A recognition method of soybean varieties based on near-infrared spectroscopy and improved deep learning model

Rapid and accurate detection of soybean varieties is of crucial importance for the successful cultivation, management, and selection of ecological types. The accuracy of the techniques currently employed is not satisfactory. Thus, in order to overcome the above shortcomings, a new recognition method of soybean varieties was proposed based on the extraction of near-infrared spectra combined with an improved LeNet-5 model. First, the original NIR data of soybean varieties were preprocessed by Multivariate Scattering Correction (MSC) for the purpose of baseline adjustment. Second, 114 characteristics from 1845 preprocessed wavenumbers were extracted by using Competitive Adaptive Reweighted Sampling (CARS) algorithm. Finally, the Joint Attention Collaborative-Sectoral Classification Model (JAC-SCM) was constructed with a detection accuracy of 99.62% for soybean varieties. The result showed that the accuracy of the proposed model improved by 16.68%, compared with traditional machine learning and deep learning models. This method provides a novel approach for crop variety recognition.