An Attention-Enhanced MobileNetV2 with Squeeze-and-Excitation Architecture for Efficient Potato Leaf Disease Detection and Classification

Abstract

Abstract Potatoes are one of the major crops eaten in developing countries; however, their production is falling due to various diseases. Early identification and detection of potato leaf diseases play a vital role in improving potato quality and quantity. Existing methods are either computationally resource intensive or lack trust in their decision-making process, which makes them difficult to deploy for real-time potato disease classification and limits its accessibility. To mitigate these limitations, this study proposed an attention-enhanced MobileNetV2 with a squeeze-and-excitation architecture, which balances high accuracy with low computational resources. This method incorporates the strength of MobileNetV2 and Squeeze-and-excitation networks. A total of 2152 images of early blight, late blight, and healthy leafs were obtained from the Kaggle public repository, which are partitioned into 70% training, 20% validation, and 10% testing and were utilized to train, validate, and test the proposed model. The MobileNetV2 backbone is utilized for feature extraction, and then a squeeze-and-attention block is used to recalibrate the feature maps by focusing on important features and suppressing irrelevant ones. Gradient-weighted Class Activation Mapping (Grad-CAM) was implemented to visualize the most relevant region of the leaf for decision-making, which increases model interpretability and user trust. The proposed model achieves a remarkable performance of 99% testing accuracy with 9.41 MB total parameters. The proposed model is suitable for real-time potato leaf disease detection and classification, which can be easily accessible to agricultural stakeholders, including farmers, and contributes to food security.

Code and data availability

The paper's phenotyping inputs are 2152 potato leaf images (early blight, late blight, healthy) obtained from a public Kaggle repository, explicitly stated as publicly available in the Declarations. No author analysis code is shared (Code Availability: Not applicable), and no trained model checkpoints are released.