Identification of sorghum variety using hyperspectral technology with squeeze-and-excitation convolutional neural network algorithms

In this study, hyperspectral technology along with a combination of squeeze-and-excitation convolutional neural networks and competitive adaptive reweighted sampling (CARS-SECNNet) was developed to identify sorghum varieties. In addition, the support vector machine (SVM) and random forest (RF) models were established for the rapid identification of sorghum varieties and compared with the CARS-SECNNet model. Two preprocessing methods, wavelet transform (WT) and multiplicative scatter correction (MSC), were applied to preprocess sorghum hyperspectral data. The average abnormal scores for WT and MSC were 1.0619 and 0.5096, respectively, indicating that MSC gave the best preprocessing results. After preprocessing using MSC, the average identification rates of sorghum variety using the of CARS-SECNNet, RF and SVM models were 99.39%, 88.71% and 93.64%, respectively. The CARS-SECNNet model was optimized up to the maximum accuracy of 99.79%, and its generalizability was validated, achieving a validation accuracy of 91.04%. This study proves that the combination of hyperspectral technology with the CARS-SECNNet model for sorghum variety identification offers higher accuracy and rapidity compared with traditional machine learning.