region could be enhanced using wavelet-based methods towards accurate estimation of the GPC across rice and wheat crops. The optimal features WF1510,4 (R (2) = 0.96) and WF1610,5 (R (2) = 0.93) exhibited similar performance for GPC estimation as compared to normalized difference protein indices (NDPIs) on calibration data, but the latter NDPI models had much poorer performance on independent validation data (NDPI1745,1780: R (2) = 0.35, RMSE = 1.83%, RRMSE = 21.70%; NDPI2080,2160: R (2) = 0.57, RMSE = 1.48%, RRMSE = 17.55%). The GPC could be estimated with high accuracies on both calibration (R (2) = 0.94) and validation (R (2) = 0.92, RMSE = 0.64%, RRMSE = 7.59%) using the SWEP extracted from the 1450-1650 nm range at Scale 5. Given their close predictive performance, the wavelet-based SWEP was less affected by the degradation of spectral resolution in the input data than the WF. The application of SWEPs could help us better understand the spectroscopy mechanism of GPC estimation from grain reflectance spectra. It also suggests that not just the commonly used red edges but also the insufficiently exploited SWIR edges are important in the entire spectral domain for quantifying vegetation properties.