以自組特徵映射網路推估蒸發量

Abstract

蒸發現象為影響水氣於水文循環分佈中重要之因素，在農業水資源管理上扮演重要角色。傳統經驗式利用氣象變數推估蒸發量，而忽略蒸發在自然界中呈高度非線性現象，故本研究利用具有分類特性的自組特徵映射網路架構一蒸發量推估模式。 本研究利用恆春氣象站的氣象變數作為模式輸入，透過自組特徵映射網路(SOM)學習，將相似特性的輸入資料聚為一類，並探討拓樸網路架構的潛在特性。自組特徵映射網路可快速有效將輸入的氣象變數分類，形成網路拓樸層，再將各聚類之中心點以線性迴歸方式與輸出層連結，可準確的推估蒸發量。另外建立強制型自組特徵映射網路(ESOM)以加強映射較為極端的案例空間，並與Modified Penman(FAO,1984)、Penman-Monteith(ICID,1994)等傳統經驗式進行比較。結果顯示，拓樸層架構能詳細說明輸入與輸出間映射的關係，且用SOM與ESOM可根據氣象變數作良好的推估；四種模式中以ESOM推估表現最好(RMSE=1.15mm/day，MAE=0.87 mm/day)，對於長期蒸發量的推估表現中，也是以ESOM表現最佳。研究再針對已建立的模式進行穩定性與適用性討論，結果顯示直接將網路用於其他地區會因區域蒸發量的差異造成模式推估值與實際觀測值有較為明顯的差異。

The phenomenon of evaporation is an important factor that affects the distribution of water in hydrological cycle and plays a key role in agriculture and water resource management. The tranditional evaporation formulas usally neglect the non-linear characteristics in the nature. In this study we propose the self-organizing map(SOM) network to estimate daily evaporation. First, the daily meteorological data from climate gauges were collected as inputs of the SOM and then classified into topology map based on their similarities to investigate their potential property. To effectively and accurately estimate the daily evaporation, the connected weights between the cluster in topology layer with output layer were trained by using the linear regression method. In addition, we bulit enforced Self-Organizing Map (ESOM) to strength mapping spaces for these extremely data and compared with Modified Penman (FAO,1984) and Penman-Monteith (ICID,1994). The results demonstrated that the topology structures of SOM and ESOM could give a meaningful map to present the clusters of meteorological variables and the networks could well estimate the daily evaporation based on the input meteorological variables used in this study. In comparing the performances of these four models, the ESOM provides the best performance (RMSE=1.15mm/day，MAE=0.87 mm/day). The ESOM performance is also well in estimating long term evaporation. We have the suitability of using these models in other areas where their evaporations are different widely from the original station, the estimation, however, are not well as the one we use in the built station. This result suggests that the network must be adequately trained before it is used to estimate the local evaporation.