AI-Driven Validation of Digital Agriculture Models

Area of Science:

• Agricultural Science
• Computer Science
• Data Science

Background:

• Digital agriculture uses artificial intelligence (AI) for crop management, but model validation is challenging.
• Neural networks offer accurate crop health assessments but lack explainability, hindering farmer trust and validation.
• Ineffective AI models can lead to counterproductive farming decisions and wasted resources.

Purpose of the Study:

• To develop an explainable AI approach for validating digital agriculture models.
• To enhance the reliability of AI-driven crop management recommendations for farmers.
• To optimize the process of field spot-checking for model validation.

Main Methods:

• Trained random forests (AI) to mimic neural network models for crop management.
• Utilized the random forest as an explainable 'white box' to analyze neural network behavior.
• Developed methods to assess test set representativeness and identify optimal spot-checking locations.
• Applied the approach to soybean defoliation assessment data.

Main Results:

• The proposed method allows for knowledge extraction from neural networks.
• It enables assessment of test set quality and identification of representative field areas.
• The approach determines optimal locations and timing for farmer spot checks.
• Reduced the need for field spot checks by up to 94% in soybean defoliation analysis.

Conclusions:

• This novel AI approach enhances the explainability and validation of digital agriculture models.
• It empowers farmers to trust and effectively utilize AI recommendations through targeted validation.
• Significant reduction in spot-checking efforts leads to increased efficiency and cost savings in precision agriculture.