A Transfer Learning-Based Hybrid Deep Learning Framework for Multi-Crop Plant Leaf Disease Classification

Abstract

Early and accurate detection of plant leaf diseases is crucial for sustainable crop management, as such diseases threaten global agricultural productivity and food security. This study introduces a hybrid deep learning framework based on transfer learning for the multi-class classification of plant leaf diseases in cassava, wheat, and tomato. The proposed architecture integrates DenseNet121 and a scaled MobileNetV2 (α = 0.35) to achieve a balance between representation power and computational efficiency. The framework incorporates attention mechanisms, compresses feature maps using 1 ×1 convolutions, and fuses them in a compact classification head using Swish activation and batch normalization. Gradient-weighted Class Activation Mapping++ (Grad-CAM++) is used for model interpretability, highlighting disease-relevant regions. Evaluated on three public datasets comprising 10,635 images across 13 disease classes and one healthy class, the model achieves perfect accuracy on cassava, approximately 95% accuracy on wheat, and up to 99.6% on tomato. Despite a compact design with ~10 million parameters, the model performs competitively and is suitable for deployment on edge devices, with inference latency under 60 ms and throughput above 20 Frame per Second (FPS).