A Novel Fusion Mechanism-Based Lightweight Efficient Convolutional Attention Network for Jute Pest Classification

Accurate and efficient identification of jute pests is essential to sustaining agricultural yield and preventing fibre-quality degradation in jute (Corchorus spp.) production. Traditional visual inspection is subjective, labour-intensive, and error-prone under variable field conditions, motivating the need for robust automated classification systems. This paper introduces a lightweight deep learning architecture, MLECNet, that integrates squeeze-and-excitation blocks with ViT and a convolutional attention module in a novel way. The proposed architecture is designed based on the information in the image, extracting features at multiple scales and initially fusing the feature maps using a depth-wise concatenation approach. The feature maps are extracted from the squeeze-and-excitation (SE) blocks, inspired by the EfficientNet architecture. The information is refined by the ViT encoder blocks attached to the initial SE blocks. The final information of these blocks is flattened and passed to the attention maps for the extraction of more refined features that classify the jute pests. Experimental evaluations were performed on a 17-class jute pest dataset comprising 17,000 augmented images. The hybrid network, containing 3.7 million learnable parameters and a model size of only 13.4 MB, achieved a classification accuracy of 92.30%, a precision of 92.40%, a recall of 92.30%, and an F1-score of 92.30% using a medium neural network classifier, outperforming individual backbones and prior state-of-the-art models such as DenseNet201, ResNet50, and VGG19. The proposed model also demonstrated superior generalisation and interpretability through comparative ablation studies, confirming the complementary advantages of convolutional, transformer, and attention-based architectures in a unified form. The results establish the proposed fusion-based framework as a computationally efficient, interpretable, and scalable solution for real-world jute pest detection, contributing toward the advancement of AI-driven precision agriculture and sustainable fibre crop management.