MSCPNet: A Multi-Scale Convolutional Pooling Network for Maize Disease Classification

Maize (Zea mays) is a critical crop for global food security and economic stability. However, it is highly vulnerable to various diseases such as northern leaf blight, common rust, and maize lethal necrosis, which can lead to significant crop losses if not detected early. Traditional CNN-based models, while effective in extracting spatial features, often fail to capture subtle multi-scale variations necessary for distinguishing between disease symptoms. These models also suffer from high computational complexity when deeper layers are introduced to handle fine-grained details. Transformer-based models, on the other hand, provide long-range dependencies but come with significant computational overhead, limiting their use in real-time agricultural applications. To overcome these challenges, we propose MSCPNet, a novel architecture that combines a truncated MobileNetV2 backbone with a Multi-Scale Convolutional PoolFormer block. The truncated backbone ensures that only essential layers for general feature extraction are retained, enhancing the model's adaptability across domains. The Multi-Scale Convolutional PoolFormer block captures both local and global dependencies through parallel convolutional branches of varying kernel sizes, while the PoolFormer module efficiently handles feature aggregation without the heavy computational cost associated with traditional attention mechanisms. This design allows the model to balance computational efficiency and high accuracy, making it highly suitable for real-time maize disease detection. Extensive evaluations on the maize leaf disease classification task yielded outstanding results, with the proposed MSCPNet achieving an accuracy of 97.44%, precision of 96.76%, recall of 97.37%, F1-score of 97.04%, and an MCC of 0.9653, with a model size of 998,084 parameters and 315,258,752 FLOPs. Furthermore, the model was evaluated on the PlantVillage dataset for tomato leaf disease classification, where it achieved an accuracy of 99.32%, precision of 99.32%, recall of 99.33%, F1-score of 99.32%, and an MCC of 0.9925. These results demonstrate the effectiveness and efficiency of MSCPNet in disease classification across different domains. © 2013 IEEE.