HybridLeafNet: A Multi-Scale Deep CNN Framework for Automated Early Fungal Disease Detection in Crop Leaves

Fungal diseases in crops represent a severe agricultural threat requiring fast and accurate automated diagnosis. This study presents an enhanced deep learning system that employs a Hybrid Convolutional Neural Network (Hybrid CNN) for classifying crop leaf images as Healthy or Infected, implemented in MATLAB. Unlike single-branch CNN architectures, the proposed HybridLeafNet combines two parallel convolutional branches operating simultaneously with different receptive field sizes — a 3×3 branch capturing fine-grained local texture patterns and a 5×5 branch capturing broader spatial disease features — whose outputs are concatenated and processed through deeper fully connected layers. The system retains the original preprocessing pipeline including resizing to 256×256, per-channel median filtering, and contrast enhancement. Training is conducted using the Adam optimizer over 100 epochs with shuffling at every epoch to prevent overfitting and improve generalization. The model additionally integrates audio feedback, playing a distinct audio cue upon classification to support non-visual user interaction. Performance evaluation employs a comprehensive set of metrics including accuracy, precision, recall, F1-score, specificity, and AUC derived from the confusion matrix. The multi-scale feature fusion strategy of HybridLeafNet significantly improves discriminative capability over single-path CNNs, yielding higher accuracy in distinguishing subtle fungal infection patterns across diverse crop leaf images.