Benchmarking Deep Learning Architectures for Multi-Class Lemon Leaf Disease Classification

Abstract

Accurate and early detection of citrus leaf diseases is essential for maintaining lemon crop health and yield. This work presents a comparative evaluation of four deep learning architectures — CNN, CNN+LSTM, ResNet50, and DenseNet121 — for multi-class classification of lemon leaf diseases using image-based data. A standardized image preprocessing pipeline was developed, including resizing, normalization, and data augmentation techniques to improve model generalization. Each model was trained and evaluated under identical conditions to ensure fair benchmarking. Experimental results demonstrated that ResNet50 outperformed all other models, achieving the highest accuracy of 90.59%, precision of 91.73%, and F1-score of 89.73%, owing to its residual learning and deep feature extraction capability. DenseNet121 followed with 85.15% accuracy, while the lightweight CNN model offered a competitive 80.20% accuracy despite its simplicity. In contrast, CNN+LSTM underperformed with only 55.45% accuracy, indicating that temporal layers were ineffective for static image classification tasks. These findings highlight the effectiveness of deep convolutional backbones — particularly residual and densely connected networks — in extracting discriminative features for plant disease detection. The study lays a robust foundation for developing hybrid or ensemble models that further enhance diagnostic accuracy in precision agriculture.