Accurate detection of full-surface ear rot in maize using hyperspectral imaging and deep learning

Area of Science:

• Plant pathology
• Agricultural engineering
• Genetics

Background:

• Maize ear rot significantly reduces crop yield and quality.
• Accurate disease severity quantification is crucial for breeding disease-resistant maize varieties.
• Current manual and RGB-based methods for disease detection are subjective and imprecise.

Purpose of the Study:

• To develop an integrated full-surface hyperspectral imaging system for precise maize ear rot detection.
• To evaluate and compare the performance of machine learning and deep learning models for classifying ear rot lesions.
• To provide an automated tool for high-throughput phenomics research in maize.

Main Methods:

• Developed an integrated line-scan hyperspectral imaging system with synchronous rotation.
• Generated non-redundant full-surface ear images using the ORB-RANSAC algorithm.
• Applied Savitzky-Golay preprocessing, genetic algorithm feature selection, and compared CNN-Bi-LSTM, RF, and other models.

Main Results:

• The convolutional neural network-bidirectional long short-term memory network (CNN-Bi-LSTM) model achieved the highest average overall accuracy of 95.61% ± 0.36%.
• CNN-Bi-LSTM demonstrated superior performance compared to traditional machine learning models like random forest (RF).
• The developed model enables high-precision pixel-level detection of Fusarium-associated maize ear rot symptoms.

Conclusions:

• The integrated hyperspectral imaging system and CNN-Bi-LSTM model offer a non-destructive, accurate method for full-surface maize ear rot detection.
• This approach addresses the limitations of traditional methods, improving the reliability of phenotypic data for resistance evaluation.
• The automated analysis software facilitates high-throughput phenomics, accelerating the discovery of maize resistance genes and contributing to food security.