應用機器學習推估水稻田土壤有機碳之時空動態與儲存潛力

Abstract

隨著工業快速發展伴隨而來的氣候變遷現象，減碳議題近年來受到國內外組織規劃與政策實施之重視，為因應減少大氣中的二氧化碳總量，以實現減緩氣候變遷的現象。我國農業部門為達到減量的目標，亦提出永續農耕方式，並嘗試提升農業土壤碳匯量。然而，目前尚未有明確訂定用於分析土壤碳匯量之動態變化的模型框架。 本研究為解決該問題，參考國外相關之方法學基礎，考量極端氣候及非氣候因子對農業土壤有機碳儲存潛量帶來不確定性，以臺灣主要糧食水稻產區－－雲嘉南為研究對象，探討未來約十五年的土壤有機碳動態分布。本文利用機器學習整合氣候、土壤條件等資料建立模型，以評估淨初級生產量及土壤有機碳之動態變化。其中包含兩階段模型，其一為淨初級生產量之時空動態模型，推估其隨時間之分布變化；其二則為地上部與地下部之數值迴歸模型，建立遙測與土壤調查資料之關係。藉由統計方法，於研究結果說明參數相關性特徵，檢視參數對於模型之顯著性，並分析環境變量對於碳潛量的分布特性。 本研究進而探討氣候變遷情境與土地管理實踐情境之碳潛量變化。於劇烈的氣候變遷條件 (SSP5-8.5) 下，其變化率相對較和緩的氣候情境 (SSP1-2.6) 對於歷史情境的23%，提升為29%，顯示土壤碳儲存潛量穩定性有下降的可能性。接著，以SSP1-2.6為基線推估未來在土地管理實踐措施下，土壤有機碳儲存潛量的變化情形，相對2000年，至2035年可提升27%以上。這些結果顯示面對不同氣候變遷情境下，土地管理實踐對於土壤碳潛量的重要性。 本研究透過衛星遙測資料、地理資訊系統與機器學習方法，呈現並說明時空間農業土壤碳潛量之連續性變化；並致力於提升對水稻田土壤碳儲存潛量變化之理解，分析環境變量對土壤碳指標的影響，以供未來相關研究與政策之參考。

Climate change has accompanies the rapid development of industrialization, and carbon reduction issues have received attention from domestic and foreign organizations in planning and policy implementation, aiming at reducing the sum of carbon dioxide, and mitigating the phenomenon of climate change. To address carbon reduction, Taiwan’s Ministry of Agriculture (MOA) has proposed sustainable agricultural practices to increase soil carbon storage. However, Taiwan lacks a clearly defined model framework for analyzing the dynamic changes in soil carbon sinks. In order to solve this problem, this study refers to foreign methodological foundations, considers the uncertainty caused by environmental variables on soil organic carbon (SOC) potential, and focuses on the primary rice-growing regions of Yulin, Chiayi, and Tainan in Taiwan. The goal is to explore the future spatiotemporal patterns of SOC distribution. This study employs machine learning to develop a model to evaluate the dynamic changes in net primary production (NPP) and SOC. The model comprises two stages: a dynamic potential change model of NPP distribution with the spatiotemporal resolution, and a regression model for aboveground and underground parts to establish a remote sensing relationship with soil survey data. Furthermore, statistical methods are used to analyze the significance of climate variables on soil carbon indicators. This study explores the changes in carbon potential under further climate change scenarios and land management practices (LMPs). Under the severe climate scenario (SSP5-8.5), the climate scenario with a relatively mild changing rate (SSP1-2.6) increased from 23% of the historical scenario to 29%, indicating that the stability of SOC storage potential has declined. Using SSP1-2.6 as a baseline, it is showed that future land management practices could increase soil carbon storage potential by up to 27% by 2035 compared to the year 2000. These results indicated the importance of land management practices for SOC in the face of climate change. This study focuses on using satellite remote sensing data, Geographic Information Systems (GIS), and machine learning methodologies to present and explain changes in spatiotemporal continuity. It is also committed to understanding changes in soil carbon storage potential in rice fields and analyzing the dynamic changes of various environmental variables on soil carbon indicators to provide a reference for future studies and policy development.