Addressing Imbalanced Dataset for Plant Disease Classification by Improving Average Recall Using Targeted-Augmentation and a Novel Recall–Precision Monitor Thawing-Guided Training

Bourzig, Dounia Kawther Dihya, Abed, Mansour, Merah, Mostefa
Journal of Crop Health · Springer Nature · 2026-11-17 · DOI: 10.1007/s10343-026-01482-z

Аннотация

Plant disease classification faces many challenges, including imbalanced datasets in wich diseased and healthy samples occur in uneven proportions. For Convolutional Neural Network (CNN) deployment, conventional metrics can overlook minority-class performance, making average recall (AR) a key measure of the model’s ability to detect underrepresented samples . The objective of this paper is to improve AR in imbalanced scenario. To address these challenges, we propose a combined approach that integrates data-level augmentation with model-level fine-tuning for binary classification. Augmentation was applied exclusively to the minority class in the training and validation sets, increasing its representation from a highly imbalanced ratio to a moderately imbalanced one while preserving realistic class distributions rather than enforcing a perfectly balanced dataset. The evaluation set was kept in its original highly imbalanced state to provide a realistic assessment of model performance under practical deployment conditions. At the model level, we introduce Recall–Precision Monitor Thawing (RPMT), a novel fine-tuning strategy that jointly leverages validation recall and precision to adaptively guide layer unfreezing and learning rate adjustment for CNNs. We evaluate the proposed RPMT under three thawing logics against two conventional thawing methods using two lightweight CNN models on a highly imbalanced cassava dataset. Results show that RPMT significantly improves AR while maintaining comparable performance across all metrics. For MobileNetV3Small, RPMT Logic‑2 achieved an AR of 0.8516, representing relative improvements of 15.3% and 34.6% over conventional thawing methods. For EfficientNetV2B0, RPMT Logic‑3 reached an AR of 0.8494, corresponding to relative improvements of 12.9% and 21.4%.

Читать у издателя →

Похожие статьи