Identification of varieties of sorghum based on a competitive adaptive reweighted sampling-random forest process

To address issues relating to the advantages of varieties of sorghum in construction, breeding, and brewing of the seed repository, a visible–near-infrared hyperspectral imaging (VNIR-HSI) non-destructive technique was proposed to detect different varieties of sorghum. The VNIR-HSI system was used to collect spectrum images for 27 types of varieties of sorghum, and the spectral data were pre-processed using Savitzky–Golay (S–G) smoothing filters, the standard normal variate (SNV), and multiplicative scatter correction (MSC). Competitive adaptive reweighted sampling (CARS) was used for dimensionality reduction. Based on full-spectrum and characteristic spectral data, classification models were developed using a random forest (RF) algorithm. The tested results indicated that precisions of calibration and prediction sets of the RF model established based on the full-spectrum reach 94.58% and 64.44%, respectively. The CARS algorithm was adopted to extract 20 characteristic wavelengths from sorghum spectra. The precisions of the calibration and prediction sets for the CARS-RF model reach 95.00% and 84.07%, respectively. Using the confusion matrix to calculate Cohen's kappa values, the calibration and prediction Cohen's kappa values for the full sample were 0.9212 and 0.9231 respectively, indicating that the evaluation results are almost identical to the correctness results. The applied model can achieve favorable effects when detecting each cultivar of sorghum. The results show that the modeling method integrating VNIR-HSI technique with CARS-RF can provide a rapid non-destructive testing method for detection of varieties of sorghum, and offer an idea for detecting cultivars of coarse cereal crops.