Hybrid Multi-Scale Neural Network with Attention-Based Fusion for Fruit Crop Disease Identification.

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Unobserved fruit crop diseases threaten global agricultural productivity and cause significant financial losses.
• Manual disease detection is inefficient, unreliable, and unsuitable for large-scale monitoring.
• Current deep learning models struggle with generalization, varying disease sizes, and edge device deployment.

Purpose of the Study:

• To develop a precise and effective fruit crop disease identification architecture.
• To overcome the limitations of existing deep learning models in plant disease detection.
• To enhance model interpretability and classification performance for real-world applications.

Main Methods:

• Proposed a Hybrid Multi-Scale Neural Network (HMCT-AF with GSAF) architecture.
• Combined a Vision Transformer with multi-scale convolutional branches for feature extraction.
• Utilized a novel HMCT-AF with GSAF module for adaptive feature combination.

Main Results:

• Achieved up to 93.79% accuracy on PlantVillage and CLD datasets.
• Demonstrated strong resilience to variations in lighting and background complexity.
• Outperformed vanilla Transformer models, EfficientNet, and traditional Convolutional Neural Networks (CNNs).

Conclusions:

• The HMCT-AF with GSAF model effectively captures scale-variant disease symptoms.
• The architecture is suitable for real-time agricultural applications on edge-compatible hardware.
• Presents a viable foundation for intelligent, scalable plant disease monitoring in precision farming.