以人工智慧建立海岸濕地土壤碳匯空間分布推估之研究

Abstract

海岸紅樹林濕地具備極高的藍碳封存潛力，為達成 2050 淨零排放目標之關鍵自然碳匯。然而，於微觀尺度下，土壤有機碳（SOC）密度呈現顯著之空間異質性，且傳統現地調查受限於高昂成本，常面臨小樣本限制，使得高解析度之空間製圖面臨嚴峻挑戰。本研究旨在透過比較純地理統計、純機器學習與混合空間模型，建構適用於微觀高噪聲環境之最佳化 SOC 空間預測框架，並鑑定關鍵環境驅動因子，以優化國家藍碳之測量、報告與驗證效率。 研究首先於紅樹林濕地現場進行大規模現地採樣，透過實驗室化學分析測定土壤理化性質與多項離子濃度。隨後執行特徵工程與主成分分析以挑選最適合參數，並採用 Bootstrap 袋外交叉驗證與漸進式樣本縮放實驗進行模型評估。結果顯示，空間座標（特別是經度）對預測模型具壓倒性之主導權，證實微觀尺度下的 SOC 分佈深受巨觀環境梯度影響。排除地理座標後，氯與鉀離子為最關鍵之化學預測因子。在演算法性能比較中，深度類神經網路受限於小樣本數據，展現嚴重的「數據依賴性」；反之，隨機森林憑藉自助聚合機制，展現極佳之預測強健性與穩定性。 最終，本研究實證「結合機器學習之外部漂移克里金（KED-ML）」為最優化之混合製圖框架。透過將 RF 萃取之非線性化學趨勢作為外部漂移項注入克里金系統，KED-ML平衡了特徵空間之化學精準度與地理空間之連續性。本研究建立之「現地採樣與實驗室分析 → RF → KED 空間差值」分析，不僅提升了預測準確度，也為未來發展低成本、高精度之國家藍碳空間調查提供科學研究基礎。

Coastal mangrove wetlands possess exceptional blue carbon sequestration potential, acting as a critical natural carbon sink for achieving 2050 Net-Zero emissions targets. However, at micro-scales (e.g., 1-hectare), Soil Organic Carbon (SOC) density exhibits extreme spatial heterogeneity. Coupled with the small-sample constraints inherent to labor-intensive field coring campaigns, high-resolution spatial mapping faces significant challenges. This study aims to construct an optimized SOC predictive framework by systematically comparing pure geostatistical models, standalone machine learning algorithms, and hybrid spatial architectures. Furthermore, it seeks to identify key environmental drivers to enhance the efficiency of national Measurement, Reporting, and Verification (MRV) protocols for blue carbon. The methodology began with systematic field soil sampling in coastal mangrove wetlands and comprehensive laboratory chemical analysis to determine physicochemical properties and ion concentrations. Subsequently, rigorous feature engineering and Principal Component Analysis were applied to select the most suitable parameters, followed by Bootstrap Out-of-Bag cross-validation and progressive sample scaling experiments. Variable importance analyses revealed that macroscopic spatial geometric gradients fundamentally dictate SOC distribution. Upon excluding geometric coordinates, Chloride and Potassium emerged as the most critical chemical predictors. Algorithmically, Deep Neural Networks exhibited severe "data hunger" under small-sample constraints, whereas Random Forest (RF) demonstrated robust predictive stability through its bootstrap aggregating mechanism. Additionally, Universal Kriging outperformed Ordinary Kriging by successfully capturing the macroscopic trend surface. Ultimately, this study confirms that Kriging with External Drift by Machine Learning (KED-ML) is the optimal hybrid mapping architecture for micro-scale blue carbon. By injecting the non-linear chemical thresholds extracted by RF as an external drift into the kriging matrix, KED-ML successfully eliminated the artificial, perfectly balancing feature-space chemical precision with geographic-space continuity. The standardized analytical pipeline developed herein—comprising systematic field/lab analysis, RF driver extraction, and KED spatial smoothing—not only substantially enhances predictive accuracy but also provides a cost-effective and scientifically defensible foundation for future national blue carbon MRV spatial verification protocols.