河川網路時空推估架構之發展並應用於頭前溪魚類與水質分布特性之探討

Abstract

河川棲地環境與魚類豐度的研究，往往受限於採樣工作的困難與相關費用高昂，難以提高資料的時空間解析度，近年來地理統計方法(geostatistical methods)的盛行，藉由已知樣點資料推估未知時空間點位的值，適合做為彌補生態領域現地採樣資料不足問題的模式工具。傳統地理統計方法大都建立在資料常態分佈與線性推估等既有假設之下，並且無法分析不確定性資料(soft data)與應用在河川網路系統。貝氏最大熵法(Bayesian Maximum Entropy method)以貝式條件機率的概念結合物理知識與不確定性資料逐步增強推估資訊，被廣泛應用於許多研究領域之非定常(non-stationary)、非同質(non-homogeneous)時空過程推估，是新興的地理統計方法。本研究利用移動平均法(moving-average construction)建立符合地理統計假設的共變異數模式，再應用貝氏最大熵法推估頭前溪環境因子之時空間分布。 何氏棘魞(Spinibarbus hollandi)為台灣特有種的淡水魚類，主要分布於台灣南部與東部河川，為當地之原生魚種；近年來因為人為放養，何氏棘魞開始出現在台灣西部河川流域，對當地河川生態造成影響，所受衝擊尤以新竹頭前溪為甚。本研究根據前人文獻，選用流速、水深等棲地參數作為分析頭前溪何氏棘魞豐度與河川環境關係的指標非生物因子，利用結構加成性迴歸模式(structured additive regression model)建立頭前溪何氏棘魞與環境因子之關係模式，期能以此作為管理河川外來魚種入侵與河川棲地生態保育暨復育的參考。 研究結果顯示，傳統共變異數模式經過移動平均法修正後，可以應用於河川網路系統並產生符合地理統計假設的共變異數矩陣，應用貝氏最大熵法推估的頭前溪環境參數結果偶有較大誤差，原因為特定測站特性與其附近提供之資訊較少所致；在頭前溪的竹林大橋，何氏棘魞較喜好水深適中，低流速的潭區作為棲息地，該結果可提供作為該物種人為移除與原生棲地保育的參考；結構加成性迴歸模式之交叉驗證結果，平均相對誤差最小約30%，最高則達400%，造成誤差之可能原因除了現地採樣誤差之外，提供訓練及驗證模式之資料偏少也是可能因素，此外，未考慮頭前溪周圍土地利用或其他人為因素亦是導致誤差偏大之另一可能原因。本研究發展之頭前溪何氏棘魞豐度與環境關係可作為淡水魚類豐度推估相關研究之初步嘗試與前驅，未來若相關模式之發展轉趨穩定與成熟，配合貝氏最大熵法河川推估模式，勢必能夠藉由預測與評估魚類物種在河川中的時空分布，協助補足採樣上的困難，並對往後外來魚種管理、河川生物多樣性管理與棲地暨物種復育計畫有所貢獻。

Stream network consists of abundant abiotic and biotic resources and there have been plenty of researches about stream systems. Nevertheless, almost all these researches of freshwater fish community structure and habitat environment have been constrained by considerable costs and hard-work sampling. Recently, many geostatistical methods have been developed and used to estimate data at unsampled sites, and thus the spatial resolution of interested data could be effectively enhanced. Geostatistics could not replace ecological studies, but serves as tool for data modelling. Traditional geostatistical methods are developed based on strict hypothesis of normal distribution and linear estimation, and that only hard data could be used and analyzed, besides, they could not be applied to estimation in stream networks. Bayesian Maximum Entropy (BME) method is a newly developed statistical method with hypothesis which is more flexible, not restrained by normal distribution and linear estimation theories; moreover, BME combines the concept of Bayesian conditional probability, physical knowledge and other soft data to gradually strengthen the information of estimation. In this study, we applied moving-average construction to integrate traditional covariance models which were in keep with geostatistical assumptions, and used BME method to estimate spatiotemporal distribution of stream habitat factors in Toucian River. Spinibarbus hollandi is a Taiwan endemic freshwater species mainly inhabitates in southern and eastern Taiwan river basins. In recent decade, S. hollandi was spread to western Taiwan streams due to anthropogenic releasing, and local stream habitat has been interfered, especially Toucian River in Hsinchu. In this study, we chose flow speed and water depth as our abiotic indicator, and applied structured additive regression (STAR) model to analyze relationships between those selected abiotic factors and invasive fish species, which was S. hollandi, in Toucian River; we hoped results of the model could serve as reference for stream manage programs including eliminating invasive species and habitat conservation. The results showed that covariance matrix for stream network could be provided using moving-average construction. In results of habitat factor estimation, there were seldom high relative errors using because of particular characteristics of certain stations and lack of information provided from the neighborhoods. Furthermore, the results indicated that pools with moderate depth and low flow speed in Toucian River were preferred by S. hollandi, and it was a little different between researches of the relation between S. hollandi and stream habitats in its native stream systems. Mean relative errors of STAR model ranged from 30% to 400%, and except potential in-site sampling error and sparsity of training data, the high mean relative errors would be caused by exclusion of human activity data in the process of model building. The relation between S. hollandi and habitat modelled in this study could be considered as a preliminary study and an outset of freshwater fish abundance modeling, and as collocating stream BME model, future studies would develop fish abundance estimation model with more robustness and precision. In the future, freshwater fish abundance estimation model would reduce sampling and monitoring costs through providing more spatiotemporal estimation results as reference data, and could be promoted to other stream networks. With the ability to estimate spatiotemporal distribution of fish species and environmental data, people may construct more proper conservation and restoration plans in the future.