Confidence and uncertainty aware deep learning for reliable grape leaf disease diagnosis under real world field conditions.

Accurate and reliable diagnosis of grape leaf diseases is essential for sustainable viticulture, enabling timely intervention, reducing economic losses, and supporting precision crop management. Although deep learning (DL) models have demonstrated remarkable classification performance, their reliability under real-world field conditions remains insufficiently explored. In particular, confidence estimates often fail to reflect true predictive correctness when models are exposed to distributional shifts, limiting their practical applicability. To address this challenge, this study proposes a confidence- and uncertainty-aware DL framework for grape leaf disease diagnosis that extends evaluation beyond conventional accuracy-based metrics. Two publicly available grape leaf datasets were employed for stratified five-fold cross-validation in binary and multiclass classification tasks, while a third independently collected dataset was reserved exclusively for leakage-free external validation. EfficientNet-B0 and MobileNetV3-Large were evaluated under four inference configurations: raw inference, temperature scaling (TempScaling), Monte Carlo dropout, and an ensemble strategy combining uncertainty estimation with calibration. External validation demonstrated that both architectures maintained discriminative capability under domain shift. Under ensemble inference, EfficientNet-B0 achieved an accuracy of 73.8%, a macro-F1 score of 73.3%, a Matthews correlation coefficient (MCC) of 0.468, and a receiver operating characteristic area under the curve (ROC-AUC) of 0.804, while MobileNetV3-Large achieved 71.8% accuracy, 71.2% macro-F1, an MCC of 0.429, and a ROC-AUC of 0.787. Despite these promising results, raw predictions exhibited substantial overconfidence, with Expected Calibration Error (ECE) values of 0.287 and 0.312 for EfficientNet-B0 and MobileNetV3-Large, respectively. TempScaling markedly improved calibration quality, reducing ECE to 0.038 and 0.042 without affecting classification performance. Ensemble inference further enhanced the balance between predictive discrimination and reliability. The results demonstrate that strong classification performance alone is insufficient for trustworthy deployment in agricultural environments. Confidence calibration and uncertainty quantification provide complementary information for identifying overconfident predictions and improving decision reliability under field variability. The proposed framework offers a reliability-oriented approach for developing trustworthy artificial intelligence systems for grape leaf disease diagnosis in precision agriculture.