Variety identification of seeds is crucial to guarantee crop quality and yield, as well as to ensure the quality and nutritional value of processed maize products. This study proposes a methodology for identifying maize seed varieties based on hyperspectral imaging (HSI) combined with deep learning. The mean spectra of the endosperm side (E1), non-endosperm side (N1), and both sides fused (F1) of maize seeds were extracted, followed by spectral pre-processing using Savitzky-Golay (SG) smoothing and multiplicative scatter correction (MSC), and feature wavelengths were selected from the spectral data using the competitive adaptive reweighted sampling (CARS) algorithm. K-nearest neighbor (KNN), decision tree (DT), support vector machine (SVM), random forest (RF), partial least squares discriminant analysis (PLS-DA), and convolutional neural networks with multi-scale feature fusion (CNN-MFF) were utilized to construct the discriminant models. The study results show that the model established using the spectra of F1 obtains better performance than E1 and N1, with an accuracy of more than 87.22 % on the prediction set. The CNN-MFF model built based on full and feature wavelengths obtained optimal results with accuracies of 97.78 % and 96.11 % on the prediction set, respectively, which proves that the CNN based on multi-scale feature fusion has better applicability and stability. In addition, visualization methods were used to demonstrate the recognition results to visualize the model's classification performance. In summary, using HSI and deep learning for the variety identification of maize seeds is feasible. The proposed method has significant potential for application in spectral analysis and can provide a reference for the online detection of seed quality in crops such as maize. Combining spectroscopic methods to analyze the distribution information of its internal nutrient elements can contribute to directing maize food processing and improving the utilization value of by-products.
扫码关注我们
求助内容:
应助结果提醒方式:
