Toward Sustainability: Trade-Off Between Data Quality and Quantity in Crop Pest Recognition

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Intelligent plant protection relies on crop pest recognition using convolutional neural networks (CNNs).
• Current deep learning methods for CNNs require large datasets, leading to unsustainable practices.
• Unsustainable practices include high data collection costs, expensive hardware, and significant power consumption.

Purpose of the Study:

• To investigate the trade-off between data quality and quantity for sustainable crop pest recognition.
• To propose a novel method for selecting high-quality data in feature space.
• To demonstrate the effectiveness of data quality-driven selection over quantity-driven approaches.

Main Methods:

• Proposed an embedding range judgment (ERJ) method operating in the feature space.
• Conducted comparative experiments to evaluate the performance of ERJ against traditional methods.
• Analyzed the impact of data quality on recognition task outcomes.

Main Results:

• Selected high-quality data, though limited in quantity, achieved performance comparable to using all training data in specific recognition tasks.
• A small subset of high-quality data outperformed a large volume of lower-quality data significantly.
• The contrast between good and bad data selection was remarkable.

Conclusions:

• Data quality is a critical factor in achieving efficient and sustainable crop pest recognition.
• The proposed ERJ method provides a foundation for data information analysis in smart agriculture.
• Highlights the need for the research community to prioritize data quality in pattern recognition and AI development.