Near infrared spectroscopy predicts crude protein concentration in hemp grain

Abstract This study evaluated near‐infrared spectroscopy (NIRS) for nondestructive crude protein (CP) prediction in hemp (Cannabis sativa L.) grain and validated the biological basis of spectral predictions. Note that 149 whole grain samples from 38 cultivars were collected from New York trials (2017–2021) and validated for CP by combustion. Seven preprocessing methods were tested using 100 training/testing splits, with standard normal variate transformation following Savitzky–Golay filtering selected as optimal. Comparing algorithms showed that partial least squares regression (PLSR) significantly outperformed support vector machines and random forest. The best preprocessing method and algorithm was applied to 1000 additional splits. Optimal models contained 12 components with mean performance of root mean square error [RMSE] = 9.94, r2 = 0.84, relative predicted deviation [RPD] = 2.5, and ratio of performance to interquartile distance [RPIQ] = 3.94. More than 99% of the models had, at minimum, the ability to distinguish between high and low values, with 93.2% capable of quantitative prediction. To validate biological relevance, a protein‐focused model was developed using three known protein absorption bands (1200–1250, 1500–1550, and 2040–2090 nm). These models had substantially reduced performance with 86% of models capable of distinguishing between high and low values but only 14% of models capable of quantitative prediction. However, this targeted approach offers evidence that NIRS predictions are biologically grounded in protein‐specific spectral features rather than spurious correlations. This research demonstrates the promise and biological validity of NIRS for hemp grain CP assessment, supporting applications in breeding programs, although applications demanding more accurate prediction will require better models.