Multi-objective Big Bang Big Crunch framework for reliable rice disease and variety classification with conditional...

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Field deployment of rice disease detectors is hindered by poor calibration and limited uncertainty estimates in lab-accurate models.
• Misclassification risks in precision agriculture necessitate improved reliability and efficiency for AI models.

Purpose of the Study:

• To propose a multi-objective Big-Bang Big-Crunch (MO-BBBC) framework for joint rice disease detection and variety classification.
• To optimize six deployment-oriented criteria: classification error, calibration quality, uncertainty estimation, model size, inference latency, and energy consumption.
• To enhance model reliability and enable real-time field application in precision agriculture.

Main Methods:

• Developed a multi-objective Big-Bang Big-Crunch (MO-BBBC) framework incorporating conditional temperature scaling.
• Implemented a lightweight, two-headed classifier on the Paddy Doctor dataset.
• Utilized Monte Carlo Dropout for uncertainty estimation and enabled real-time inference.

Main Results:

• Achieved 90.6% disease accuracy and 97.9% variety accuracy.
• Improved calibration significantly compared to post-hoc baselines.
• Demonstrated robust out-of-distribution detection and real-time inference speeds on CPU/GPU.

Conclusions:

• The MO-BBBC framework effectively balances accuracy, efficiency, and reliability for field deployment.
• Conditional temperature scaling enhances model calibration and preserves reliability.
• The framework narrows the gap between prototype validation and practical application in precision agriculture.