ST-CFI: Swin Transformer with convolutional feature interactions for identifying plant diseases

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Global food security is threatened by population growth and limited arable land.
• Early and accurate plant disease identification is crucial for minimizing crop losses and boosting agricultural output.

Purpose of the Study:

• To introduce the Swin Transformer with Convolutional Feature Interactions (ST-CFI), a deep learning framework for plant disease detection.
• To evaluate the ST-CFI model's performance across diverse plant disease datasets.

Main Methods:

• The ST-CFI model integrates Convolutional Neural Networks (CNNs) and Swin Transformers for local and global feature extraction.
• An inception architecture and cross-channel feature learning are employed for enhanced feature extraction.
• The model was tested on five datasets: PlantVillage, Plant Pathology 2021, PlantDoc, AI2018, and iBean.

Main Results:

• The ST-CFI model achieved high accuracy, including 99.96% on PlantVillage, 99.22% on iBean, 86.89% on AI2018, and 77.54% on PlantDoc.
• The model demonstrated robustness and generalization capabilities across different datasets.
• High accuracy and F1 scores, with low loss values, confirm the model's effectiveness in learning discriminative features.

Conclusions:

• The ST-CFI model represents a significant advancement in early and accurate plant disease detection.
• This framework is a valuable tool for precision agriculture, contributing to agricultural sustainability and productivity.
• The integration of CNNs and Transformers enhances feature extraction for improved plant disease identification.