類神經網路推估蒸發量:(I)結合動態因子分析與(II)使用衛星資料

Evaporation Estimation using Artificial Neural Networks: Based on (I) Dynamic Factor Analysis and (II) Satellite Images

孫 維(Wei Sun)

指導教授 ： 張斐章

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2011年)

https://doi.org/10.6342/NTU.2011.00864

摘要

蒸發量實為水循環、水資源管理以及農業灌溉重要的一環，蒸發量之測量以A型蒸發皿為主，而一般於蒸發量之推估主要以物理經驗式做推估之方式，但其準確率仍有相當大之改進空間。本研究主要可分為兩部分，第一部分為介紹結合動態因子分析(DFA)以及倒傳遞演算法(BPNN)之新型模式-BD Model，此模式大大的提升推估準確率。動態因子分析(DFA)首次被應用於蒸發量以及蒸發趨勢的推估，由本研究結果顯示，動態因子分析可以有效的建立測站間蒸發量之共同趨勢，並藉由赤池資訊準則(AIC)值做為評估標準，篩選蒸發量模式所需之氣象輸入因子。最後將所選擇之氣象輸入因子以及動態因子分析所得之資訊做為倒傳遞演算法之輸入，並推估A型蒸發皿之蒸發量。研究結果顯示，BD模式於蒸發量之推估有相當良好之準確性。
研究第二部分為使用遙測影像推估台灣全島之蒸發量，並建立台灣之蒸發量地圖 (Evaporation Map)。在此部分使用Landsat5以及Landsat7之衛星影像產品-EVI植生指數以及表面溫度做為模式之輸入，並使用調適性網路模糊推論系統(ANFIS)做為核心推估模式推估全台蒸發量。此部分研究提供不同於以往的方式推估大面積、大範圍之蒸發量，其推估值雖不如第一部分之BD模式來的準確，但RMSE誤差仍在1mm/day左右，此誤差值對於水資源管理為可容許範圍。
本研究提供了兩種不同的方式推估蒸發量，第一部分之BD模式，在有足夠的氣象資料下，可以提供更為精準之推估。而第二部分使用調適性模糊推論系統結合衛星影像則可以做大範圍蒸發量之推估，縱使該區域沒有氣象測站仍可以有效推估蒸發量。期望本研究之兩種蒸發量推估方式可以使水資源管理更加確實且有效率。

關鍵字

蒸發量 ； 類神經網路(ANN) ； 動態因子分析(DFA) ； 調適性網路模糊推論系統 (ANFIS) ； 倒傳遞類神經網路(BPNN) ； EVI植生指數 (EVI) ； 地球資源技術衛星Landsat.

並列摘要

Evaporation is one of the major elements in the hydrological circle and an important reference to the management of water resources and agricultural irrigation. To efficiently explore the mechanism and spatial distribution of evaporation, the study consisted of two parts, in which the first part proposed a hybrid model (BD) combining Back-Propagation Neural Networks (BPNN) and Dynamic Factor Analysis (DFA) to improve the accuracy of evaporation estimation, and the second part made use of the satellite images to establish the spatial distribution of evaporation covering whole Taiwan.
In the first part, the DFA was first applied to investigate the influence of meteorological variables on evaporation. In addition, the common trend extracted from evaporation observations at each gauging station was obtained by evaluating the corresponding AIC (Akaike’s information criterion) values. Furthermore, the explanatory meteorological variables highly related to evaporation were also identified through the DFA. Finally, the BPNN was used for accurately estimating evaporation based on the selected explanatory meteorological variables and DFA estimation, and the performance of the constructed BD model was compared with that of empirical formulas. Results demonstrated that the proposed BD model has excellent applicability and reliability in terms of the accuracy of evaporation estimations.
The second part aims to construct an effective evaporation estimation model that possesses the ability to present the spatial distribution of evaporation in Taiwan. To achieve this goal, the remote sensing images obtained from Landsat 5 and Landsat 7 satellites were used as inputs to the Adaptive Network-Based Fuzzy Inference System (ANFIS). The image products included Enhanced Vegetation Index (EVI) and surface temperature with a sample size of 342. Results obtained in this phase indicated that the ANFIS model can easily perform the variation of evaporation estimations in space and accurately capture the trend of evaporation with errors of about 1 mm/day, which is acceptable for relative applications. Overall, the estimations of evaporation were achieved in this study in the aspect of point and regional estimations through BD and ANFIS approaches, respectively. The performance demonstrated that both models are of great stability and reliability in evaporation estimation, which are capable of providing valuable information for water resources management.

並列關鍵字

Evaporation ； Artificial Neural Network (ANN) ； Dynamic Factor Analysis (DFA) ； Adaptive Network-Based Fuzzy Inference System (ANFIS) ； Backpropagation Neural Network (BPNN) ； Enhanced Vegetation Index (EVI) ； Landsat.

參考文獻

Akaike, H., 1974. New look at statistical-modle identification. IEEE Transactions on Automatic Control, AC19(6): 716-723.

Antar, M.A., Elassiouti, I. and Allam, M.N., 2006. Rainfall-runoff modelling using artificial neural networks technique: a Blue Nile catchment case study. Hydrological Processes, 20(5): 1201-1216.

Behzad, M., Asghari, K., Eazi, M. and Palhang, M., 2009. Generalization performance of support vector machines and neural networks in runoff modeling. Expert Systems with Applications, 36(4): 7624-7629.

Beven, K., 1979. Sensitivity analysis of the Penman-Monteith actual evapotranspiration estimates. Journal of Hydrology, 44(3-4): 169-190.

Boegh, E., Soegaard, H. and Thomsen, A., 2002. Evaluating evapotranspiration rates and surface conditions using Landsat TM to estimate atmospheric resistance and surface resistance. Remote Sensing of Environment, 79(2-3): 329-343.