土壤有機碳變化之整合統計分析

Abstract

氣候變遷加劇下，土壤作為重要碳匯，其儲存有機碳的潛力備受重視。多年生作物系統因具備持續輸入碳與低土壤擾動的特性，被視為有效的碳農業策略之一。本研究以全球多年生作物系統資料集為基礎，分析不同變化情境與土地利用轉變下對土壤有機碳儲量變化的影響，並結合機器學習與因果推論方法進行評估，作為自然碳匯策略之量化依據。

運用逐步迴歸、LASSO、隨機森林與梯度提升樹等模型進行重要變數篩選，以SHAP值解釋關鍵因子對土壤有機碳變化預測之貢獻；接續以變異數分析與穩健彎曲線迴歸辨識土壤採樣深度與作物種植年數對土壤有機碳變化的影響趨勢與轉折點；最終結合四種加權方法—逆機率加權、協變數平衡傾向分數、穩定平衡權重與能量平衡權重—建構因果模型以估計不同變化情境與土地利用對土壤有機碳的平均處理效應。

結果顯示，最濕季節降雨量、初始土壤有機碳含量為影響土壤有機碳變化之重要因子，持續種植多年生作物於第13.4年出現高峰，低土壤擾動下平均處理效應為6.7%；農田轉作的情境下，土壤有機碳變化率平均提升約4.3%，顯著提升土壤有機碳儲量，顯示其潛在增碳效益。

綜上所述，氣候與初始土壤條件為影響多年生作物系統下土壤有機碳累積的重要因子。透過機器學習與因果推論方法整合之分析框架，不僅驗證多年生作物持續種植具備顯著增碳潛力，亦指出土地利用轉作情境下，農田轉作多年生作物能有效提升土壤有機碳變化率，為自然碳匯策略與碳農業政策的推動提供實證依據。

Soils represent a major leverage point for climate mitigation, functioning as significant reservoirs of organic carbon. Perennial cropping systems, characterized by continuous carbon inputs and low disturbance, are considered promising nature-based solutions. This study uses a global dataset of perennial systems to evaluate soil organic carbon (SOC) responses under different transition conditions and land-use types. The objectives of this study are to combine machine learning algorithms and causal inference to identify key drivers and estimate treatment effects.

Utilizing stepwise regression, LASSO, random forest, and gradient boosting, we identified key predictors of SOC change. SHAP (Shapley Additive Explanations) values were used to interpret variable contributions. Temporal and depth-related trends were examined via ANOVA and robust bent line regression to identify structural inflection points. A causal inference framework incorporating four weighting methods—inverse probability weighting (IPW), covariate balancing propensity scores (CBPS), stable balancing weights (SBW), and energy balancing weights (EBW)—was used to estimate average treatment effects on the relative change in SOC.

Results show that wettest-quarter precipitation and initial SOC stock are dominant drivers. SOC accumulation peaked at 13.4 years under continuous perennial cropping, with an average treatment effect of 6.7% in low-disturbance systems. Cropland-to-perennial transitions increased SOC by 4.3% on average. These findings highlight the importance of climate and baseline soil conditions for SOC accumulation. The integrated analytical framework demonstrates the value of combining machine learning and causal inference to evaluate carbon farming strategies. This study offers empirical evidence supporting perennial cropping as an effective approach to enhance terrestrial carbon sequestration.