利用機理性模型和機器學習預測陸地生態系土壤有機碳的穩定性

Abstract

土壤碳封存被認為是實現長期碳儲存與碳中和的潛在重要策略。本研究針對未施行主動減碳管理措施之農地、森林與草地生態系，系統性分析土壤有機碳（SOC）的穩定性與釋放行為，透過整合現地量測之土壤 CO2 通量（F）、由分解速率常數（k）推估之平均停留時間（MRT），並建立 k 與 F 之預測模型。結果顯示，在 0.2 m 表層土壤中（n = 4,963），超過 82% 的觀測值其 MRT 低於 10 年，四分位距（Q1–Q3）集中於 2–8 年，顯示表層 SOC 週轉速率偏快。相關分析與模型結果一致指出，溫度為控制 CO2 排放最關鍵之環境因子，當溫度上升 10 °C 時，F 與 k 分別增加約 1.6–2.3 倍。相較之下，碳氮比（CN）、黏土含量與 pH 雖可能影響 SOC 穩定性，但因資料量有限且時間解析度不足，其效應難以量化。模型選擇結果顯示，CN 與 pH 分別為農地與森林 CO2 排放之重要控制因子。外部驗證顯示，約 80–90% 的預測值落於三倍誤差範圍，顯示機理模型具有良好預測能力；相較之下，機器學習模型雖可提高模型準確度，但伴隨明顯過度參數化。將模型應用於台灣尺度後發現，不同縣市森林（n = 784）與農地（n = 102）之 MRT 介於 2.2–4.1 年，顯示表層 SOC 具有高度快速週轉特性。其空間分布進一步以 ArcGIS 呈現，證實本模型可有效整合於碳管理與策略規劃中。整體而言，本研究指出，在缺乏主動管理措施下，農地、森林與草地之表層 SOC 難以作為有效的長期碳儲存來源；而環境因子對 MRT 的關係分析，仍受限於場址差異性，以及缺乏高品質且具高時間解析度之 CO2 通量與土壤性質資料。

Soil sequestration has been suggested as a potentially important strategy for long-term carbon storage and achieving carbon neutrality. This study investigates the stability and release of soil organic carbon (SOC) cropland, forest, and grassland without active carbon mitigations. This is achieved by examining field collected CO2 flux (F), deriving mean residence time (MRT) from decomposition rate constant (k) and constructing predictive models of k and F. Results show that for SOC within the upper 0.2 m soil layer (n = 4,963), more than 82% of observations exhibit MRT values below 10 years, with interquartile (Q1–Q3) ranges of 2–8 years across croplands, forests, and grasslands. Correlation analysis and models both identify temperature as the most critical factor on CO2 efflux across the ecosystems, with a 10 °C increase resulting in a 1.6–2.3 fold rise in F and k. while carbon to nitrogen ratio (CN), clay content, and pH are potentially important determinants of carbon stability, their effects are less quantifiable as their data are more limited and inadequately resolved temporally. Model selection suggests that CN and pH may be important for surficial CO2 efflux in croplands and forests, respectively. External validation shows that 8090% of in situ prediction fall within a threefold error range, indicating good performance of the mechanistic model. Machine learning based models are found to be statistically inferior to mechanistic models, with better accuracy achieved at the cost of substantial over-parameterization. Application of the developed mechanistic models reveals rapid surface SOC turnover in Taiwan, with MRT ranging from 2.2 to 4.1 years across forest (n=784) and cropland (n=102) soils in different counties. The assessed MRT are spatially mapped using ArcGIS, illustrating the mechanistic models can be readily integrated into carbon management practice or strategic planning. Overall, this study demonstrates that surface SOC in cropland, forest, and grassland is not an effective long-term carbon storage without active mitigative management practices. Further delineation of environmental factors on MRT is constrained by site-specificity and the lack of high-quality and high-resolution temporal data of both CO2 efflux and soil properties.