Attention mechanisms and transfer learning for robust peach leaf damage classification under domain shift

Researchers developed an AI-powered image classification system for detecting peach leaf damage using deep learning and attention mechanisms, achieving 93.3% accuracy on a benchmark dataset. The study demonstrates that EfficientNet models with attention modules provide robust generalization across different farming environments, addressing a critical need in automated agricultural disease diagnosis.

This research addresses a genuine agricultural challenge: climate change has intensified both abiotic stress and pest pressures on crops, making manual diagnosis increasingly difficult and time-consuming. The visual similarity of different foliar symptoms creates ambiguity even for experienced agronomists, particularly when conditions vary across multiple orchards. The study's contribution lies not merely in high accuracy rates but in demonstrating practical robustness through domain shift evaluation—testing models on locally-collected data that differs from the training distribution.

The architectural findings reveal that EfficientNet models outperformed alternatives, with the addition of Convolutional Block Attention Modules (CBAM) providing measurable improvements in minority class performance. This attention mechanism innovation matters because real-world agricultural datasets often feature imbalanced damage categories, making standard accuracy metrics misleading. The transfer learning results—achieving 93% macro F1-score on locally-collected images—validate that the approach generalizes beyond controlled datasets.

For the agricultural technology sector, this research supports the growing trend of deploying edge AI solutions on farms. Automated leaf damage detection systems could reduce input costs, improve pesticide targeting, and enable earlier intervention. The emphasis on robustness across field conditions addresses a persistent barrier to agricultural AI adoption: models trained on curated datasets frequently fail in production environments with variable lighting, camera angles, and disease presentations.

Future development should focus on expanding the local evaluation dataset and testing on diverse peach varieties and geographic regions. Integration with field robotics and decision-support systems would accelerate practical deployment, transforming reactive disease management into proactive crop health monitoring.

• →EfficientNetB5 with CBAM attention mechanisms achieved the best performance at 93.3% accuracy on benchmark peach leaf damage classification.
• →Transfer learning strategies successfully addressed domain shift, reaching 93% macro F1-score when tested on locally-collected field images.
• →Attention modules improved robustness for minority damage classes, addressing a critical challenge in imbalanced agricultural datasets.
• →The study demonstrates that deep learning models can generalize effectively across varying field conditions when properly designed and fine-tuned.
• →Automated leaf damage detection represents a practical tool for early pest and disease intervention in climate-stressed agricultural systems.