Cross-Dataset Generalization of Bangladeshi Rice Leaf Disease Classifiers: Benchmark, Diagnosis, and Mitigation

Yet, little is known about how well deep learning models for Bangladeshi rice leaf disease recognition perform when tested across different datasets, leaving an important question unanswered. Datasets differ in camera type, background, class set, and annotation standards; thus, high in-dataset accuracy on one of them does not guarantee high performance on the others. How much does the performance degrade? Is acquisition related? Which techniques can mitigate the effect? We have normalized three Bangladesh-based datasets to a consistent class taxonomy, frozen the splits, and studied cross-dataset transfer performance via this models, MobileNetV2, EfficientNet-B0, and ResNet50 in six directions of ordered sharing classes. The in-dataset macro-F1 mean is 0.719, while the cross-dataset macro-F1 mean is 0.436 (mean gap 0.387). The strong data augmentation increases the cross-dataset macro-F1 mean from 0.436 to 0.503 (mean gap +0.070, Wilcoxon p≈0.0002 on 18 cells). Leave-one-dataset-out technique improves performance of the target model in only 3 out of 9 directions. Adaptive Batch Normalization improves results in only 5 out of 18 pairs. A ResNet50 bucket study attributes almost all the gains to the background randomization.