| 摘要: | 河川為地球主要淡水來源，對生態系統與人類活動具關鍵功能。臺灣受中央山脈阻隔，短陡河川一方面承擔公共用水、農業灌溉與工業用水，另一方面維繫水生棲地與沿海生態。濁水溪、高屏溪等流域分別支撐嘉南平原及高屏地區之農業與產業發展，也是區域民生用水與環境安全的重要命脈，然而河川水質受工業、農業、畜牧與民生污水等多源污染影響，導致治理決策不確定性。本研究首先整合環境部全臺水質監測資料，結合土地利用，進行因子分析，以主成分萃取並採用Varimax轉軸，將水質與土地利用間高度相關的資訊濃縮為少數「潛在污染因子」得出因子一以亞硝酸鹽氮、硝酸鹽氮與銅、鋅、鎳及工業土地利用類呈現正向負荷，反映工業排放之氮與重金屬複合污染。因子二以氨氮，生化需氧量、總磷與大腸桿菌群和森林與溶氧量呈強負負荷。顯示森林覆蓋可提升溶氧並抑制生活與畜牧污水相關有機污染。接續以氨氮(NH₃–N)為核心，利用倒傳遞神經網路(BPNN)建立濃度迴歸模型，交叉驗證評估模型之泛化能力；結果顯示平均驗證表現為MAE=1.18、RMSE=1.98、R²=0.44，對數轉換後之觀測與預測散佈關係亦呈現一定解釋力(R²=0.519)，顯示模型可掌握主要濃度變異結構並具備實務可用性。另為治理溝通，將迴歸預測值依環境部飲用水水質標準(0.1 mg/L)門檻轉為超標風險指標，其判讀準確率85%、精確率89%、AUC=0.895，並可重現超標比例(實測59.4%、預測60.7%)。綜合結果顯示，氨氮風險西岸都市化區較高且呈熱區，森林與降雨具稀釋與淨化效益，人造建成區與特定工業用地推升負荷，可作為監測分佈點與治理優先序依據。;Rivers are a primary source of freshwater on Earth and play indispensable roles in both ecosystems and human society. In Taiwan, the Central Mountain Range forms steep and short river systems that not only supply domestic, agricultural, and industrial water demands, but also sustain aquatic habitats and coastal ecosystems. Major basins such as the Zhuoshui River and Gaoping River support agricultural and industrial development in the Chianan Plain and the Kaohsiung–Pingtung region, serving as vital lifelines for regional livelihoods and environmental security. However, river water quality is influenced by multiple pollution sources—including industrial effluents, agricultural runoff, livestock wastewater, and domestic sewage—thereby increasing uncertainty in management and decision-making.
This study first integrated nationwide water quality monitoring data from the Ministry of Environment (MOENV) with land-use information and conducted factor analysis using principal component extraction with Varimax rotation. Highly correlated patterns between water quality and land use were condensed into a limited number of latent pollution factors. Factor 1 showed positive loadings for nitrite nitrogen, nitrate nitrogen, copper, zinc, and nickel, together with industrial land-use categories, indicating combined nitrogen and heavy-metal contamination associated with industrial discharges. Factor 2 exhibited strong negative loadings for ammonia nitrogen, biochemical oxygen demand, total phosphorus, and Escherichia coli, as well as forest coverage and dissolved oxygen, suggesting that forested areas enhance dissolved oxygen and suppress organic pollution linked to domestic and livestock wastewater.
Subsequently, focusing on ammonia nitrogen (NH₃–N), a backpropagation neural network (BPNN) was developed to construct a concentration regression model, and cross-validation was applied to evaluate its generalization performance. The results showed an average validation performance of MAE = 1.18, RMSE = 1.98, and R² = 0.44. The scatter relationship between observed and predicted values after logarithmic transformation also demonstrated explanatory power (R² = 0.519), indicating that the model captures major concentration variability and is practically applicable. To facilitate risk communication for governance, the regression outputs were further converted into an exceedance risk indicator using the MOENV drinking-water quality threshold of 0.1 mg/L. The resulting classification achieved an accuracy of 85%, a precision of 89%, and an AUC of 0.895, while successfully reproducing the exceedance proportion (observed: 59.4%; predicted: 60.7%). Overall, the findings indicate that NH₃–N risk is higher and forms hotspot patterns in urbanized areas along Taiwan’s western corridor. Forest cover and rainfall provide dilution and purification effects, whereas built-up areas and specific industrial land-use types increase pollution loads. These results can support monitoring site allocation and help prioritize management actions. |
