Recognition of maize seed varieties based on hyperspectral imaging technology and integrated learning algorithms

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Maize seed purity is crucial for crop yield, but traditional identification methods are inefficient.
• Developing rapid and accurate techniques for maize seed purity assessment is essential.

Purpose of the Study:

• To propose a novel method for rapid and accurate maize seed purity identification using hyperspectral imaging and ensemble learning.
• To evaluate the performance of different preprocessing, dimensionality reduction, and classification techniques.

Main Methods:

• Hyperspectral imaging (400-1000 nm) was used to collect spectral data from maize seed endosperm.
• Data preprocessing included Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and Savitzky-Golay First Derivative (SG1).
• Dimensionality reduction was performed using Iteratively Retains Informative Variables (IRIV) and Competitive Adaptive Reweighted Sampling (CARS).
• Recognition models were built using k-nearest neighbor (KNN), support vector machine (SVM), linear discriminant analysis (LDA), and decision tree (DT).
• A Random Subspace Ensemble Learning (RSEL) model was developed with LDA as the base classifier.

Main Results:

• Multiplicative Scatter Correction (MSC) and Iteratively Retains Informative Variables (IRIV) showed the best performance in preprocessing and dimensionality reduction, respectively.
• Linear Discriminant Analysis (LDA) achieved the highest precision among base classifiers.
• The optimized MSC-IRIV-RSEL model improved precision from 0.9333 to 0.9556 and the Kappa coefficient from 0.9174 to 0.9457.

Conclusions:

• Hyperspectral imaging technology combined with subspace ensemble learning provides a promising new approach for maize seed purity recognition.
• The developed method offers a significant improvement in accuracy and efficiency over traditional techniques.