Hyperspectral Imaging Combined With Deep Transfer Learning for Rice Disease Detection

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Rice diseases pose a significant threat to global food security.
• Accurate and rapid disease detection is crucial for effective crop management.
• Developing individual disease classifiers for each rice variety is time-consuming and resource-intensive.

Purpose of the Study:

• To introduce and evaluate deep transfer learning methods for cross-variety rice disease detection using hyperspectral imaging.
• To compare the performance of fine-tuning, deep CORAL, and DDC methods in transferring knowledge between rice varieties.
• To explore multi-task transfer learning strategies for enhanced detection accuracy.

Main Methods:

• Hyperspectral imaging (HSI) was used to collect data from four rice varieties.
• A convolutional neural network (CNN) served as the base for three deep transfer learning techniques: fine-tuning, deep CORAL, and DDC.
• Twelve transfer tasks were designed to assess model performance across different rice varieties.
• Saliency maps were employed to visualize key spectral features identified by the CNN.

Main Results:

• Fine-tuning demonstrated the best transferable performance, achieving over 88% accuracy in most transfer tasks.
• Deep CORAL outperformed DDC, with accuracies exceeding 80% in four transfer tasks.
• Both pair-wise and multi-task transfer learning strategies showed promising results.
• Saliency map analysis indicated overlapping important wavelength ranges with and without transfer learning.

Conclusions:

• Deep transfer learning effectively enables rice disease detection across different varieties, reducing the need for variety-specific models.
• Hyperspectral imaging combined with deep transfer learning presents a viable, efficient, and cost-effective method for field-based rice disease detection.
• This integrated approach holds significant potential for improving precision agriculture and disease management strategies in rice cultivation.