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

Technical Summary: Attention Mechanisms and Transfer Learning for Robust Peach Leaf Damage Classification under Domain Shift

Problem Statement
The paper addresses the challenge of automated peach leaf damage classification in precision agriculture, specifically focusing on the issue of domain shift. While Convolutional Neural Networks (CNNs) have demonstrated strong performance on curated public datasets, they suffer significant performance degradation when applied to images captured under different field conditions (e.g., variations in illumination, camera setup, and background complexity). This problem is particularly acute for perennial fruit crops like peaches, where existing datasets are small, heterogeneous, and rarely validated against locally collected field imagery. The study highlights that manual diagnosis is challenging due to visually similar foliar symptoms caused by abiotic and biotic stresses, necessitating automated models capable of generalizing across heterogeneous environments.

Methodology
The authors propose an image-based classification methodology involving two distinct experimental phases:

1. Benchmarking on Public Data:

   • Dataset Construction: A benchmark dataset was created by manually annotating publicly available images from PlantDoc and a dataset by Farman et al. The final dataset comprises 1,366 peach leaves categorized into six classes: bacterial spot, abiotic stress, mechanical stress, mite presence, chewing insect damage, and healthy leaves. The dataset is notably imbalanced, with the "healthy" class dominating.
   • Model Evaluation: Eleven CNN architectures were evaluated, including MobileNetV2/V3, EfficientNet (B0, B3, B5), ResNet (50, 101), VGG (16, 19), InceptionV3, and DenseNet121.
   • Attention Integration: The Convolutional Block Attention Module (CBAM) was integrated into selected backbones (EfficientNetB3, InceptionV3, EfficientNetB5) to enhance feature discrimination. CBAM was attached to the output of the last convolutional stage before global average pooling.
   • Training Strategy: Models were trained using a two-stage approach: feature extraction with frozen backbones followed by partial fine-tuning (unfreeze ratio of 0.3). Class-weighted categorical cross-entropy loss and data augmentation were employed to mitigate class imbalance.

2. Domain Adaptation via Transfer Learning:

   • Local Dataset: A new local dataset of 180 images was collected from a commercial peach orchard in Jumilla, Spain. Due to seasonal constraints, this dataset contained only four of the six original classes: abiotic stress, mechanical stress, healthy, and chewing insect damage.
   • Fine-Tuning Strategies: Three transfer learning strategies were assessed to address the domain shift between the public source domain and the local target domain:
     1. Feature Extraction: Training only the classification head.
     2. Partial Fine-Tuning: Fine-tuning the last layers.
     3. Full Fine-Tuning: Fine-tuning the entire network.

Key Contributions

1. A rigorous performance benchmark of eleven CNN architectures on a manually annotated public dataset of 1,366 peach leaf images across six classes.
2. A systematic evaluation of CBAM integration, assessing its impact on minority class detection and overall classification performance.
3. The creation and public release (on Zenodo) of a novel local field dataset of 180 images, used as a target domain for evaluating cross-domain generalization.
4. A comparative analysis of three transfer learning fine-tuning strategies to mitigate domain shift.

Results

• Public Dataset Performance: The EfficientNet family achieved the best overall performance. EfficientNetB0 reached 92.9% accuracy, while EfficientNetB5 provided the best minority-class detection. DenseNet121 also performed strongly with 92.6% accuracy.
• Impact of CBAM: The integration of CBAM improved performance in several backbones, particularly EfficientNetB5 and InceptionV3. EfficientNetB5 + CBAM emerged as the top-performing model on the public dataset with 93.3% accuracy. CBAM specifically enhanced recall for minority classes, such as mite presence, though benefits were architecture-dependent.
• Domain Shift Adaptation: On the local dataset, feature extraction consistently produced the weakest results. The optimal fine-tuning strategy varied by architecture.
  • EfficientNetB3 + CBAM was identified as the best architecture for handling domain shift, achieving a macro F1-score of 93% and 94.6% accuracy on the local domain after full fine-tuning.
  • MobileNetV3Large and DenseNet121 also showed significant improvements after full fine-tuning, with MobileNetV3Large increasing its F1-score by over 135%.
  • The study found that while CBAM generally improved adaptation efficiency, its effect was not universally positive; for instance, InceptionV3 + CBAM suffered performance degradation under full fine-tuning.

Significance and Claims
The paper claims that attention mechanisms, specifically CBAM, improve robustness on minority classes and enhance generalization across varying field conditions. It demonstrates that performance on large-scale pretraining benchmarks does not directly translate to robustness in target domains, emphasizing the need for systematic evaluation of fine-tuning strategies. The authors conclude that EfficientNet-based architectures, particularly when enhanced with CBAM and subjected to appropriate transfer learning strategies, offer a robust solution for supervised domain adaptation in agricultural disease classification under low-data regimes. The work highlights that model capacity alone does not guarantee better transferability, and that the interaction between attention mechanisms and backbone architectures must be carefully evaluated.