FusionDLNet: A Feature Fusion-Based CNN Approach for Citrus Leaf Disease Identification

Abstract

Translate Article

In this study, we introduce an innovative Fusion-Based Deep Learning Network (FusionDLNet) architecture designed specifically for detecting diseases in citrus plants. This network utilizes two parallel convolutional pathways, each featuring different kernel sizes to enhance feature extraction. The first pathway uses 3×3 convolutional filters, while the second uses 5×5 filters, allowing the model to capture both detailed and broader spatial features. Before putting images into the model, they undergo several pre-processing steps including rescaling, and normalization. These techniques aim to boost the generalization capabilities of the model. Subsequently, features extracted from both pathways are merged and directed through a fully connected layer that uses a softmax activation function for classification. The FusionDLNet model has been evaluated on citrus datasets and Sweet Orange leaf datasets and has shown superior performance against ResNet-50, VGG16, and VGG19. It reflects approximately 21%, 18%, and 23% improvement in accuracy metrics on Citrus leaves dataset and 17%, 13%, and 10% improvement in accuracy metrics on the Sweet Orange leaf dataset. The comparative results showcase the validity of the proposed architecture towards improving representation and classification of features for detecting diseases in plants. This proposed model presents a valuable tool for smart agricultural practices by facilitating identification of diseases that can lead to improved crop management and improved yield optimization