Peanut yield and grade prediction in Georgia, USA: integrating management, climate, and remote sensing data with explainable AI

Predicting agricultural yields and quality is essential for optimizing decision-making, food security, economic planning, and resource management. Although many studies have compared algorithms or variables for predicting yield or quality, the information on peanut ( Arachis hypogaea L.) is limited. To address this, our study aims to: a) compare the performance of various machine learning (ML) models for predicting peanut yield and grade across diverse variable groups (GV); b) select the most accurate ML and GV combination; c) identify the most important factors driving outcomes using SHapley Additive exPlanations (SHAP); and d) assess the spatial generalization of the best models using a Leave-One-Site-Year-Out (LOSYO) cross-validation. A total of 540 ML were trained comparing 18 different ML with 15 GV, being these: management (M), weather (W), soil (S), and remote sensing (R) datasets and their combination. Management information was recovered from surveys performed from 2017 to 2019 on more than 200 peanut farms from Georgia, USA and open-source related data was retrieved. Results indicated that Cubist-rule and support vector machine performed better than other models achieving the lowest root mean squared error values (816 kg ha −1 for yield, 1.52 for grade). Among the different GV, M + S and M + R performed better than others for yield and grade, respectively. SHAP revealed irrigation, geographic location, and soil properties as key drivers of yield, whereas vegetation indices and management decisions influenced grade. The study underscores the value of interpretable ML models for optimizing peanut production under variable environmental conditions, offering actionable insights for farmers and agronomists.