DLCPD-25: A Large-Scale and Diverse Dataset for Crop Disease and Pest Recognition

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate crop pest and disease identification is vital for global food security.
• Existing deep learning datasets lack the scale, diversity, and real-world complexity needed for robust model development.
• Current datasets often use controlled conditions, failing to represent natural environments.

Purpose of the Study:

• Introduce DLCPD-25, a novel, large-scale, and diverse benchmark dataset for training deep learning models.
• Address the limitations of existing datasets by incorporating realistic environmental conditions and class distributions.
• Facilitate the development of generalizable agricultural diagnostic systems.

Main Methods:

• Constructed DLCPD-25 by integrating 221,943 images across 23 crop types and 203 pest/disease/healthy classes.
• Included images from diverse online sources and extensive field collections.
• Utilized state-of-the-art self-supervised learning models (MAE, SimCLR v2, MoCo v3) for pre-training on the DLCPD-25 dataset.

Main Results:

• Pre-trained models on DLCPD-25 demonstrated strong performance in learned representations.
• Linear probing evaluation showed the SimCLR v2 framework achieved 72.1% top accuracy and 71.3% Macro F1 score.
• The dataset's realistic complexity, including uncontrolled field images and a natural long-tail distribution, proved effective.

Conclusions:

• DLCPD-25 is a valuable and challenging resource for training generalizable deep learning models.
• The dataset supports the creation of more robust and accurate agricultural diagnostic systems.
• This work advances the development of AI-powered solutions for crop protection and food security.