Machine learning reveals heterogeneous climate impacts and weakening agricultural productivity growth

Abstract

Climate change may influence agricultural productivity through nonlinear and regional effects of temperature, precipitation, and solar radiation. Current studies of agricultural total factor productivity (AgTFP) are either regional or reliant on econometric models, with limited evidence on whether machine-learning methods can better capture climate-productivity relationships on a global scale. This study developed a global machine-learning model to predict country-level agricultural total factor productivity (AgTFP) with a total of 110 climate predictors were constructed from growing-season conditions, three-month moving averages, and annual summaries of temperature, precipitation, and solar radiation. Out of the thirty-six regression methods tested, Random Forest provided the strongest global test performance (R² = 0.502, RMSE = 16.98). A Shapley Additive exPlanations (SHAP) analysis on the global model showed that minimum temperature had the greatest SHAP importance (28.7%), followed by solar radiation (23.8%). On a regional level, test R^2 ranged from 0.311 in Africa to 0.658 in Asia, while the dominant climate family differed across regions. Under MRI-ESM2.0 SSP2-4.5 projections for the years 2015–2100, 129 of 175 countries exhibited increasing AgTFP trends and 46 exhibited decreasing trends; of those trends, 109 increases and 25 decreases were statistically significant. In comparison with historical trends, 61.1% of countries are projected a slower rate of increase and 21.8% reversed from historical increase to a projected decrease. These findings indicate that Climate-AgTFP relationships are nonlinear and geographically differentiated and productivity growth is expected to weaken relative to historical trends.