Counterfactual Based Approaches for Feature Attributions of Stress Factors Affecting Rice Yield

Abstract

Agriculture is a crucial sector in many countries, particularly in India, where it significantly influences the economy, food supply, and rural livelihoods. The increased integration of Deep Learning (DL) and Machine Learning (ML) into agriculture has enabled substantial advancements in predicting crop yields and analyzing factors affecting them. The counterfactual reasoning framework of DICE outperforms LIME and DICE in offering finer insights into feature importance and the relative impact of different factors on yield prediction. DICE provided the clearest causal insights, demonstrating how adjustments to attributes like sandy alfisols and surface texture could lead to significant changes in crop yield by affecting water retention and nutrient availability. SHAP ranked features like phosphate and potash based on their average importance across the dataset, offering a global view of influential factors but lacking in-depth causal understanding. LIME provided localized insights on immediate influences, such as average rainfall and nitrogen content, although it fell short in revealing broader causal interactions essential for targeted agricultural interventions. The findings highlight the significance of counterfactual explanations in agricultural ML models, as they provide a robust understanding of causal relationships, going beyond correlation-based attributions. The study provides understandable and practical insights, allowing for focused actions to enhance productivity and adaptability in agriculture. By improving the interpretability of agricultural machine learning models, the research ultimately supports the creation of predictive systems that strengthen sustainable practices and economic development within the agricultural industry.