Quantifying and Mitigating Domain Shift in Peach Leaf Damage Classification: Attention Mechanisms and Fine-Tuning Strategies

Deep learning provides a practical framework for crop damage assessment from imagery, supporting early decision-making in agricultural management. The gist: CBAM-enhanced EfficientNetB5 achieved the best overall performance with an F1 score of 0.936, demonstrating that attention mechanisms improve robustness on minority classes and enhance generalization across varying field conditions under domain shift.

Problem Context and Motivation

Automated image-based classification of plant damage is a key application of computer vision in precision agriculture, offering a scalable and objective alternative to manual field inspection. A persistent challenge is that models trained on curated public datasets suffer significant performance degradation when applied to images captured under different field conditions, known as domain shift. This problem is particularly acute in perennial fruit crops like peaches, where existing datasets are small, often heterogeneous in composition, and rarely validated against locally collected field imagery. Peach leaf damage classification involves six visually similar damage categories—including bacterial spot, abiotic stress, mite presence, mechanical damage, chewing insect damage, and healthy leaves—with severe class imbalance and limited publicly available data.

Benchmark Dataset Construction

A benchmark dataset was constructed through manual annotation of publicly available images comprising 1,366 peach leaves categorized into six damage types. The public datasets used included the PlantDoc dataset (containing 103 images for peach leaves) and a second dataset from Mendeley Data containing 1,033 images of peach fruits, leaves, and stems captured under diverse lighting and environmental scenarios. After selection and manual annotation using the CVAT image annotation tool, 1,366 individual leaves were annotated and labeled. The final class distribution was unbalanced, with the 'healthy' class dominating at 951 images compared to minority classes like 'chewing insect' (45) and 'bacterial spot' (144).

Model Architectures and Attention Mechanisms

Multiple convolutional neural network architectures were evaluated, including EfficientNet family models (B0, B3, B5), DenseNet121, InceptionV3, ResNet50/101, and VGG family models. The paper systematically evaluated the Convolutional Block Attention Module (CBAM) integrated into selected backbones to assess its impact on minority class detection and overall classification performance. CBAM extends the Squeeze-and-Excitation module by additionally incorporating spatial attention, enabling the model to focus on informative regions both channelwise and spatially. The integration of CBAM improved several backbones, particularly EfficientNetB5 and InceptionV3, with the CBAM-enhanced EfficientNetB5 emerging as the top-performing model with 93.3% accuracy.

Transfer Learning Strategies for Domain Adaptation

To evaluate adaptability under realistic conditions, a local dataset of 180 images across four classes was collected in a commercial orchard in Jumilla, Murcia, Spain (the target domain). Three fine-tuning approaches were assessed to address domain shift:

1. Feature extraction: Training only the head while freezing the backbone.

2. Partial fine-tuning: Fine-tuning the last layers of each model.

3. Full fine-tuning: Fine-tuning the entire model in order to explore whether modifying weights leads to achieving better performance in the new domain.

The results showed that Feature extraction consistently produces the weakest results across nearly all evaluated models, confirming that freezing the pretrained backbone is insufficient for this transfer learning task. Furthermore, FTA emerges as the most effective strategy for the majority of highperforming architectures, with EfficientNetB3 + CBAM achieving an F1-macro score of approximately 0.93 together with a remarkable performance improvement (∆F1 = 0.368).

Experimental Setup and Evaluation Metrics

Models were trained under a unified setup using the Adam optimizer, employing a two-stage training process: Stage 1 involved training for 100 epochs with the backbone frozen, incorporating class weights to address class imbalance. Stage 2 continued for an additional 120 epochs with the backbone partially unfrozen according to an unfreeze ratio of 0.3 and a starting learning rate of 1 × 10−5. The loss function used was the class-weighted categorical cross-entropy loss (L), designed to mitigate the impact of class imbalance by assigning higher importance to minority classes, using the formula:

L = − (1/N) Σ X i Σ C c w c y i,c log(ˆy i,c)

Evaluation metrics considered were Accuracy, F1-score (weighted and macro-averaged), Precision (macro), Recall (macro), and Macro F1. The paper emphasized the importance of macro variants to account for class imbalance.

