Abstract: This study aimed to achieve rapid and non-destructive testing of unknown wheat hardness samples’harness. Monte Carlo cross-validation statistical rules were used to identify wheat hardness spectral data and exclude abnormal samples. First derivative method was used to preprocess the spectral data, to obtain a representative prediction set and correction set, spectral physicochemical value symbiotic distance method was used to divide the set of wheat spectral data, and obtain the prediction set samples. First derivative method was used to preprocess the spectral data to eliminate high-frequency noise, baseline drift, sample background and other irrelevant information in the acquired spectral data, thereby to reduce the impact of various non-target factors on the detection model. Based on competitive adaptive reweighting algorithm, wavelength variables that were useful to the model were screened, thereby improving the stability and predictability of the prediction model. Finally, a partial least squares method of wheat hardness prediction model was established. The results showed that the CARS-PLS model evaluation parameters R and RMSEP reached 0.884 3 and 0.543 6, respectively, indicating that this model, based on near infrared spectroscopy, could accurately predict wheat hardness.