綿の葉の病気を解釈可能かつ正確に分類するための注意機構強化型CNNアンサンブル

Precise and timely identification of cotton leaf diseases is essential for sustaining crop yield and quality, yet manual inspection remains time-consuming, labor-intensive, and prone to error. Existing automated approaches are limited by insufficient dataset diversity, inconsistent evaluation practices, limited use of explainable AI (XAI), and high computational cost. To address these challenges, we propose an attention-enhanced CNN ensemble, namely CottonLeafNet, which integrates lightweight convolutional neural networks for accurate cotton leaf disease classification across two publicly available datasets. CottonLeafNet achieves state-of-the-art performance, obtaining 98.33% accuracy, a macro F1-score of 0.9833, Cohen's kappa of 0.9800, a mean PPV of 0.9838, and an NPV of 0.9967 on Dataset D1, with an inference time of 0.51 s per image. On Dataset D2, it reaches 99.43% accuracy, a macro F1-score of 0.9942, Cohen's kappa of 0.9924, a mean PPV of 0.9943, and an NPV of 0.9981, with a 0.40 s inference time. Moreover, a unified eight-class dataset created by merging both datasets yields a test accuracy of 99.08%. Robustness analysis under artificially induced class imbalance further confirms the model's stability, with consistently strong macro F1-scores. To evaluate the generalization capability of the proposed CottonLeafNet, we conducted cross-dataset experiments, and the results indicate that the model maintains moderate performance even when trained and tested on different datasets. Gradient-Weighted Class Activation Mapping (Grad-CAM) visualizations demonstrate that CottonLeafNet reliably attends to disease-relevant regions, enhancing interpretability. Finally, real-time feasibility is validated through a web-based deployment achieving ≈1 s inference per image. These results establish CottonLeafNet as an accurate, robust, interpretable, and computationally efficient solution for automated cotton leaf disease diagnosis.