EgyPLI: A Real-life Annotated Image Dataset for Egyptian Plant Leaf Identification

Area of Science:

• Computer Vision
• Agricultural Science
• Machine Learning

Background:

• Lack of geographically diverse datasets hinders generalized plant identification models.
• Existing datasets often lack real-world variability, limiting model robustness.
• Egypt's agricultural sector needs localized, representative data for automated systems.

Purpose of the Study:

• Introduce the Egyptian Plant Leaf Image Dataset (EgyPLI) as a novel resource.
• Provide a geographically diverse and realistic dataset for plant identification research.
• Facilitate the development of robust deep learning models for plant classification and disease diagnosis.

Main Methods:

• Curated a dataset of 3,588 real-world leaf images from eight plant species.
• Captured images under varying conditions (viewpoints, lighting, clutter) reflecting natural environments.
• Annotated and preprocessed images, including healthy and diseased leaves, for consistent standards.

Main Results:

• Evaluated dataset performance using ResNet50, VGG16, and a custom CNN.
• Achieved high accuracies: 61.67% (ResNet50), 96.81% (VGG16), and 99.22% (custom CNN).
• Demonstrated the dataset's suitability for training effective plant identification models.

Conclusions:

• EgyPLI addresses a critical gap in plant leaf image datasets, particularly for Egypt.
• The dataset's realism supports training robust models for practical deployment.
• EgyPLI enables advancements in automated plant classification, diagnosis, and health assessment.