Explainable Machine Learning Approach Quantified the Long-Term (1981–2015) Impact of Climate and Soil Properties on Yields of Major Agricultural Crops Across CONUS

Abstract

A comprehensive understanding of the long-term data on the crop, soils, environment, climate, and production management would facilitate efficient data-driven decision-making in agriculture production under changing climate. We have employed an explainable machine learning algorithm (random forest model coupled with LIME; Local Interpretable Model-Agnostic Explanations framework) using multi-decadal (1981–2015) data on climate variables, soil properties, and yield of major crops across the Coterminous United States (CONUS). This data-driven approach explained the multi-faceted factors of crop production for corn, soybean, cotton, and wheat under field conditions by leveraging agricultural informatics. We attempted to show how crop yields can better be correlated and explained when production input varies along with changing climatic/environmental and edaphic conditions. Our findings suggest Growing Degree Days (GDDs) as important climatic factors, while water holding capacity is one of the dominant soil properties in interpreting crop yield variability. Our findings will facilitate growers, crop production scientists, land management specialists, stakeholders, and policy makers in their future decision-making processes related to sustainable and long-term soil, water, and crop management practices.