Pre-symptomatic detection of wheat stem rust using hyperspectral imaging and deep learning.

Area of Science:

• Plant Pathology
• Agricultural Engineering
• Computer Science

Background:

• Wheat stem rust (Puccinia graminis f. sp. tritici) poses a significant threat to global wheat yields.
• Early detection of wheat stem rust is crucial for effective disease management and preventing widespread crop loss.

Purpose of the Study:

• To evaluate the efficacy of hyperspectral imaging combined with deep learning for pre-symptomatic detection of wheat stem rust.
• To compare the performance of different deep learning models for early disease identification.

Main Methods:

• A time-series hyperspectral dataset was collected from wheat plants post-inoculation (4-9 days post inoculation).
• Seven deep learning models were assessed, with the top three undergoing weighted cross-entropy optimization.
• Model interpretability was analyzed using input gradient analysis, SHAP attribution, and vegetation index screening.

Main Results:

• Weighted optimization improved F1-scores by 10.0%-18.4%.
• The best model achieved high F1-scores (0.94 at DPI 4, 0.99 at DPI 5) for pre-symptomatic detection, preceding visible symptoms.
• The 480-550 nm spectral region was key for pre-symptomatic detection, while the 750-870 nm region provided general disease information.

Conclusions:

• Hyperspectral imaging and deep learning accurately detect wheat stem rust before symptoms are visible under experimental conditions.
• This approach provides a foundation for developing field-scale early warning systems for wheat diseases.