Hyperspectral imaging combined with CNN for maize variety identification

Area of Science:

• Agricultural Science
• Image Processing
• Machine Learning

Background:

• Maize is a major global food crop with numerous varieties that are difficult to distinguish visually.
• Accurate identification of hybrid maize varieties is crucial for seed production and agricultural management.

Purpose of the Study:

• To develop a rapid and effective non-destructive method for identifying hybrid maize varieties.
• To leverage hyperspectral imaging technology and convolutional neural networks (CNNs) for this identification task.

Main Methods:

• Hyperspectral images (900-1700nm) of 735 maize seeds from seven hybrid varieties were acquired.
• Spectral reflectance data underwent Savitzky-Golay (SG) Smoothing and Maximum Normalization (MN) preprocessing.
• Feature wavelengths were selected using Competitive Adaptive Reweighting Algorithm (CARS) and Successive Projection Algorithm (SPA), then mapped to 3D image features.
• A five-layer CNN was employed to classify the maize varieties based on these 3D features.

Main Results:

• The optimal maize variety identification model achieved 96.65% accuracy on the test set.
• The model demonstrated a detection frame rate of 1000 Fps/s in a GPU environment.
• The best performing model utilized an input feature dimension of 768 and a layer depth factor (d) of 1.0.

Conclusions:

• The proposed hyperspectral imaging and CNN-based method enables rapid, accurate, and non-destructive identification of maize varieties.
• This approach offers a promising new strategy for identifying seeds of maize and other crops.