Identification of maize seed varieties based on stacked sparse autoencoder and near-infrared hyperspectral imaging technology

Maize seed variety identification is the key to improving the quality and yield of maize. The study aimed to investigate a stacked sparse autoencoder combined with a cuckoo search (CS) optimized support vector machine (SSAE-CS-SVM) to meet the identification requirements of accurate detection. First, the near-infrared (NIR) (871.61-1766.32 nm) hyperspectral data of maize seeds were processed using Savitzky-Golay (SG) combined with standard normal variables (SNV). Subsequently, SSAE, SAE, principal component analysis (PCA), and competitive adaptive reweighted sampling were employed for feature extraction. Finally, the recognition model was constructed using the SoftMax regression model, SVM, and CS optimized SVM (CS-SVM). The results indicated that the SSAE-CS-SVM model achieved satisfactory performance (the training set and testing set accuracies were 99.45% and 95.81%, respectively). This study confirmed the great potential of combining NIR hyperspectral imaging technology with deep learning algorithms for maize seed variety identification. Practical applications The identification of maize seed varieties is extremely important for ensuring seed purity and improving maize quality and yield. A method based on SSAE and NIR hyperspectral imaging technology was studied to identify maize seed varieties. The method proposed in this paper could identify maize seed varieties non-destructively and accurately, which provided a new way for the online detection of seed varieties.