Enhancing climate prediction accuracy with machine learning to promote agricultural carbon emission reduction

Abstract Agricultural greenhouse gas emissions exhibit strong nonlinear responses to climate conditions, yet conventional numerical simulation approaches fail to meet the demands of regional-scale emission reduction decision-making. To address the problem of blind agricultural carbon management caused by insufficient climate prediction accuracy, this paper proposes a physically constrained ensemble learning framework. The framework integrates satellite remote sensing, ground-based monitoring, and reanalysis datasets to construct a spatiotemporally adaptive climate prediction structure. A bidirectional coupled climate–carbon emission interaction model is established to quantify the dynamic effects of temperature and precipitation on soil respiration, paddy field methane emission, and nitrogen fertilizer-induced nitrous oxide emission. A prediction-driven multi-objective optimization algorithm is further developed to support intelligent control of cropping systems and agronomic management for emission minimization. Experiments conducted on a winter wheat–summer maize rotation site in the North China Plain demonstrate that, compared with conventional empirical models, the proposed framework reduces the root mean square error of monthly temperature prediction, raises the precipitation skill score, and achieves a 14.7% reduction in carbon emission intensity per unit yield through optimized agricultural management, offering technical support for climate-smart agriculture.