Machine Learning Driven Optimization of Carbon Sequestration in Intercropping Systems Using XGBoost Modeling and Partial Dependence Analysis

This study integrates XGBoost modeling with Partial Dependence Plots to optimize carbon sequestration. The experiment was conducted using a split-split plot design replicated thrice with the main plot compressing tillage practices viz. , minimum tillage (M ₁ ) and conventional tillage (M ₂ ), subplot consists of row ratios viz. pigeonpea + maize (1:2 ratio) (R ₁ ), pigeonpea + maize (1:3 ratio) (R ₂ ), and sole pigeonpea (R ₃ ) and sole maize (R ₄ ), sub-subplot consists residue management practices viz. on farm produced biochar application (S ₁ ), on farm produced residue application (S ₂ ), and control with no biochar or residue application (S ₃ ). Machine learning models of XGBoost accurately predicted CSQ while Artificial Neural Network performed best for SP and CF, confirming nonlinear relationships. PDPs showed available nitrogen increased productivity from 3.0 to 3.4 Mg ha⁻¹ y⁻¹ (150–300 kg ha⁻¹), while SOC reduced it from 3.8 to 3.2 Mg ha⁻¹ y⁻¹ (0.425–0.525%). Sensitivity analysis identified SOC, moisture, and enzyme activities as key drivers, with SOC enhancing sequestration by ~ 1.5 Mg C ha⁻¹ and improving footprint by ~ 1.2 Mg CO₂-Ce ha⁻¹. These insights enable targeted, efficient soil management for productivity and climate benefits. Machine learning-based Partial Dependence Plot (PDP) and sensitivity analysis (SA) revealed SOC, BD, and microbial parameters as key drivers of these outcomes.