TeaLeafVision: An Explainable and Robust Deep Learning Framework for Tea Leaf Disease Classification

TeaLeafVision: Explainable and Robust Deep Learning for Tea Leaf Disease Classification

Overview of Objective and Scope

The paper "TeaLeafVision: An Explainable and Robust Deep Learning Framework for Tea Leaf Disease Classification" (2604.07182) addresses the challenge of automating tea leaf disease identification using deep convolutional neural networks (CNNs) and explainable AI (XAI) techniques. The central motivation is to mitigate the limitations of labor-intensive and error-prone manual disease assessment, promoting timely intervention in economically significant tea agriculture. The authors frame their approach around three axes: accuracy in disease classification, interpretability of model decisions, and robustness to input perturbations, targeting real-world application reliability.

Dataset Construction and Preprocessing Pipeline

The study utilizes the teaLeafBD dataset, comprising 5,278 high-resolution RGB images labeled into seven classes: six disease categories (brown blight, gray blight, green mirid, helopeltis, red spider, tea algal leaf spot) and healthy leaves. Imaging conditions reflect variable field settings, introducing realistic noise. Images are uniformly resized (224×224), normalized, and the dataset is split into training (70%), validation (20%), and test (10%) partitions. Data augmentation (flipping, rotation, zoom) and class oversampling are implemented to counteract class imbalance and enhance generalization.

CNN Architectures and Training Methodology

Three pre-trained CNNs are systematically evaluated: DenseNet201, MobileNetV2, and InceptionV3. Each model is fine-tuned with Adam optimizer and categorical cross-entropy loss, and early stopping is used to prevent overfitting. DenseNet201 leverages densely connected blocks to facilitate feature reuse and gradient propagation, while MobileNetV2 and InceptionV3 emphasize computational efficiency and multi-scale feature extraction, respectively.

Adversarial training, simulating variations in input through controlled perturbations, is incorporated during fine-tuning. This regularization aims to increase robustness to noise and model stability in field deployments.

Quantitative Results and Robustness Evaluation

DenseNet201 achieves a test accuracy of 99%, outperforming MobileNetV2 (94%) and InceptionV3 (92%) by a significant margin. Precision, recall, and F1-score metrics remain uniformly high across all disease classes for DenseNet201, with the model displaying minimal confusion between visually similar categories. MobileNetV2 and InceptionV3 show comparatively higher rates of misclassification in ambiguous cases, particularly within the green mirid and helopeltis categories.

Adversarial robustness is substantiated by evaluating performance under varying perturbation strengths ($\epsilon$). DenseNet201 retains validation accuracies above 97.8% across all levels and peaks at 99.15% for $\epsilon=0.12$, confirming resilience to mild adversarial attacks and confirming suitability for practical deployment on field-acquired images.

Model Interpretability: Grad-CAM and Occlusion Sensitivity

Addressing the black-box nature of deep CNNs, the authors employ Grad-CAM to generate class-discriminative localization maps. These visualizations consistently target biologically meaningful disease regions—including lesions, discolorations, and texture anomalies—aligning with domain expert expectations. Occlusion sensitivity analysis further corroborates this: systematic masking of disease-specific pixels precipitates large drops in prediction confidence, emphasizing the model’s reliance on pathological evidence rather than spurious background cues.

These XAI techniques provide actionable transparency for agronomists and end-users, mitigating skepticism regarding automated decisions in mission-critical agricultural scenarios.

System Prototyping and Real-World Integration

A prototype web application has been released, showcasing practical integration. The application enables users to upload tea leaf images and returns disease classification alongside Grad-CAM heatmaps, streamlining adoption by non-technical stakeholders such as field technicians and farmers. The interface operationalizes the research contributions, demonstrating feasibility for widescale precision agriculture deployments.

Comparative Analysis and Positioning with Related Work

DenseNet201, as configured in TeaLeafVision, surpasses the strongest prior results for automated tea disease recognition, notably exceeding the 97.3% accuracy of YOLOv7 [1], 90.16% of LeafNet [2], and the 91.3% achieved by ShuffleNet-based approaches [3]. The study distinguishes itself from prior works through comprehensive attention to robustness and explainability rather than focusing solely on raw accuracy metrics. Existing studies generally lack systematic adversarial evaluation and integration of interpretability frameworks, limiting their practical transferability for decision-critical agricultural settings.

Limitations and Future Research Trajectories

The principal limitations identified are the modest dataset scale and lack of independent external validation. Although adversarial testing demonstrates resistance to perturbations, further scrutiny is necessary under severe domain shifts (e.g., different imaging devices, locations, seasonal variations). Future efforts should prioritize multi-source dataset aggregation, deployment on edge devices for real-time inference, and investigation into few-shot or self-supervised paradigms for rapid adaptation to emergent pathogens.

Conclusion

TeaLeafVision (2604.07182) introduces a robust and explainable CNN-based framework, establishing a new state-of-the-art for tea leaf disease classification with a reported test accuracy of 99% using DenseNet201. The combination of adversarial training, Grad-CAM interpretability, and a production-ready prototype demonstrates strong translational promise in precision agriculture contexts. Future work extending data diversity and field validation is essential to fully realize the potential for broad-scale agricultural impact.