Related Experiment Videos
An uncertainty-aware evaluation framework based on hierarchical vision transformers for robust cross-domain plant
Tripti Shrivastava1, Ayush Kumar Agrawal1, Abhinav Shukla1
1Department of Information Technology and Computer Science, Dr. C. V. Raman University, Bilaspur, India.
Scientific Reports
|May 27, 2026
Summary
This study introduces an uncertainty-aware Hierarchical Vision Transformer (HViT) for reliable plant leaf disease detection. The framework improves accuracy and calibration in real-world agricultural settings, enhancing precision agriculture.
Area of Science:
- Computer Vision
- Plant Pathology
- Precision Agriculture
Background:
- Deep learning models excel at plant leaf disease classification but struggle with domain shift and lack reliable uncertainty estimation.
- Real-world agricultural applications require robust models that perform accurately under varying field conditions.
Purpose of the Study:
- To develop and evaluate an uncertainty-aware cross-domain framework for plant leaf disease classification.
- To improve the reliability, robustness, and calibration of deep learning models in precision agriculture.
Main Methods:
- Implemented a Hierarchical Vision Transformer (HViT) integrating multi-scale feature learning.
- Incorporated Monte Carlo Dropout for predictive uncertainty estimation and temperature-based calibration.
- Conducted bidirectional cross-domain evaluation using controlled (New Plant Diseases Dataset) and field (PlantDoc) datasets.
Main Results:
- Achieved high accuracy: 97.8% on controlled data and 93.6% on field data.
- Significantly improved model calibration, evidenced by lower Expected Calibration Error (ECE), Negative Log-Likelihood, and Brier score.
- Demonstrated enhanced robustness to domain shift with reduced performance degradation and stable uncertainty.
Conclusions:
- Integrating uncertainty estimation and calibration within a hierarchical transformer framework offers a more reliable solution for agricultural disease diagnosis.
- The proposed approach provides a deployment-ready tool for real-world precision agriculture, addressing limitations of existing deep learning models.
