Title: Leveraging Transfer Learning and Attention Mechanisms to Enhance Peach Leaf Damage Classification Amidst Domain Shift

Abstract

Artificial intelligence offers a viable framework for evaluating crop damage using imagery, thereby facilitating proactive decision-making in agricultural management. In peach orchards, the escalating impacts of climate change have intensified both abiotic stressors and biotic threats, such as diseases and pests. These factors frequently generate foliar symptoms that are visually indistinguishable, complicating manual diagnosis. This challenge is particularly pronounced when assessing fields with diverse environmental conditions, underscoring the necessity for automated systems capable of strong generalization.

To address this, we developed an image-based classification method for detecting damage on peach leaves. We constructed a benchmark dataset by manually annotating publicly available images, comprising 1,366 leaves categorized into six distinct damage types. We assessed the performance of various deep learning architectures. The EfficientNet family delivered superior results: EfficientNetB0 attained 92.9 percent accuracy, while EfficientNetB3 reached 91.5 percent. Notably, EfficientNetB5 demonstrated the most effective handling of minority classes. Additionally, DenseNet121 achieved an accuracy of 92.6 percent.

We further investigated the impact of integrating the Convolutional Block Attention Module (CBAM). This integration enhanced performance for several backbone networks, most notably EfficientNetB5 and InceptionV3, though it yielded minimal or detrimental effects on others. The combination of CBAM with EfficientNetB5 resulted in the highest overall accuracy of 93.3 percent.

To test robustness against realistic domain shifts, we gathered a local dataset consisting of 180 images across four categories and applied transfer learning techniques. We experimented with three distinct fine-tuning strategies. Among these, EfficientNetB3 paired with CBAM performed best within the local domain, achieving a macro F1-score of 93 percent post-transfer. Ultimately, our findings indicate that models incorporating attention mechanisms exhibit greater resilience for minority classes and offer superior generalization capabilities across varying field environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC