Efficient Attention Guided Fusion Model for Plant Disease Classification

Abstract

Early detection of diseases in plants has been identified as a critical factor for ensuring the maintenance of productivity, preventing economic losses, and promoting sustainable agriculture. Traditional manual approaches for diagnosing diseases are time-consuming, subjective, and inappropriate for large-scale and real-time agriculture. In order to overcome the limitations of traditional approaches, the CNN–CBAM–MobileViTNet has been proposed, an efficient attention-guided network by the fusion of Convolutional Neural Networks (CNNs), Convolutional Block Attention Module (CBAM), and Mobile Vision Transformer (MobileViT) for plant diseases recognition.The CNN component is effective in capturing local visual patterns like lesions, discoloration, and texture. The CBAM component is effective in refining the feature representations by focusing on disease-related spatial areas and useful channels. The MobileViTNet branch is useful in capturing contextual relationships from the leaf areas through lightweight transformer blocks. The CNN–CBAM—MobileViTNet is tested on an enhanced dataset with 38 classes of plant diseases and health conditions, splitting data into 70% training, 15% validation, and 15% testing. Significantly, extensive experimental analysis reveals that the test accuracy is 99%, with high precision, recall, and [Formula: see text]1-score values. Training-validation curves show that the model converges stably with little overfitting, while ROC analysis shows high classwise discrimination ability of the model. Hence, the CNN–CBAM–MobileViTNet model is reliable and may be used for real-world applications in smart agriculture and automatic crop disease monitoring systems.

Plant phenotyping relevance

植物葉の病徴を画像から分類するCNN・注意機構・Transformer融合モデルを開発し、38クラスの病害・健全状態で性能評価しており、病害状態の表現型取得・推定が中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.