人工智慧於河川生態水文系統之管理

Abstract

近年來由於全球氣候變遷的因應及生態環境復育意識的提昇，人類在追求物質的滿足之餘，慢慢開始重視與生態、環境的共存關係；河川流量管理即為一兼顧人類使用需求及河川生態系統需求之理念，將生態觀念融入河川流量經營管理之中，以達到人類與河川生態系統共存的理想；台灣生態水文指標系統考量台灣特有之水文地文條件，以自然河川的生態河川流態特性做為水資源管理的生態參考指標，為同時兼顧人類用水及河川生態之多目標水資源管理的重要參考。 本論文為一系列跨學科領域(生態水文系統)分析探討之研究，從時間及空間面向深入探討錯綜複雜的生態學與水文學間之關係，提出創新的分析技術，如結合統計方法、變異數分析、人工智慧及多目標最佳化搜尋等，皆廣泛的被應用在此研究中，本研究目的主要有:(1)台灣生態水文系統特性之探討與其於流量管理之應用；(2)建立台灣地區魚類生態水文地形之空間特性分區；(3) 探討台灣生態水文特性對魚類群聚於時間之影響；(4)應用類神經網路建立台灣生態水文指標對魚類生物多樣性之探討；(5)建立考量人類需求及提升生物多樣性之多目標最佳化水資源管理。 本研究首先透過迴歸分析整合生態水文因子與地形因子間之關係，以了解台灣生態水文與地形之相關性，建立水文-地形-生態間之交互關係，並應用新穎的聚類分析方法，探討台灣生態水文特性分區；接著利用變異數分析，探討TEIS所反映之台灣地理、地形特性及各指標在空間分佈上之差異，如不同集水區間或同集水區上游和下游監測站間，TEIS指標是否有差異，以提供水資源管理者在流量管理之重要參考依據。 在溪流生態學的研究中，空間與時間尺度及連續性等觀念經常被提出來探討及進行研究，因此在本研究中除了空間上的探討外，亦利用移動平均法深入瞭解水文資料對魚類生態資料在時間上的影響關係，並採用點二相關係數分析，將歷史流量資料與魚類資料結合產生一個水文-生態矩陣，討論TEIS與魚類生態資料之關係，此部分研究結果顯示河川流態資料有顯著影響之時間期距與台灣魚類平均壽命之長度相吻合。 然基於對大自然有限的了解，往往無法得知河川生態系統的實際需求，也無法明確地以數值或方程式的方式表示，因此本研究最後應用人工智慧技術，架構結合自組特徵映射網路與輻狀基底函數類神經網之複合式類神經網路:自組特徵輻狀基底類神經網路，並透過此模式利用台灣生態水文指標推估河川生態系統中之魚類生物多樣性，結果顯示此模式不但能有系統的分類歸納河川流量資料，並且能快速、有效率又精確的推估魚類之生物多樣性；本研究最後基於考量河川流態於提升生物多樣性的概念，發展以考量生物多樣性為原則之多目標水資源管理策略，具體建立提升河川生物多樣性及考量人類用水之永續水資源最佳化管理模式。

In response to global climate change and the raise of eco-environmental restoration concept, the equity between ecosystems, environment and human beings gains increasing attention for the past years. The concept of streamflow regime management is to incorporate ecological sustainability into flow regime management by taking the needs of both human and river ecosystems into consideration. Hydrologic indicator systems provide a summary of hydrologic conditions that are considered important to organism maintenance and ecosystem sustainability. The characteristics of ecological flow regimes can be used as reference indicators for multi-objective water resources management, which reflect the water demands of both human consumption and ecological systems. Physiographic and climatic conditions make Taiwan an ideal location for the examination of eco-hydrological issues in a sub-tropical climate. In this dissertation, a series of research is conducted and devoted to the interdisciplines of ecology and hydrology in Taiwan. The major purposes of this dissertation are to explore the complex relationship between ecology and hydrology in spatial as well as temporal aspects and to apply the research results to integrated water resources management. Therefore, innovated analytical techniques, such as a combination of statistical methods, multivariate analyses, artificial neural networks (ANNs) and multi-objective optimization methods, are used comprehensively in eco-hydro studies. The Taiwan Eco-hydrologic Indicator System (TEIS) was developed particularly to evaluate ecological flows in Taiwan and is adopted throughout this dissertation. The objectives of relational research consist of: (i) investigate the eco-hydrological characteristics of Taiwan rivers and assess the applicability of those characteristics to flow management; (ii) classify the spatial characteristics of the eco-hydro-geomorphic systems with respect to fisheries; (iii) assess the temporal eco-hydrological characteristics of fish communities; (iv) apply ANNs to estimate the eco-hydrological characteristics and fish bio-diversity in rivers; and (v) implement the multi-objective optimization of water resources management by taking the equity between human needs and riverine biodiversity into consideration. The first study of this dissertation identifies the eco-hydro-geomorphic relationship between TEIS statistics and geomorphic factors based on the regression analysis, and then a novel clustering analysis is used to divide Taiwan into several eco-hydrological regions based on TEIS statistics. The analysis of variance (ANOVA) is adopted in the second study of this dissertation to explore the spatial significance of eco-hydro systems based on TEIS statistics, which would reflect the geographical and geomorphological features of Taiwan. This analysis can provide water resources managers with an important reference for flow management. The assessment of riverine ecology is presented in the third study of this dissertation, which investigates issues such as the concepts of spatial-temporal scales. The moving average method is used to build the temporal relationship between hydrological data and fishery data, and then a FISH-TEIS matrix can be obtained from the point-biserial correlation analysis that combines historical flow data and fishery data. This analysis indicates that a flow regime defined by a 4-year record (the existing year and three antecedent years) is adequate to characterize the natural flow regime for Taiwan fisheries, which is consistent with the ecology and life spans of indigenous fish species. Due to the limited understanding of nature, it is difficult to get acquainted with the actual demands of river ecosystems or represent the systems by numerical methods and/or formulas. Therefore, the fourth study in this dissertation uses artificial intelligence techniques to build up a hybrid ANN that combines the self-organizing feature map (SOM) and the radial basis function neural networks (RBFNNs) into the self-organizing radial basis network (SORBN) for estimating fish bio-diversity based on TEIS statistics. The results show that this model not only can categorize stream flow data but also can estimate fish bio-diversity quickly, efficiently and precisely. Finally, the concept of improving riverine biodiversity is implemented in the fifth study of this dissertation to develop sustainable water resource management by considering both human and ecosystem needs.