Key Findings on Attention and Transfer Learning

The results demonstrated that "CBAM consistently improves performance on Mite Presence, one of the most challenging minority classes, primarily by increasing recall in InceptionV3 and EfficientNetB5.

As a fastidious and diligent researcher, I have analyzed this paper's methodology, findings, and future directions. The core contribution lies in developing a robust framework for peach leaf damage classification under challenging domain shift conditions by combining advanced attention mechanisms (CBAM) with systematic transfer learning fine-tuning strategies.

Here are the specific improvements that can be made to AI systems based on this research:

1. The system should incorporate a multi-stage decision pipeline integrating both feature enhancement and domain adaptation.

2. The model architecture should be dynamically selected based on the complexity of the input data and resource constraints (e.g., using MobileNet for edge deployment, EfficientNetB5 for high-accuracy cloud processing).

3. The system must utilize a class-weighted categorical cross-entropy loss function during training to explicitly penalize misclassification of minority classes (like Mite Presence or Chewing Insects), ensuring the model learns discriminative features rather than simply overfitting to the majority Healthy class.

4. The feature extraction strategy should be replaced by a dynamic Transfer Learning Fine-Tuning (TL-FT) strategy, specifically favoring Fine-Tune All for deep architectures (like DenseNet121 or EfficientNetB5) and Fine-Tune Last for lighter models, to optimally balance the extraction of generalized features from ImageNet with adaptation to local field conditions.

5. The system should dynamically adjust the attention module's configuration (CBAM) based on class performance; for instance, if the model struggles with a specific minority class (e.g., Mechanical Stress), the spatial and channel attention weights should be prioritized during training to focus on those specific visual cues.

6. The final deployed system can perform real-time diagnosis in diverse field conditions (varying illumination, background clutter) by leveraging the robust, domain-adapted weights learned from local data, significantly reducing the need for manual intervention.

This improved AI system can:

1. Identify and classify six distinct peach leaf damage types with high accuracy (up to 93.6% F1-macro score in testing).

2. Generalize its classification performance effectively across different field conditions and geographical locations (domain shift robustness).

3. Specifically detect challenging, visually similar minority classes such as Mite Presence and Chewing Insect damage, where baseline models often fail due to class imbalance.

4. Adapt quickly to new local agricultural environments with limited labeled data by leveraging the insights from transfer learning fine-tuning strategies, ensuring high performance on target domain imagery.

5. Provide actionable early detection for fruit and tree management, enabling timely intervention against pests and abiotic stresses, thereby reducing economic losses in peach orchards.

Deep learning models for crop damage assessment are typically trained and validated on curated public imagery, yet their behaviour when deployed in real orchards remains poorly quantified. This work measures and mitigates that gap for peach leaf damage classification, where climate-driven abiotic and biotic stresses produce visually similar foliar symptoms. A benchmark of 1366 manually annotated peach leaves covering six damage types was assembled from public sources, and a second, independently acquired dataset of 180 field images across four classes was collected in a commercial orchard as an unseen target domain. Eleven convolutional backbones and three attention-enhanced variants were compared; CBAM-EfficientNetB5 achieved the best source-domain performance (93.3% accuracy, 0.849 macro F1). Applied directly to the target domain, source-trained models lost on average 0.21 macro F1 points (26.5% relative), with 12 of 14 architectures degrading, confirming that benchmark performance substantially overestimates field behaviour. Three fine-tuning strategies were then evaluated as mitigation: feature extraction proved insufficient in nearly all cases, whereas full fine-tuning recovered performance, with CBAM-EfficientNetB3 reaching 0.9459 accuracy and 0.9297 macro F1 on the local domain. Attention mechanisms improved minority-class recall and adaptation efficiency, but did not by themselves confer robustness to domain shift. The results establish a transferability baseline for peach leaf diagnosis and quantify the adaptation cost of moving from public benchmarks to operational orchards.