An explainable vision transformer with transfer learning based efficient drought stress identification.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Science

Background:

• Early detection of drought stress is crucial for minimizing crop loss.
• Non-invasive imaging techniques capture subtle plant changes, valuable for machine learning.
• Vision transformers (ViTs) offer advanced feature extraction for complex image analysis.

Purpose of the Study:

• To develop an explainable deep learning pipeline for drought stress detection in potato crops.
• To leverage ViTs for analyzing aerial imagery to identify drought-affected plants.
• To provide an interpretable solution for farmers to improve crop management.

Main Methods:

• Applied two ViT-based approaches: ViT-SVM synergy and an end-to-end ViT classification.
• Utilized aerial imagery for drought stress detection in potato crops.
• Employed attention maps for model interpretability, visualizing stress signatures.

Main Results:

• Achieved high accuracy in identifying drought stress using ViT models.
• Demonstrated ViT's capability in capturing intricate spatial relationships indicative of stress.
• Attention maps successfully highlighted key drought stress features in aerial images.

Conclusions:

• The proposed explainable ViT pipeline offers a robust method for drought stress detection.
• The approach provides valuable insights into plant features associated with drought stress.
• This interpretable AI solution supports informed decision-making for effective crop management.