[Construction and Driving Factors Analysis of a Machine Learning-based Prediction Model for Net Carbon Sink in...

Intelligent prediction of agricultural net carbon sink and the mechanistic analysis of its driving factors are of great significance for promoting carbon reduction and sequestration policies in China's agricultural and rural sectors under the "dual carbon" goals. Based on the calculated data of agricultural net carbon sinks and panel data of 31 provincial regions in China from 2000 to 2022， multiple machine learning algorithms were used to construct a prediction model for agricultural net carbon sinks. The SHAP values and PDP plots were employed to reveal the response characteristics of the driving factors of the prediction model to agricultural net carbon sinks. The results show that：①The GWO-RF model constructed in this study demonstrated high prediction accuracy and stability for the agricultural net carbon sink （MSE = 0.04%， MAPE = 7%， R² = 0.984）. ② The importance ranking of the model's driving factors was as follows： effective irrigated area > cultivated land area > provincial characteristics > fertilizer application intensity > regional education level > urbanization level > agricultural mechanization level. ③The 2D PDP results of pairwise interactions among the three important driving factors， namely effective irrigated area， cultivated land area， fertilizer application intensity， on agricultural net carbon sinks showed that： First， the interaction between the effective irrigated area and cultivated land area was weak when the effective irrigated area was less than 1 600×10³ hm² and strong when it was greater than this value； second， the interaction between the effective irrigated area and fertilizer application intensity was weak， and the influence of the effective irrigated area was dominant； and third， the interaction between cultivated land area and fertilizer intensity was weak when the cultivated land area was less than 1 600×10³ hm² and strong when it was greater than this value. ④ The 3D PDP visualization of the above three driving factors on agricultural net carbon sinks showed that the effective irrigated area had the most significant impact on agricultural net carbon sinks， almost dominating the entire process. The underlying reason is that the intensity and method of irrigation can significantly affect the potential of crop carbon absorption and soil carbon sinks. The research results provide a new method and novel approach for the prediction of agricultural net carbon sinks， also providing decision-making references for the government and relevant departments in formulating agricultural carbon emission reduction and sequestration plans and policies.