以類神經網路建立環境因子與魚類群聚之間的關係

Abstract

臺灣位於亞熱帶季風區，再加上該地區地形及環境之影響，臺灣地區的降雨量分佈豐枯懸殊，水資源管理相當困難。近年來台灣地區隨著工商業發展，經濟及民生活動增加，用水需求大量提升。 為了因應農業、工業、民生等用水需求，於各大流域廣建水庫。然而大型水工結構物卻改變了原本河川之自然環境，阻隔了生態系統間之連結；加上人類活動所產生之民生與工業污水多未經處理即排入河川中，嚴重污染河川之水質，亦對河川生態環境系統造成極大之衝擊。 新店溪流域於2005年至2012年間，有長期且連續之流量、水質及魚類生態之調查；且於水庫的建造及相關的極端事件(颱洪、乾旱事件)，亦有詳細之記載。本研究將探討新店溪之流量、水質及水工結構物對魚類生態系統之影響，以了解流量、水質、生態三者間之交互作用，並分析極端事件對於河川生態系統之衝擊，以利日後水資源管理之有效規劃與永續經營。 本研究首先依據台灣生態水文指標系統(Taiwan Ecohydrologic Indicator System, TEIS)，將流量資料轉換成月尺度之流態，再以統計方法探討極端事件對於河川魚種組成之影響。接著運用Gamma檢定篩選影響河川魚種相似度之關鍵水質及流態因子，並以類神經網路之自組特徵映射組織網路(Self-Organizing Map, SOM)探討關鍵因子與魚種相似度之間的交互關係，最後以倒傳遞類神經網路(Back Propagation Neural Network, BPNN)及調適性網路模糊推論系統(Adapted Network-Based Fuzzy Inference System, ANFIS)建置魚種相似度推估模式。 根據新店溪流域的研究結果顯示：採樣資料受到極端事件的影響並不明顯，但也有可能是因調查時距過長，無法看出極端事件有無影響。兩測站間採樣結果的差值顯示懸浮固體、電導度、及流量當月十日最大值及最小值為影響河川魚種相似度之重要因子。SOM的聚類結果顯示水質相近之測站有較高的魚種相似度，而流態相近之測站的魚種相似度高，然而在魚種相似度低的情況，流態則無明顯差異。ANFIS的推估模式，可以將輸入的資料特性以隸屬函數呈現，透過與SOM的連結，可建立環境因子與魚種相似度的規則庫。BPNN的推估模式可較ANFIS更準確地計算兩地的魚種相似度。本研究建立了魚種相似度之推估模式，考量河川流態、水質及水工結構物對魚種組成之影響，期能提供未來水資源管理及河川生態保育規劃之參考依據。

Being located in the subtropical monsoon zone along with the topographically steep gradients, Taiwan suffers from extremely unevenly distributed rainfall in both space and time. This condition makes the use and management of water resources particularly difficult. The fast economic growth and urbanization has led to dramatically increased water demands in recent years. Large-scale reservoirs were therefore constructed to satisfy water needs of agricultural, industrial and public sectors, and effective reservoir operations were managed to maximize the use of limited water resources. However, the natural physical environments have been altered by reservoirs and hydraulic facilities. Additionally, the lack of wastewater monitoring programs has deteriorated water quality in rivers. Being concerned with ecological impacts made by the changes of physical and chemical conditions in rivers, this study aims to explore the complicated relationships of river flow, water quality and fish community, examine the impacts of extreme events on riverine fishes, and investigate the influence of hydraulic facilities on eco-hydrological environments for making an environmental- and ecological-friendly water resources management. In this study, the Shindien River was chosen as the investigative area because long-term (2005-2012) records of river flow, water quality and fish distribution in the river were available. The construction and operational histories of reservoirs and the records of extreme events (typhoons and droughts) were also available. Daily river flows were converted into a flow regime based on Taiwan Ecohydrologic Indicator System (TEIS). Then statistical analyses were conducted to explore the impacts of extreme events on fish species composition. Following that, Gamma Test (GT) was used to determine key factors affecting fish species similarity with respect to water quality parameters and flow regime variables, and the Self-Organizing Map (SOM) was applied to exploring the non-linear relationships among key factors and fish species similarity. Finally, two predicting models were constructed to estimate the fish species similarity by using the Back Propagation Neural Network (BPNN) and the Adapted Network-Based Fuzzy Inference System (ANFIS). The results of statistical analyses indicated that extreme events did not show statistically detectible changes in the surveyed fish species composition, which might be because the time interval of two consecutive surveys was too long. Such phenomena therefore could not clarify the impacts of extreme events on fish species composition. The results of the GT indicated that differences in suspended sediments, electrical conductivity, 10-day maximum flow and 10-day minimum flow of the surveyed month between two sampling sites were obvious and therefore these four variables were identified as key factors affecting fish species similarity. The clustering results of the SOM indicated that two sampling sites with similar water quality showed high fish species similarity, so as two sampling sites with similar flow regimes. Nevertheless, the flow regimes of two sampling sites with low fish species similarity did not make significant difference. The ANFIS model provided advantages of using membership functions to display the characteristics of input data, and when linking with the SOM, it was flexible in developing knowledge-based rules to rationally suggest the effects of environmental factors on fish species similarity. The BPNN model showed better prediction accuracy than the ANFIS. In sum, this study systematically explored the complex influences of flow regime, water quality and reservoirs on fish species composition and well constructed fish similarity prediction models, which can provide useful information for effective and sustainable river ecosystem management.