A method of maize seed variety identification based on near-infrared spectroscopy combined with improved DenseNet model

Abstract

The development of a real-time online system for rapid and non-destructive identification of seed varieties can significantly improve production efficiency in modern agriculture. Near-infrared spectroscopy technology has become one of the commonly used techniques in seed variety identification due to its fast and non-destructive characteristics. However, existing convolutional neural networks are difficult to reflect the complex nonlinear relationships of the near-infrared (NIR) spectrum, resulting in poor modeling performance, and their high model complexity is not conducive to real-time online identification tasks. Therefore, this study proposes a maize seed variety identification method using near-infrared spectroscopy technology and lightweight deep learning network (BAC-DenseNet). First, a total of 750 samples from 5 different types of maize seeds were taken as the research object. The spectral data were pre-processing using the SGD-SNV, and the identification accuracy was improved by an average of 15.78 %. Then, the attraction–repulsion optimization algorithm combined with Laplacian Eigenmaps (AROA-LE) was used to perform dimension reduction on the pre-processed data, and the dimensionality was reduced from 1845 to 66. Finally, a lightweight deep learning network model (BAC-DenseNet) was constructed based on DenseNet-121 network with layer pruning and the introduction of batch channel normalization (BCN), self-attention and convolution mixed module (ACmix) and convolutional block attention module (CBAM). The experimental results show that the proposed BAC-DenseNet model has an identification accuracy of 99.33 %. Compared with the original network and seven other classical deep learning models, the proposed method has an average improvement of 2.83 %, 3.52 %, and 3.47 % in accuracy, Kappa, and MCC, respectively. Meanwhile, Params, Size, and FLOPs decreased by an average of 9.09 M, 35.08 MB, and 88.66 M, respectively. This method offered high accuracy and reliability in maize seed variety identification, which can provide qualitative indicators for the breeding, planting, and management of maize seed varieties. This study can provide a reference method for variety identification of other agricultural products.