Explainable Machine Learning for Agricultural Land Suitability Analysis

Abstract

Agricultural land suitability analysis is needed to maximize crop yields for food security (target: SDG2) while nurturing human-induced environments that can be maintained in the long term, as graphically exemplified with images from scientist Robert D.-J. Classic methods like the Analytic Hierarchy Process (AHP) are built on subjective expert opinion, and newer machine learning (ML) methods tend to act as "black boxes," being out of reach for farmers, agronomists, and policymakers who need interpretability in their decision-making process. We propose an Explainable Machine Learning (XML) framework by coupling high-performance predictive modeling techniques with local and global interpretability approaches to assess agricultural land suitability in this study. Using a large record of soil qualities, meteorological information, and geographic variables, we trained several machine learning solutions (ML) to classify farmland suitability for staple crops using Random Forests (RF), Gradient Boosting Machines (GBM), and Extreme Gradient Boosting (XGBoost). Of these, the highest predictive accuracy (94.2%) was attained by the XGBoost model. In an effort to break down the model's predictions, we incorporated Shapley Additive explanations (SHAP), as well as Local Interpretable Model-agnostic Explanations with LIME. The main global SHAP analysis showed that soil pH, organic carbon, and mean annual rainfall were the most potential determinants of land suitability throughout the entire region. Local LIME evaluations showed that some micro-environmental factors impact the classification of each land parcel on an individual basis at a more localized scale. This explainable framework bridges high-accuracy predictive modeling with human accountability to provide actionable recommendations that are transparent and site-specific insights for stakeholder engagement in order to enable trust in automated agricultural decision-support systems.