pone.0340807.t008 -

Deploying rice disease detectors in the field remains challenging because models that are accurate in the lab are often poorly calibrated and provide limited uncertainty estimates, raising the risk of costly misclassification. This paper proposes a multi-objective Big-Bang Big-Crunch (MO-BBBC) framework that jointly performs disease detection and variety classification while optimizing six deployment-oriented criteria: classification error, calibration quality, uncertainty estimation, model size, inference latency, and energy consumption. The proposed framework presents conditional temperature scaling, an adaptive scheme that mitigates over-calibration and preserves reliability. The framework is implemented in Python on a lightweight two-headed classifier and evaluated on the Paddy Doctor dataset, MO-BBBC base framework achieves 90.6% disease accuracy and 97.9% variety accuracy; improves calibration to (% better than strong post-hoc baselines); achieves micro-AUC of 0.994/0.999 and micro-AP of 0.961/0.994 (disease/variety); delivers robust OOD detection (AUROC = 0.887/0.886); and supports real-time inference at ms and ms per 64-sample batch on CPU/GPU with Monte Carlo Dropout uncertainty. The resulting Pareto set enables practitioners to trade accuracy for efficiency and reliability, narrowing the gap between prototype validation and field deployment in precision agriculture.