Leakage-Proof Reliability-Aware Deep Learning Framework for Rice Leaf Disease Diagnosis

Source abstract

Timely and accurate diagnosis of rice leaf diseases is significant in order to achieve reduced loss and precision agriculture management. In this paper, a leakage-aware and reliability-oriented deep learning framework is proposed to classify rice leaf diseases based on the Mendeley rice leaf disease dataset. In order to avoid the problem of similar and duplicate images in the training and validation set, the framework incorporates perceptual hashing and stratified group K-fold cross-validation. The evaluation measures the accuracy of the in-domain classification, robustness to controlled image corruption, evaluation of the Expected Calibration Error and temperature scaling methods, selective prediction using risk-coverage analysis, external-dataset validation, and computational efficiency profiling. The in-domain evaluation resulted in an almost perfect score. The accuracy achieved by the cross-validation method applied to the folds evaluated showed high values, while the external UCI rice leaf dataset presented significantly lower values, with an accuracy of 54.17% and macro-F1 of 48.98%. This contrast suggests that high in-domain accuracy should be considered with caution if there is a domain shift. In addition, efficiency and reliability analysis indicated that mid-scale convolutional models, especially ResNet-based models, achieved a better trade-off between predictive performance and computational cost compared to larger high-capacity models. The suggested assessment scheme in the proposed framework then focuses not only on accuracy but also on deployment-oriented evaluation in the context of rice leaf disease diagnosis with the help of AI.