人工神經網絡和長短期記憶模型探討森林生態系通量特徵

Abstract

摘要 在過去的幾十年裡，渦流相關係數技術已經被廣泛運用於觀測不同的地表-大氣交互作用，然而，由於設備故障或者低風速所造成的弱渦流現象，目前仍有大約20~60%的通量測量數據的缺失，在所有資料補遺的方法當中，機器學習(machine learning, ML)是一個強大的工具，可以簡單地建立輸入和輸出之間的非線性關係，並在許多通量的研究上面被廣泛利用。 本研究中，選擇以台灣亞熱帶地區棲蘭通量站，以探討機器學習在該地區的效能，本研究以溫度，土壤溫度，相對溼度，淨輻射通量四個參數作為主要分析對象，但也加入風速、風向、能見度、光合有效輻射做進一步的討論，其中主要使用人工神經網絡(ANN)和長短期記憶(LSTM)來預測生態系統當中的CO2交換和潛熱通量(LE)。並使用決定係數R2、平均絕對誤差MAE、均方根誤差RMSE等指標對於機器學習進行分析，後續研究使用偏微分(PaD)來探討不同參數的貢獻，以了參數之間的關係，最後將缺漏值進行填補，並和未填補資料進行比較及分析，以了解是否在季節性或或者整體上面有落差。 ANN和LSTM的初步結果顯示，在各種參數的組合之下使用八個參數可以得到較高的R2，CO2的部分分別為0.74和0.71，LE的部分為0.71和0.67，在日夜分開測量可以發現夜間無法有效預測，然而在PaD當中顯示Rn對於兩個參數的貢獻量最大，分別為62%和35%，而在填補後可以發現在冬季缺值較多的季節四分位明顯縮小，機器學習對於棲蘭地區有良好的預測能力。

Abstract The eddy-covariance technique has been widely applied to quantify the surface-atmosphere interactions over different landscapes in the past decades. However, nowadays there are still about 20 to 60% missing data in the flux measurement because of equipment failure or the weak turbulence caused by low wind speed. Among all the gap-filling methods, machine learning (ML) is a powerful tool to simply establish the non-linear relationship between the input and output parameters and has been broadly utilized in many flux studies. However, very little attention to ML was given to the analysis of multi-landscape comparison. In this study, the CLM flux station in the subtropical region of Taiwan is chosen to investigate the effectiveness of machine learning in this region. This study uses artificial neural networks (ANN) and long short-term memory (LSTM) to predict the CO2 exchange and latent heat flux (LE) in the ecosystem. In the subsequent study, partial derivation (PaD) was used to investigate the contribution of different parameters to understand the relationship between parameters, and finally, the missing plants were filled and compared with the unfilled data to understand whether there is any seasonal or overall discrepancy. The preliminary results of ANN and LSTM show that using eight parameters under various combinations of parameters can obtain higher R2, 0.74 and 0.71 for CO2 and 0.71 and 0.67 for LE, respectively, and in separate measurements of day and night it can be found that nighttime is not effectively predicted, however, in PaD it shows that Rn has the largest contribution to two parameters, 62%, and the PaD shows that Rn contributes the most to both parameters, with 62% and 35% respectively, and after filling the quartiles, it can be found that the quartiles shrink significantly in the winter season when there are more missing values, and the machine learning has good prediction ability in CLM.