發展反向混合模型以解析河川中溶解性營養鹽端源濃度和輸出量貢獻

Abstract

人為活動 (人口和土地利用) 與水文情勢為控制營養鹽(C/N)的來源、動態和輸出的重要因素。為了確定營養鹽流佈的關鍵控制因子，需要提出一種能穩定解析不定端源的方法來了解河流營養鹽來源。本研究包括三個部分：1) 開發混合模型反向解法：採用新的逐步階層式自動變數選擇算法，以蒙地卡羅方法生成混合模型的係數集，搭配四個模型效能指標：AIC, BIC, adjust R2和VIF以多目標柏拉圖最適化決定可接受的參數，以提供高維穩定問題的解法。此流程寫成新的R套件：InvEMMA。2) DIN (Dissolved inorganic nitrogen) 輸出的土地利用控制。本研究選在台灣北部的都市化山地小河流域 (淡水河)，以2002-2005年河川DIN負荷，計算關鍵土地利用類別及輸出係數。結果表明，在21種土地利用類別中，有5種被自動選為DIN出口控制關鍵因素：森林，農業 (包括水稻，農田，果園)和城市地區。森林中的DIN產量 (521.5 kg-N km-2 yr-1) 比全球河川DIN的產量高約2.7倍。在城市地區，點源與非點源之間的差異分析表明，未經處理的污水(非點源)約佔人類關聯負荷的93％，導致城市的出口係數較高。3) 水文情勢控制河川溶解碳動態。河流溶解碳(包括DOC (dissolved organic carbon)和DIC (dissolved inorganic carbon))的運輸是連接陸地和水生碳儲層的關鍵過程，但其來源（地表、地中或地下）與流路（受地表逕流、地下水流或地下水），深受水文情勢控制。這項研究監測了2014-2016間台灣西南部曾文水庫上游三個山地小河日尺度及兩次颱風事件期間的DOC和DIC濃度。結果表明，兩個颱風僅在三天內就貢獻了年度DOC和DIC通量的15.0-23.5％和9.2-12.6％。選擇的三個主要端源會在事件發生期間隨流動狀況而變化其組成。採用HBV水文模式模型，並結合端源混合模型，確定來自不同端源組成部分的DOC和DIC傳輸量，地表端源貢獻事件尺度DOC達40-48%，地下水端源貢獻DIC約44-57%。在估算全球碳收支時，應考慮台灣山地小河獨特的模式，其特徵是高溶解碳通量和颱風帶來的大量運輸。總而言之，本研究開發了一種自動選擇程序，以查找河中DIN，DOC和DIC出口過程中的關鍵控制因素。這有助於我們避免在判斷變量時出現系統性的人為錯誤。因此，可以根據鑲嵌景觀特徵或動態流分量估算合理的模型結構和特定的養分產量。這種方法很好地評估了河流的養分負荷並提高了對養分傳輸過程的理解。

Anthropogenic activities (population and land use) and hydrological regime regulate the source, dynamics, and output of major nutrients (C/N) at the watershed scale. To determine the critical (less systematic human error) control factors of riverine nutrients, a relatively reliable method is needed to be proposed. This study consists of three parts: 1) Develop the mixed model inverse analyzing method: adopt a new stepwise and hierarchical automatic variable selection algorithm, generate the coefficient sets of the mixing model with the Monte Carlo method, and combine four model performance indicators: AIC, BIC, adjust R2 and VIF with multi-objective Pareto frontier optimization to provide stability solutions to high-dimensional problems. This process is wrapped as a new R package: InvEMMA. 2) Land use control of DIN (dissolved inorganic nitrogen) export. The study selected an urbanized small mountainous river watershed (Danhui River) in northern Taiwan, calculated key land use categories and export coefficients based on the river DIN load from 2002 to 2005. The results show that 5 of 21 land-use categories auto-selected as DIN export controlling key factor: forest, agriculture (including paddy, cropland, orchard), and urban area. DIN yield in the forest (521.5 kg-N km−2 yr−1) was ~2.7-fold higher than the global riverine DIN yield. In an urban area, the analysis of differentiation between point and non-point sources showed that the untreated sewage (non-point source), accounting for ~93% of the total human-associated load, resulted in a high export coefficient of urban. 3) The hydrological regime controls the dynamics of riverine dissolved carbon. The transportation of dissolved carbon, including DOC (dissolved organic carbon) and DIC (dissolved inorganic carbon) in rivers, is a crucial process connecting terrestrial and aquatic carbon reservoirs. The study monitored the southwestern part of Taiwan from 2014 to 2016. The results showed that the two typhoons contributed 15.0-23.5% and 9.2-12.6% of the annual DOC and DIC flux in only three moments. The three main end-members selected will change their proportion during the event. Using the HBV hydrological model coupling with the mixing model to determine DOC and DIC transport quantities from different end-members components. The surface end-member contributes 40-48% of the event scale DOC, and the groundwater end-member contributes about 44-57% of the DIC. When increasing the global carbon budget, consideration should be given to the unique model in Taiwan's SMR, characterized by high dissolved carbon flux and massive transportation caused by typhoons. All in all, this study has developed an automatic selection program to find the critical control factors in the export process of DIN, DOC, and DIC. This process helps us avoid systematic human error when judging variables. Thus, a reasonable model structure and specific nutrient yield can be predetermined according to the expected landscape characteristics or dynamic flow components. This method provides a good assessment of the riverine nutrient load and improves the nutrient transport process's understanding.