CropDeep: The Crop Vision Dataset for Deep-Learning-Based Classification and Detection in Precision Agriculture

Area of Science:

• Computer Vision
• Agricultural Science
• Machine Learning

Background:

• Precision agriculture requires advanced intelligence for economic potential and production efficiency.
• Deep learning applications in agriculture necessitate large, domain-specific crop vision datasets.
• Existing datasets often lack the diversity and complexity of real-world agricultural conditions.

Purpose of the Study:

• To introduce the CropDeep dataset for species classification and detection in agriculture.
• To establish a robust benchmark for deep learning models in challenging agricultural environments.
• To evaluate the performance of state-of-the-art deep learning models on crop-related tasks.

Main Methods:

• Collected 31,147 images across 31 crop classes in greenhouses using various cameras and equipment.
• Annotated over 49,000 instances, focusing on visually similar species and periodic changes.
• Conducted baseline experiments using deep learning classification and detection models.

Main Results:

• Deep learning classification achieved over 99% accuracy.
• Deep learning detection accuracy reached 92%, indicating dataset complexity.
• YOLOv3 demonstrated potential for agricultural detection tasks.

Conclusions:

• The CropDeep dataset provides a valuable resource for advancing deep learning in agriculture.
• Current deep learning models show high performance in crop classification but require improvement for detection.
• Further research and model development are needed to optimize deep learning for crop production and management.