應用類神經網路推估低緯度濕地二氧化碳通量變化

Abstract

目前濕地研究多集中於高緯度的泥炭地(peatlands)，低緯度濕地的研究相對較少，然而低緯度濕地面積占全球濕地面積之70%，若以泥炭地之研究結果進行全球濕地碳通量的推估，可能會與實際情形產生誤差。在碳收支的研究中，渦度相關法(eddy-covariance method)被認為是最直接而準確地量測生態系統通量的方式，但受限於此法對量測環境的要求，故無通量塔處之碳通量變化趨勢須以推估的方式進行。在眾多推估方式中，類神經網路(artificial neural network)在各種不同生態系統的推估研究中，已證實具有良好的模擬能力，能夠精確掌握氣象資料與碳通量資料之間的關聯性。 本研究使用類神經網路作為推估工具，採用關渡塔一(GDP-T1)、關渡塔二(GDP-T2)、佛羅里達站(US-Esm)三個低緯度濕地測站資料(包括氣象及二氧化碳通量)作為輸入及輸出參數，配合倒傳遞演算法(back propagation algorithm)進行模式建立與推估。模型訓練完成後，計算相關係數(Correlation coefficient)、均方根誤差(RMSE)、平均絕對百分比誤差(MAPE)等指標，以討論推估結果與觀測結果之間的誤差情形，以及模型推估二氧化碳通量變化模式之能力。GDP-T1及GDP-T2站最佳結果出現在日間模型(R=0.89及R=0.87)，US-Esm站點最佳結果出現在夜間模型(R=0.62)。跨站模擬最佳情況可高達R=0.73，此結果顯示未來以類神經網路應用於大範圍推估為可行的。 本研究所建立之模型可應用於濕地生態系統進行二氧化碳通量數據的推估，提高全球濕地碳收支推估精確程度，並進一步研究無通量塔處之二氧化碳通量變化特徵，分析相似生態系統二氧化碳通量的變異模式，如季節或年際變異等，克服通量塔設置之空間限制及節省儀器架設成本。在面對氣候變遷時，此模型的推估能力也可增加對於未知風險的了解，作為評估未來變化趨勢的依據之一。

In wetland studies, few attentions have been given to low-latitude wetland ecosystem presently, but it accounts for about 70% of the global wetland area. Therefore, it’s very important to consider the contribution of this significant portion on global carbon (C) budget. In the past decades, eddy-covariance method has been widely applied in many C budget studies at the ecosystem scale, but there are still several limitations affecting the performance of EC methods. In order to overcome the abovementioned limitations, many linear or non-linear statistical techniques are applied to fill the measurement gap. Among various methods, the Artificial Neural Network (ANN) method is considered to be an excellent means to identify the complex non-linear relationship between the CO2 flux and meteorological variables. In this study, a back-propagation ANN model was applied to quantify CO2 flux at three low-latitude wetland sites (Guandu Nature Park Tower One (GDP-T1), Guandu Nature Park Tower Two (GDP-T1), and Florida Everglades short hydroperiod marsh (US-Esm)) in East Asia and the US. Meteorological variables were used as the input parameters to train the ANN to predict the CO2 exchange. The best results of the GDP-T1 (R=0.89) and GDP-T2 (R=0.87) occurred in the simulation of the daytime (DT) model, and that of the US-Esm (R=0.62) in the nighttime (NT) models. The cross-site simulation was feasible, the best result could up to 0.73 in terms of R. This model provided a quick, efficient, and highly accurate estimation, and could be conducted to estimate the dynamics of CO2 flux where there is no direct in-situ flux measurement. The simulation capability is helpful to characterize the spatial/temporal variations in low-latitude wetland ecosystems, and improve the quantification of global C budget.