Leakage-Aware and Reliability-Centered Deep Learning for Bean Leaf Disease Classification Under Bangladeshi Field Conditions

Abstract— While automated, image-based recognition of bean leaf diseases has been shown to aid timely crop management, results from previous studies are hard to interpret in the presence of near duplicate images in training and test sets. The authors provide a leakage-aware assessment of deep learning models for bean leaf disease classification using a dataset of 1,201 images from four classes from the 2026 bean leaf disease field image dataset in Bangladesh. Duplicate images were found using a hash-based technique and perceptual-hash detection and duplicate groups were limited to a single subset through a group-aware partitioning procedure; a test on a naive random split verified that, without such safeguards, duplicated leakages would significantly overestimate the test performance. We fine-tune four different architectures (ResNet-50, EfficientNet-B0, ConvNeXt-Tiny, and ViT-B/16) with the same configuration and evaluate their performance on the leakage-safe test set. The model with the highest accuracy, Macro-F1, and balanced accuracy (1.0000) was chosen as the best one: ConvNeXt-Tiny. ViT-B/16 was significantly worse with accuracy 0.9000. The model selected achieved an accuracy of ≥ 0.9833 after eight synthetic image perturbations, low ECE (low Brier score, 0.0432/0.0060, respectively), and did not benefit from TTA. Since there were no misclassifications in the test set, uncertainty-based error detection was not evaluated, and the robustness seen was due to synthetic corruptions and not to validated field variability. Based on the results, using duplicate-aware dataset auditing and splitting the dataset into groups significantly influences the accuracy of reported performance in bean leaf disease recognition, and external validation with data collected independently in the field is necessary for a wider generalization of these results.
Keywords— Bean Leaf Disease Classification, Deep Learning, Data Leakage, Near-Duplicate Detection, Leakage-Safe Evaluation, ConvNeXt-Tiny, Test-Time Augmentation, Uncertainty Calibration.